Two-stage quality adaptive fingerprint image enhancement using Fuzzy C-means clustering based fingerprint quality analysis

Image and Vision Computing 83-84 (2019) 1–16

Contents lists available at ScienceDirect

Image and Vision Computing journal homepage: www.elsevier.com/locate/imavis

Two-stage quality adaptive fingerprint image enhancement using Fuzzy C-means clustering based fingerprint quality analysis夽 Ram Prakash Sharma* , Somnath Dey Discipline of Computer Science and Engineering, Indian Institute of Technology Indore, India

A R T I C L E

I N F O

Article history: Received 23 February 2018 Received in revised form 11 December 2018 Accepted 21 February 2019 Available online 9 March 2019 Keywords: Biometrics Fingerprint image quality Fuzzy C-means clustering Fingerprint image enhancement Fingerprint matching

A B S T R A C T Fingerprint recognition techniques are dependent on the quality of fingerprint images. An efficient enhancement algorithm improves the performance of recognition algorithms for poor quality images. Performance improvement of the recognition algorithms will be more if the enhancement process is adaptive to the fingerprint qualities (wet, dry or normal). In this paper, a quality adaptive fingerprint enhancement algorithm is proposed. The proposed fingerprint quality assessment (FQA) algorithm assigns the appropriate quality class of dry, wet, normal dry, normal wet, and good quality using Fuzzy C-means clustering technique to each fingerprint image. It considers seven features namely, mean, moisture, variance, uniformity, contrast, ridge valley area uniformity (RVAU), and ridge valley uniformity (RVU) to cluster the fingerprint images into suitable quality class. Fingerprint images of each quality class undergo through a two-stage fingerprint quality enhancement (FQE) process. In the first stage, a quality adaptive preprocessing (QAP) method is used to preprocess the fingerprint images. Next, fingerprint images are enhanced with Gabor, short-term Fourier transform (STFT), and oriented diffusion filtering (ODF) based enhancement techniques in the second stage. Experimental evaluations are performed on a quality driven database of FVC 2004. Results show that the performance improvement of 1.54% to 50.62% for NBIS matcher and 1.66% to 8.95% for VeriFinger matcher are achieved while the QAP based approaches are used in comparison to the current state-of-theart enhancement techniques. In addition, the experimentation is also performed on FVC 2002 database to validate the robustness and efficacy of the proposed method. © 2019 Elsevier B.V. All rights reserved.

1. Introduction In the recent era, biometric based authentication is popular due to its uniqueness among individuals. Many countries have extensively deployed it for the identification of persons such as US Visitor and Immigration Status Indicator Technology (US-VISIT) [1], India’s unique identification card by Unique Identification Authority of India (UIDAI) [2], and UK Iris Recognition Immigration System (IRIS), etc. Different biometric traits such as fingerprints, face, iris, voice, and signature are used for automatic identification of an individual in entry control, ATMs, e-passports [3], etc. Fingerprint-based recognition [4,5] is one of the most reliable ways of biometric authentication because of their universality, distinctiveness, and accuracy. However, fingerprint recognition algorithms are highly dependent on the quality of the acquired fingerprint images [6]. Fingerprint quality analysis can

夽 This paper has been recommended for acceptance by Sinisa Todorovic. * Corresponding author. E-mail addresses: [email protected] (R.P. Sharma), [email protected] (S. Dey).

https://doi.org/10.1016/j.imavis.2019.02.006 0262-8856/© 2019 Elsevier B.V. All rights reserved.

be utilized for acquisition of good quality fingerprint images, fingerprint image enhancement [7], and liveness analysis for fake fingerprint detection [8], etc. Due to some environmental and physical conditions of user’s skin, sometimes it may be quite difficult to acquire a good quality fingerprint image. Further, most of the fingerprint recognition algorithms utilize minutiae points for matching [9]. A poor quality image may output spurious minutiae points and ignore some genuine minutiae points which may degrade the recognition performance. Therefore, a robust fingerprint enhancement algorithm is required for the low-quality images resulting into accurate detection of minutiae points. Several on-going and past efforts have been made for the enhancement of fingerprint images [10,11]. Some of these techniques are based on spatial domain methods such as contextual filters [10], Gabor filters and its variations [12,13], and oriented diffusion filtering (ODF) [14]. Other enhancement methods utilize frequency domain methods such as log-Gabor filters [15], short-term Fourier transform (STFT) [16], directional Fourier transform, and simple Fourier analysis [17]. Single stage processing either in frequency or in spatial domain for enhancement of low-quality fingerprint images is not enough due to the difference in ridge structure of individuals. To address this

2

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

limitation, some of the recent methods [18,19] combine both the frequency and spatial domain methods to obtain better enhancement results. However, these methods depend on the accurate extraction of contextual information (ridge frequency and orientation). It is quite difficult to precisely extract ridge orientation and frequency information from low-quality fingerprint images. Due to the erroneous computation of these information, enhanced fingerprint images can produce some spurious minutiae points and ignore some genuine minutiae points. Thus, a method providing adequate enhancement for high or good quality fingerprint images may not be able to give proper enhancement for low or bad quality fingerprint images. Therefore, it is desirable that enhancement process should be adaptive to the fingerprint quality. Hence, the primary objective of our work is to design a quality adaptive enhancement method for fingerprint images of different qualities. In order to alleviate the limitations of the existing enhancement algorithms, a quality adaptive fingerprint enhancement method is proposed in this work. Fig. 1 shows the schematic diagram of our proposed method concerning traditional enhancement methods. It shows that our proposed method preprocesses the fingerprint images based on its quality nature (good or bad, dry or wet) before the final enhancement. This helps in accurate extraction of fingerprint properties (orientation and frequency) from different qualities of fingerprint images. On the contrary, existing enhancement methods use similar enhancement for all fingerprint images irrespective of their quality nature. The proposed technique enhances the fingerprint images based on their quality classes (Qc ) which can be dry, wet, normal dry, normal wet or good. The proposed method works in two phases. In the first phase, a fingerprint quality assessment (FQA) method is designed to cluster fingerprint images into different quality classes. In FQA, a set of eleven features namely moisture (MI), mean (M), variance (V), ridge valley area uniformity (RVAU), ridge line count (RLC), uniformity, contrast, radial power spectrum (RPS), ridge valley uniformity (RVU), Gabor, and Gabor-Shen is extracted from the fingerprint image. The best feature subset among all the extracted features is obtained using a novel one-way analysis of variance (ANOVA) statistical test. The obtained best feature subset includes uniformity, contrast, mean, variance, moisture, RVAU, and RVU. These features are the most discriminative for quality clustering of fingerprint images. The selected feature subset is fed as input to the Fuzzy C-means clustering algorithm which clusters fingerprint images into different quality classes. In the second phase, two-stage fingerprint quality enhancement (FQE) is performed which comprises of a quality adaptive preprocessing (QAP) method followed by either Gabor or STFT or ODF enhancement techniques. After FQA, each quality cluster of fingerprint images is fed to the QAP which comprises of parameter adaptive unsharp masking filter, contrast limited adaptive histogram equalization (CLAHE), and Gaussian smoothing. The preprocessed

images are further enhanced with the state-of-the-art enhancement methods (eg. Gabor/STFT/ODF). The proposed fingerprint enhancement scheme preprocesses the images adaptively based on their quality classes. These quality classes represent dry, normal dry, good, normal wet, and wet quality natures of fingerprint images. In a nutshell, the major contributions of the proposed work are summarized as follows: • A novel set of quality features (moisture, mean, variance, RVAU, and RLC) is proposed to analyze dry, normal dry, good, normal wet, and wet quality nature of fingerprint images. • The proposed FQA method investigates joint contribution of the proposed quality features and few existing quality features (uniformity, contrast, RPS, RVU, Gabor, and Gabor-Shen) to assess fingerprint image quality. • A novel one-way ANOVA statistical test based feature selection method is proposed to obtain the set of discriminative features which can accurately differentiate fingerprint images based on their characteristics (dry, normal dry, good, normal wet, and wet). • An objective methodology to cluster fingerprint images into different quality classes using Fuzzy C-means clustering algorithm has been proposed. • A QAP approach is proposed to enhance the fingerprint images which preprocesses the images according to their quality classes. • The proposed two-stage FQE approach achieves significant performance improvement by using QAP as front-end with the Gabor/STFT/ODF fingerprint enhancement algorithms. The rest of the paper is organized as follows. In Section 2, an overview of related works on fingerprint quality enhancement and assessment is presented. Section 3 describes the proposed method. The experimental results of FQA and FQE are thoroughly discussed and presented in Section 4. Finally, conclusion is drawn in Section 5 with a glimpse of future research direction.

2. Related works The existing fingerprint enhancement algorithms are generally classified in two types: spatial domain and frequency domain enhancement techniques. Further, there are several methods [18-21] in the literature which use two stages of enhancements. Some of these existing techniques [18,19] use spatial and frequency domain methods for efficient and precise extraction of contextual information (ridge orientation and frequency) to enhance the low-quality fingerprint images. Further, some of the fingerprint enhancement

Existing Enhancement Techniques Enhancement algorithm

Matching

Quality analysis

Quality based preprocessing

Original fingerprint image Enhancement algorithm

Matching

Proposed Enhancement Technique Fig. 1. Schematic diagram of the proposed approach with respect to the traditional enhancement methods.

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

methods [22,23,11] use quality (dry, wet or good) information of fingerprint images in combination with spatial or frequency domain filters. Therefore, all the existing enhancement methods are divided into four categories namely, spatial domain, frequency domain, two-stage, and quality adaptive (spatial or frequency) fingerprint enhancement techniques. 2.1. Fingerprint enhancement techniques Salient features of the few existing fingerprint enhancement techniques are shown in Fig. 2. Here, we discuss few relevant fingerprint enhancement techniques of these categories present in the literature. • Spatial domain enhancement methods: O’Gorman et al. [10] proposed a technique to design contextual filters for fingerprint image enhancement. The contribution of their work is to quantify and justify the functional relationships between fingerprint image features (ridge orientation, ridge width, valley width) and filter parameters (filter mask size and filter parameters). For efficient enhancement, the filter is precomputed in 16 different directions to enhance the fingerprint image by applying oriented matched filter masks. The most popular and widely used spatial domain fingerprint image enhancement algorithm is Gabor based enhancement algorithm which is proposed by Hong et al. [12]. Gabor filter employs the frequency and orientation information for enhancement of images. Filter parameters are adapted for each block according to the ridge frequency and orientation to preserve the actual ridge-valley structure. Any error in the computation of ridge orientation and ridge frequency for the blocks of fingerprint image may cause an adverse effect on the enhancement process which influences the recognition performance. Another variation of the traditional Gabor filter for fingerprint enhancement is proposed by Gottschlich et al. [13], which used curved Gabor filter for enhancing curved ridge valley structures. Curved Gabor filter locally adapts their shape to the direction of flow and enables the selection of parameters. Gottschlich et al. [14] proposed an orientation diffusion filtering based enhancement technique. Orientation diffusion filtering utilizes local orientation of the fingerprint ridge and valley flow, followed by locally adaptive contrast enhancement. Few other spatial domain fingerprint enhancement techniques are presented in Refs. [24-26]. The spatial domain enhancement methods don’t perform well for the lowquality fingerprint images as the computation of accurate and precise ridge orientation and frequency is quite difficult due to corrupt ridge-valley structure. • Frequency domain enhancement methods: Wang et al. [15] proposed a frequency domain implementation of Log-Gabor filter. Log-Gabor filter efficiently improves the contrast between

3

ridges and valleys, and preserves the ridge structure of fingerprint images since it does not involve any DC component. The most widely used fingerprint enhancement approach in the frequency domain is proposed by Chikkerur et al. [16]. They utilized STFT analysis and contextual filtering in Fourier domain for fingerprint quality enhancement. This algorithm has an advantage that it computes all the intrinsic images (ridge orientation, ridge frequency, and foreground region mask) simultaneously using STFT analysis and utilizes full contextual information for enhancement. A top-down iterative filtering based fingerprint enhancement is proposed by Zhu et al. [17]. A new filter is formed by combining the filters proposed in Refs. [27,28] to enhance the fingerprint image. This method initially filters the entire image and then divides it into sub-images for further enhancement using the same method until the sub-image size reaches to a pre-defined threshold. This scheme is not well suited for high curvature fingerprint images which contain fragmentary ridges. Other variations of the frequency domain fingerprint enhancement methods can be found in Refs. [29,30]. Majority of the frequency domain enhancement methods also require the computation of intrinsic properties (ridge orientation and ridge frequency) which are not estimated accurately in low-quality fingerprint images. • Two-stage enhancement methods: In this category, Yang et al. [18] proposed an efficient two-stage, spatial and frequency domain enhancement scheme by learning from the fingerprint images. In the first stage, a spatial ridge compensation filter is used to enhance ridge areas and contrast of local ridges. The second stage comprises a frequency bandpass filter which is separable in angular and radial frequency domain. The parameters of the bandpass filter are learned from the input image and the first stage enhanced image. Ghafoor et al. [19] proposed a method based on 2-fold local adaptive contextual filtering. Initially, frequency domain filtering is performed using a bandpass filter followed by local directional filtering in the spatial domain to enhance the fingerprint images. Shifei et al. [21] proposed a method which combines the Gabor filter with classification dictionaries for enhancement. Their enhancement process is divided into two phases. In the initial phase, fingerprints are enhanced using the Gabor filter. In next phase, classification dictionaries based on spectra diffusion are used for enhancement. Shifei et al. [20] designed a 2D adaptive Chebyshev bandpass filter (ACBF) with orientation-selective in the frequency domain for fingerprint enhancement. In their method, fingerprint images are first enhanced by Gabor filter and histogram equalization. The enhanced fingerprint images are further enhanced based on spectra diffusion using the 2D ACBF with orientation-selective. These methods are just a combination of frequency and spatial domain methods. Therefore, their drawbacks remains the same as frequency and spatial domain methods. Additionally, the

Fingerprint Enhancement Spatial

Frequency

Two-stage

Contextual filters [10] Gabor filters [12] Curved gabor filter [13] Orientation diffusion filtering [14] Anisotropic filters [20]

Log-Gabor filters [15] STFT [16] Fourier analysis [17] Wavelets [25] Discrete cosine transform [26]

Spatial ridge compensation filter + bandpass filter [18] Bandpass filter + directional filter [19]

Fig. 2. Existing fingerprint enhancement techniques.

Quality-based Gabor filter on high to low quality region [11] Morphological processing [29] Parameter adaptive matched directional filter [30] Gaussian matched filter on high to low quality blocks [31]

4

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

complexity of these methods is more as these methods require processing in both the frequency and spatial domain. • Quality adaptive enhancement methods: In this category, there are few fingerprint enhancement methods [22,23,11] presented in the literature which consider quality adaptive aspect of fingerprintimages.Yunetal.[22]proposedamethodthatfollows an adaptive preprocessing approach to enhance the fingerprint quality appropriately. Different morphological preprocessing stages are designed to enhance oily and dry fingerprint images. Oily or dry nature of fingerprint images are determined using five features namely, mean, variance, block directional difference, ridge-valley thickness ratio, and orientation change. Although, the method processes the fingerprint images based on their quality nature, its performance is not comparable to the other well-known methods [12,16]. Bartunek et al. [23] proposed an adaptive fingerprint enhancement method. This method is based on spatial contextual filtering using parameter adaptive matched directional filters. Turroni et al. [11] proposed a methodwhichreliesonthequalityoftheinputfingerprintimage. They proposed a scheme which determines the high and lowquality regions using the combined image after convolution with Gabor filter bank and local ridge flow homogeneity. Fingerprint enhancement is done selectively using contextual filtering starting from high-quality regions to low-quality regions. Sutthiwi et al. [31] proposed quality diffusion based iterative fingerprint enhancement method by using a Gaussian-matched filter. First, enhancement is performed on the high-quality blocks using Signal to Noise ratio (SNR). Next, these enhanced blocks are fed to unreliable low-quality regions for effective enhancement. Since, most of the quality based methods are iterative in nature, their computational complexity is high. Nowadays, deep learning techniques particularly based on convolution neural network (CNN) have been successfully used in various applications [32,33]. Recently, Li et al. [34] proposed the use of deep CNN for latent fingerprint enhancement. Their enhancement model FingerNet includes one encoding convolutional part and two decoding deconvolutional parts. The enhanced latent fingerprint image is generated as output from the pixel-to-pixel and end-to-end learning based trained network. A comparative study of the several fingerprint enhancement techniques can be found in Refs. [35,36]. 2.2. Fingerprint quality analysis methods To assess the quality nature of fingerprint images, some fingerprint quality analysis methods [37-39] are presented in literature. Lim et al. [37] proposed a block quality classification method to classify the fingerprint image blocks into very good, good, bad, and very bad quality clusters. Their feature vector comprises of directional strength, sinusoidal local ridge/valley pattern, ridge/valley uniformity, and core occurrences of a fingerprint image sub-block. Tabassi et al. [39] proposed a classifier based method which defines the quality as the degree of separation between match and non-match distributions of a given fingerprint. It classifies the quality of fingerprint image into five levels: poor, fair, good, very good, and excellent using features vector consisting of clarity of ridges and valleys, size of the image, and a measure of number and quality of minutiae. Munir et al. [38] proposed a hierarchical k-means clustering based fingerprint quality classification method which classifies fingerprint images as dry, wet, and normal. A set of frequency (energy concentration) and statistical features (mean, uniformity, smoothness, image inhomogeneity, etc.) are utilized to classify fingerprint images in different quality class. Some other fingerprint quality analysis methods can be found in Refs. [40-42]. There are few methods in the literature which assess the impact of fingerprint quality nature on recognition performance [43-46].

Olsen et al. [43] analyzed the impact of fingerprint skin moisture on the recognition performance of the system. They have collected an in-house data set of fingerprint images while varying the moisture level of a fingertip using cream, alcohol, and water. Experimental evaluations on the collected data set show that a controlled level of moisture is required in order to achieve good biometric performance. Moisture level above or below a threshold (wet and dry) causes degradation in the performance of a biometric recognition system. Labati et al. [44] proposed a method which detects a set of non-ideal conditions of a fingertip such as dirt, grease, carrying a shoulder bag, carrying a handbag, and normal in Automated Border Control (ABC) systems. These conditions degrade the quality of the acquired fingerprint images which may degrade the performance of the recognition system. Their proposed method detected these non-idealities using Histogram of Oriented Gradients (HOG), mean, variance, gray-level co-occurrence matrix, and ridge orientation features. A feedback message is displayed to improve the condition of fingertip or sensor and provide the fingerprint impression again. Krishnasamy et al. [45] evaluated the performance of two recognition algorithms when matching dry-finger to dry-finger, and dry-finger to wet-finger. They evaluated the performance on a self-made data set containing dry and wet fingerprints. Their results exhibit that the verification errors rates for both the recognition algorithms are increased when the gallery samples are dry and probe samples are wet. Patil et al. [46] studied the challenge of recognizing dry or wet fingerprints for the authentication in an attendance system. They have utilized rotational filtering to enhance the dry and wet fingerprint images which improves the performance of the biometric attendance system. 3. Proposed method Analysis of the traditional prior works [12,14,16] show that the existing fingerprint enhancement algorithms are not adequate to enhance different low-quality (dry, wet, normal dry, and normal wet) fingerprint images. To overcome the demerits of these methods, a new and effective quality adaptive enhancement method is proposed. The block diagram of the proposed method is illustrated in Fig. 3. Proposed enhancement method works in two phases: 1) Fingerprint quality assessment and 2) Fingerprint quality enhancement. In the first phase, an FQA algorithm is designed to classify fingerprint images into suitable quality class Qc , where c represents the quality class (dry, normal dry, good, normal wet, and wet). The second phase comprises the two-stage FQE scheme in which the fingerprint images are preenhanced by QAP in the first stage. Later, in the second stage, QAP based enhanced fingerprint images are fed to the Gabor/STFT/ODF enhancement techniques for further enhancement. A detailed description of each processing block is provided in the following sections. 3.1. Fingerprint quality assessment (FQA) In FQA, quality relevant features of the fingerprint images are computed. Thereafter, the most discriminative features are selected from all the extracted features which are fed to a Fuzzy C-means clustering algorithm as input. The fingerprint images are classified into five quality classes Qc where c can be dry, normal dry, good, normal wet, and wet. Properties of these defined quality classes are as follows: • Dry images are formed because of low pressure on the scanner surface or dryness of skin which in turn produces broken ridges [22,38] as shown in Fig. 4 (a). A large number of false minutiae can be detected from these images due to the broken ridges. • Normal dry quality fingerprint images contain scratchy ridges whose thickness is less than the subsequent valley region as shown in Fig. 4 (b). In these images, most of the regions have

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

5

Dry Normal dry Biometric samples

Feature extraction

One-way ANOVA feature selection

Fuzzy C-means clustering

Good Normal wet

Phase 1: Fingerprint Quality Assessment (FQA) Wet

Second Stage Enhancement

Enhanced biometric samples

First Stage Enhancement

Gabor/STFT/ODF enhancement

Gaussian smoothing

Traditional Enhancement Methods

Contrast limited adaptive histogram equalization (CLAHE)

Unsharp masking (Parameter tunning Rquality, Aquality)

Quality Adaptive Preprocessing (QAP)

Phase 2: Fingerprint Quality Enhancement (FQE)

Minutiae extraction

Biometric references

BOZORTH3

Comparison score

Fig. 3. Block diagram of the proposed two-stage fingerprint enhancement method using fingerprint quality analysis.

medium ridge-valley contrast [38] and contain a sufficient number of minutiae points for accurate recognition. • Good quality fingerprint images are those which have clearly separated ridge-valley structure [22,38] so that a minutiae extraction algorithm is able to operate well, as shown in Fig. 4 (c). These images have precisely located minutiae points which aid to improve the performance of a recognition system. • Normal wet fingerprint images have dark and hazy ridges with less valley regions as depicted in Fig. 4 (d). These images contain sufficient number of minutiae points [38] for accurate recognition. • Wet fingerprint images are too dark and don’t have clear ridgevalley separation which makes it quite difficult to separate ridge-valley structure in them. In these images, most of the valley regions are filled with moisture due to high pressure or oily skin [22,38] as shown in Fig. 4 (e). 3.1.1. Feature extraction Quality of a fingerprint image is usually measured in terms of clarity of ridge-valley structure. Different states of this ridge-valley structure form different quality classes of fingerprints. These different classes of quality are caused by many factors such as user-specific

factors like age, social customs, gender, and injuries or environmental issues such as temperature and humidity, or user-sensor interaction based on the indoor or outdoor use of sensors, dust on sensors, etc. Therefore, some relevant features which exhibit properties of the defined quality classes are extracted to classify the fingerprint images into different quality classes. Local variance based segmentation approach [47] is followed to find the foreground region of the fingerprint image. All of these features are computed in a 16 × 16 block-wise manner in the region of interest of the fingerprint image. We have selected a block of size 16 × 16 to accommodate at least one ridge valley pair since a ridge valley pair is 8–12 pixel wide in a 500 dpi fingerprint image [9]. We have extracted the following five features namely, moisture, mean, variance, RVAU, and RLC from the fingerprint images where moisture, RVAU, and RLC are computed from the binary image obtained using Otsu’s method [48]. Feature computation for the foreground blocks (FB) of a fingerprint image is carried out as follows: (i) Moisture (MI): Moisture level on a fingerprint may influence the dry or wet nature of a fingerprint image. Therefore, moisture of a fingerprint image should be computed effectively to assess its quality. The moisture level of a block is computed as

Fig. 4. Fingerprint images of different qualities from FVC2004 DB1 data set. (a) Dry, (b) Normal dry, (c) Good, (d) Normal wet, and (e) Wet.

6

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

the percentage of unwanted ridge pixels in the valley region of a fingerprint block. The percentage of unwanted ridge pixels is obtained by subtracting the percentage of average number of ridge pixels (k) in good quality blocks (marked by the human experts) from the percentage of total number of ridge pixels in that block. The value of k is obtained by averaging the percentages of ridge (black) pixels in 500 good quality blocks classified by human experts. Averaging the moisture level of all foreground blocks of a fingerprint image provides the overall moisture of the fingerprint image, I. Dry fingerprint images have moisture level in negative, which indicates low moisture level on the fingertip. This low moisture level causes less contact with sensor platen, leading to broken ridges. On the contrary, a positive moisture level of wet fingerprints indicates that some of the valley regions are local filled with dark pixels. The moisture (MIFB ) for m × n size foreground block FB is computed using Eq. (1). ⎛

local MIFB

⎞ j=n i=m  1 =⎝ FB(i, j)[FB(i, j) = 0]⎠ × 100 − k (1) m×n i=1 j=1

k=



⎫ ⎧ c=500 i=m   j=n  ⎪ ⎪ 1 ⎪ ⎪ ⎪ ⎪ B (i, j)[B (i, j) = 0] cgood ⎪ ⎪ ⎨ c=1 m×n i=1 j=1 cgood ⎬ ⎪ ⎪ ⎪ ⎪ ⎩

500

⎪ ⎪ ⎪ ⎪ ⎭

× 100

local where, MIFB represents the moisture level of the foreground block FB, FB(i, j) is the value at pixel location (i, j) of a binary block FB, and m × n is the count of all pixels in the block (i.e. 162 = 256). Here, we obtain k value as 51.25% using Eq. (2) where Bcgood represents a good quality block marked by human experts. Global moisture level (MIglobal ) of the fingerprint image is computed using Eq. (3) where |FB| is number of foreground blocks in a fingerprint image.

1 |FB|

i<=X,j<=Y



local MIFB(i,j)

i=m j=n 1  FB(i, j) m×n

(4)

i=1 j=1

Global mean (Mglobal ) of the fingerprint image is computed using Eq. (5) where |FB| represents the count of number of foreground blocks in the fingerprint image.

Mglobal =

1 |FB|

i<=X,j<=Y



i=1,j=1

local MFB(i,j)

i=m j=n 1  local 2 (FB(i, j) − MFB ) m×n

(6)

i=1 j=1

V global =

1 |FB|

i<=X,j<=Y



local VFB(i,j)

(7)

i=1,j=1

(iv) Ridge valley area uniformity (RVAU):In general, a good quality fingerprint image contains equal regions of ridges and valleys with identical thickness throughout the foreground region of the fingerprint image. But, it is found that the distribution of ridge and valley pattern across different foreground blocks in a fingerprint image is not uniform in real scenarios. Therefore, determining RVAU in a fingerprint image will help to predict the overall quality of the fingerprint image. The ratio of ridge area versus valley area is computed using total number of ridge and valley pixels in a binary image. RVAU local (RVAUFB ) of a foreground block FB is computed using Eq. (8).

i=1 local RVAUFB =

|m × n| −

B(i, j)[FB(i, j) = 0]

j=1 i=m  j=m 

(8) FB(i, j)[FB(i, j) = 0]

i=1 j=1

Global RVAU (RV AUglobal ) of a fingerprint image is computed using Eq. (9) where |FB| represents the count of number of foreground blocks in a fingerprint image.

RVAU global =

1 |FB|

i<=X,j<=Y



local RVAUFB(i,j)

(9)

i=1,j=1

(3)

i=1,j=1

local where, MIFB(i,j) is the moisture level of a foreground block, and (i, j) represents the horizontal and vertical index of 16 × 16 size blocks. (ii) Mean (M): Mean of the fingerprint image is considered only for the foreground area which shows the overall gray level of the image. Fingerprint images of different qualities have different mean distributions which make it a good feature for local quality prediction. Mean (MFB ) of each foreground block FB of size m × n in a fingerprint image is computed using Eq. (4).

local MFB =

local VFB =

 i=m j=n (2)

MIglobal =

(iii) Variance (V): Variance is calculated to obtain overall uniformity of gray level in foreground region of a fingerprint image. local Variance (VFB ) of each foreground block FB is computed using Eq. (6), and overall variance (Vglobal ) of fingerprint image is computed using Eq. (7).

(5)

(v) Ridge line count (RLC): Counting the number of ridge lines in each foreground block provides an indication of the block quality which constitutes to the overall quality of the fingerprint image. The motivation for considering RLC as a candidate feature to predict fingerprint quality is that a good fingerprint contains a constant number of ridge lines in almost all the foreground blocks. It is possible that some of the foreground blocks may have zero ridge lines in a dry image due to less contact with sensor platen. Similarly, wet fingers exhibit opposite effect where the fingerprint image looks very dark and some of the foreground blocks may contain only 1 ridge of abnormal thickness. Therefore, evaluation of RLC is required as it is a good indicator to determine the overall quality of the fingerprint image. RLC for foreground block of a fingerprint image is computed using the ridge map of the block with one-pixel thickness. The ridge map is obtained by thinning morphological operation which is first rotated vertically using orientation estimation method adapted from Ref. [39] to make the ridge lines vertical. The orientation of each block is computed using the method proposed in Ref. [39], where the numerical gradient of the block is determined using the finite central difference for all interior pixels in the x-direction (dx) and y-direction (dy). The covariance matrix (CFB ) for the foreground block is computed

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

using principal components analysis. The expression of CFB is defined in Eq. (10).

CFB =

      1  dx  a c = dx dy dy c b m ∗ n m∗n

(10)

Orientation of a foreground block (thetaFB ) is computed using Eq. (11). thetaFB = tan

−1



c c2 + (a − b)2



a−b c2 + (a − b)2

(11)

Next, the block is rotated using the orientation of the block in such a way that the ridges are at vertical position. Further, thinning morphological operation is performed to count the pixel bit-flips (white to black and black to white) for each row of a block. This gives the count of twice the number of ridges present in each row of a block. The half of the maximum of local these counts is considered as RLC of the block (RLCFB ) which is defined in Eq. (12).

local RLCFB

⎛ ⎞ j=n  1 = ×max i ⎝ FB(i, j)[b− > w||w− > b]⎠ for i = 1 to m 2 j=1

7

image. The procedure for computation of RVU is inherited from Olsen et al. [50]. Gabor: The computation of fingerprint quality using a Gabor filter bank with different orientations is done at local level. The strength of the Gabor response for a fingerprint block containing regular ridge-valley pattern will be high for one or few filters having orientations similar to block orientation. On the other hand, Gabor responses for all orientations will be low and constant for the block containing unclear ridgevalley structure. The standard deviation of the Gabor filter bank responses is computed which indicates the Gabor quality of the block. The Gabor feature computation procedure is followed as provided in Ref. [50]. Gabor-shen: Gabor-shen measures the fingerprint quality using the Gabor filter bank responses obtained on each block and their standard deviation. Using a pre-defined threshold, each block is classified as foreground, background, and poor or good quality. The ratio of the number of poor foreground blocks to the number of foreground blocks indicates the GaborShen quality measure for a fingerprint image. The Gabor-Shen quality is computed by following the method proposed in Olsen et al. [50]. As a result, we define a 11-dimensional feature vector (F) comprising moisture, mean, variance, RVAU, RLC, uniformity, contrast, RPS, RVU, Gabor, and Gabor-shen features. This feature vector is fed as input to feature selection unit to select the most discriminative feature subset for quality clustering of fingerprint images.

(12) Average RLC for the entire fingerprint image (RLCglobal ) is computed using Eq. (13).

RLC global =

1 |FB|

i<=X,j<=Y



local RLCFB(i,j)

(13)

i=1,j=1

(vi) Other features: Apart from the proposed features, fingerprint quality is influenced by some other features defined in the literature. Therefore, to make fingerprint quality clustering more accurate, few other features are also evaluated. In this work, we will cluster fingerprint images in dry, normal dry, good, normal wet, and wet classes. Therefore, features which influence these quality classes are considered. For this purpose, two additional features namely, uniformity and contrast from Ref. [49] and four other features namely, radial power spectrum (RPS), ridge valley thickness uniformity (RVU), Gabor, and Gabor-shen from Ref. [50] are considered in the candidate feature vector for quality prediction of fingerprint images. The description of these features are provided as follows: RPS: Radial power spectrum is a global feature which is computed on the entire fingerprint image. It measures the maximum signal strength in a defined frequency band of the Fourier spectrum. Ridges can be approximated by a sine wave. Hence, regions of consistent ridge structure exhibit high energy concentration in a narrow frequency band. More details of RPS can be found in Olsen et al. [50]. RVU: Ridge valley thickness uniformity measures the consistency of the widths of ridges and valleys. A fingerprint image with clear ridge-valley structure will exhibit a constant ratio between ridge and valley width across all the blocks. On the other hand, this ratio will vary for the fingerprint images with corrupted or partially corrupted ridge-valley structure (dry or wet fingerprint images). The standard deviation of the RVU ratio of each block will indicate the quality of the fingerprint

3.1.2. Feature selection One-way ANOVA test is used to analyze the features based on their discriminative capability to distinguish different quality classes. This test outputs the set of discriminative features which can classify the fingerprint images in suitable quality class. One-way ANOVA computes the dissimilarity between the clusters and similarity within the cluster to estimate that the defined classes are similar with probability p. A good feature exhibits more dissimilarity between the classes and less dissimilarity within the class. Therefore, to show the statistical significance of each feature, a null hypothesis is tested. Null hypothesis affirms that there is no significant difference between means of the different classes for a particular feature (fi ). Acceptance or rejection of the null hypothesis depends on whether the probability value p is less or more than the significance level (a = 0.05). If the probability value is greater than the significance level a, the null hypothesis is accepted indicating that there is no significant difference between the different classes for the given feature. This feature can be rejected from the feature vector F. If the probability value is less than the significance level a, the null hypothesis is rejected which indicates that this feature is able to discriminate the given classes with more than 95% significance level. One-way ANOVA test signifies that there is some statistical difference in all quality classes for each feature, but it doesn’t indicate where this difference lies or whether there is enough statistical difference in each pair of the quality classes. To gain more insight into this, Tukey’s honest significant difference (HSD) test is performed. Tukey’s HSD test identifies whether there is a significant difference between mean feature values for each pair (e.g. (D, ND) and (ND, G)) of the quality classes. The final feature vector Fs is selected after the feature selection phase and fed to the Fuzzy C-means clustering to classify the fingerprint images into an appropriate quality class. 3.1.3. Quality clustering using Fuzzy C-means The Fuzzy C-means clustering algorithm is used for quality clustering of fingerprint images. Fuzzy C-means classifies fingerprint images into five quality classes Qc (dry, normal dry, good, normal wet, and

8

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

wet). It assigns n membership values (between 0 and 1) to each image corresponding to n cluster centers unlike k-means clustering where each image exclusively belongs to one cluster. If properties of a image are similar to a cluster center, its membership value for that cluster will be more and the image will be assigned to the cluster which has the maximum membership value for that image. The summation of the membership values assigned to each image corresponding to different clusters is always 1. 3.2. Two-stage fingerprint quality enhancement (FQE) The proposed FQE works in two stages. In the first stage, a quality adaptive preprocessing is performed which adapts its parameter based on the quality of the fingerprint images. The second stage of FQE utilizes Gabor, STFT, and ODF enhancement algorithms to enhance the fingerprint images using their contextual (orientation and frequency, etc.) information. Description of both stages is given in the following sections. 3.2.1. Quality adaptive preprocessing (QAP) An adaptive fingerprint preprocessing method with different image characteristics is always better than uniform filtering for all images. Conventional filtering of dry fingerprint images may remove black pixels of ridges which are thinner than adjacent ridges. Similarly, removing black pixels in valley region of oily images using conventional filtering may eradicate thinner valleys. Adaptive filtering adjusts parameters of the filter according to the characteristics (dry, wet, and good) of the fingerprint image. The first stage of the proposed FQE scheme is QAP which is illustrated in Fig. 3. QAP method processes the fingerprint image based on its quality nature before the enhancement process. A detailed description of each processing block is narrated in the following. 3.2.1.1. Unsharp masking. The initial step of the preprocessing method utilizes unsharp masking with quality adjustable parameter values for enhancing fingerprint images of each quality class. Unsharp masking filter enhances the brightness of ridges and makes them sharper in dry and normal dry fingerprint images. On the contrary, it clears the valley region of wet and normal wet images by adjusting its radius and amount parameters. Radius parameter controls the size of the region around edge pixel which is affected by sharpening while amount parameter controls how much darker or brighter the pixel will be made. As good quality fingerprint images already have bright ridge pixels and clear valley regions, unsharp masking filter parameters for them should be adjusted in such a way that it should not degrade their quality. Table 1 shows the NFIQ 2.0 scores for 100 good quality fingerprint images (original and enhanced using different parameter values of unsharp masking filter) marked by the human experts. Experimentally, it is observed that radius = 2 and amount = 1 for unsharp masking filter assist in improving the ridgevalley clarity of good images, resulting in higher NFIQ 2.0 scores. Optimal parameter values for dry, normal dry, normal wet, and wet fingerprint images are obtained from the parameter values of good images. As dry fingerprint images have very thin and less bright ridges with clear valley region, the radius of unsharp mask filter should be less and amount should be more to make ridge pixels brighter. Conversely, in wet fingerprint images, radius should be more and amount of unsharp effect should be less to make valley region clear. To obtain these parameter values, we utilize a fingerprint image with highest membership value (m) assigned by Fuzzy C-means clustering. Along with the m values, coefficients are decided for each quality class. The quality gap between dry and good fingerprint images is more as compared to normal dry and good images. Therefore, increase or decrease in parameter value for dry should be more than normal dry images. Same is true for the wet and normal wet fingerprint images. The coefficient for dry and wet fingerprint images cd/w is 1, and for normal

dry and normal wet fingerprint images the coefficient cnd/nw is 0.5. The radius and amount parameter values for dry and normal dry fingerprint images are computed using Eqs. (14) and (15), respectively. Rd/nd = Rg − md/nd × cd/nd

(14)

Ad/nd = Ag + md/nd × cd/nd

(15)

Here, quality can be dry or normal dry based on fingerprint characteristics, md/nd is the membership value for dry or normal dry fingerprint image. The value of md/nd would be in the range of [0,1], and coefficient values for dry and normal dry is cd = 1 and cnd = 0.5, respectively. Rg = 2 and Ag = 1 are radius and amount for good quality fingerprint images. Similarly, radius and amount parameter values for wet and normal wet fingerprint images are obtained from Eqs. (16) and (17), respectively. Coefficient values for wet and normal wet are cw = 1 and cnw = 0.5, respectively. Rw/nw = Rg + mw/nw × cw/nw

(16)

Aw/nw = Ag − mw/nw × cw/nw

(17)

3.2.1.2. Contrast limited adaptive histogram equalization (CLAHE). After the quality adaptive unsharp masking, the contrast of the fingerprint image is enhanced using CLAHE. CLAHE is more suitable than normal histogram equalization for improving the local contrast of the fingerprint images due to the fact that amplification of noise is less in CLAHE as compared to normal histogram equalization. Contrast limited adaptive histogram equalization enhances the contrast of fingerprint image while preserving the brightness which enhances edges in different regions of the fingerprint image. 3.2.1.3. Gaussian smoothing. The last preprocessing step in the first stage of enhancement process is Gaussian smoothing. Gaussian smoothing is computationally efficient, and gives higher significance to the pixel near the edges (ridges) which avoids blurring effect near the ridges. Therefore, Gaussian smoothing is appropriate to reduce the effect of noise caused by contrast enhancement using CLAHE. 3.2.2. Second stage enhancement Enhanced fingerprint images obtained from quality adaptive preprocessing in the first stage are fed to the second stage enhancement. Second stage enhancement employs well-known Gabor [12], STFT[16], and ODF[14] enhancement algorithms to further enhance the fingerprint images. The details of these algorithms can be found in Refs. [12,16,14]. 4. Experimental results In this section, experimental evaluations of the proposed method are presented. Section 4.1 provides a brief description of databases and experimental methodologies followed in this work. Section 4.2 presents results of the feature selection using One-way ANOVA statistical test. In Section 4.3, results of FQA phase are presented with a comparative study with some well-known FQA algorithms [51,38]. Results of FQE phase are reported with a comparative study with some well-known fingerprint enhancement techniques [12,16,14] in Section 4.4. 4.1. Database and experimental methodology In order to assess the efficacy of the proposed FQA and FQE methods, experiments are conducted on FVC 2004 database which contains fingerprint images of low and varying quality [52]. We

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

9

Table 1 Average quality scores of 100 good quality fingerprint images (Original and enhanced using parameter adaptive unsharp masking based QAP) given by NFIQ 2.0.

Average NFIQ 2.0 score

Original

QAP (R = 1, A = 2)

QAP (R = 1.5, A = 1.5)

QAP (R = 2, A = 1)

QAP (R = 2.5, A = 0.5)

62.86

64.34

67.21

69.43

67.13

have also conducted experiments on FVC 2002 database [53] to test the robustness of the proposed approach. The FVC 2004 database is acquired to give a challenging benchmark for state-of-the-art recognition algorithms than previous fingerprint verification competition databases [54]. FVC 2004 database consists of four data sets, namely, DB1, DB2, DB3 and DB4. Acquisition of fingerprints in DB1 to DB3 is done in different sessions by varying conditions to enforce the challenging image quality characteristics for recognition while DB4 consists of synthetically generated fingerprint images. The images of DB1 and DB2 data sets are acquired using an optical sensor, DB3 with the thermal sweeping sensor, and DB4 with SFinGe v3.0 sensor. Each data set consists of 100 fingers with 8 impressions which make each data set of 800 fingerprint images. Similar to FVC 2004 database, FVC 2002 database also contains 4 data sets (DB1, DB2, DB3 and DB4) of 800 images in each of them. The fingerprint images in FVC 2002 database are not acquired in varying image quality conditions. Fingerprint verification tests are performed as per FVC protocol [52] to ensure the comparability of results with Refs. [12,14,16]. To obtain genuine match rate, each sample of a subject in DB1 to DB4 data sets is compared with other samples. Hence, there are 8C2 × 100 = 2800 genuine comparisons. False match rate is obtained by comparing the first sample of a subject with the first sample of other remaining subjects. Therefore, total imposter comparisons are 100C2 = 4950. The detail description of FVC protocol and computation of equal error rate (EER) can be found in Ref. [55]. The EER points to a system’s operating point at which it correctly recognizes genuine and imposter user with equal probability. National Institute of Standard and Technology (NIST) biometric image software package (NBIS [56]) and Verifinger minutiae extractor with matcher are employed for fingerprint verification. In NBIS package, minutiae are extracted using MINDTCT package, and templates are matched using BOZORTH3 matcher. 4.2. Feature selection using one-way ANOVA statistical test One-way ANOVA statistical test is performed to select the most discriminative features which can differentiate the defined quality classes. Evaluations are performed on 200 fingerprint images of each quality class obtained from all four data sets (DB1 to DB4) of FVC 2004 database [52]. Feature vector F is obtained from fingerprint images of each quality class as given in Section 3.1.1. Results of one-way ANOVA with the probability value p for the individual features are shown in Table 2. The p-value obtained for each feature (fi ) is less than 0.05, therefore, null hypothesis is rejected for all features. Rejection of null hypothesis reflects that means of each feature for the defined quality classes are not same. This signifies that a considerable difference lies in all five defined quality classes which can be observed from the feature distribution obtained by one-way ANOVA for each quality class as shown in Fig. 5. Further, Tukey’s HSD test is performed to identify whether there is any significant difference between mean feature values for each pair (e.g. (D, ND) and (ND, G)) of the quality classes or not. Probability values p corresponding to each pair of quality classes for the individual features are reported in Table 3. It can be observed that the probability values (p) for some pairs of the quality class is more than the a value (p > 0.05). This indicates that mean feature values of these pairs are not significantly different from each other. Hence, these features are not well suited for quality clustering of fingerprint images, and they can be removed from feature vector F. Probability values (p) for (good-normal dry), (good, normal wet), and (normal dry, normal

wet) pairs of RLC is 0.9182, 0.6188, and 0.9781, respectively which are more than the significance level (a = 0.05). Therefore, the null hypothesis is accepted for these pairs as per RLC feature, stating that the mean value of RLC feature for fingerprint images in these pairs of classes is similar. It can also be verified in Fig. 5 (f) that RLC values for fingerprint images in these pairs of clusters fall in similar range. Therefore, the RLC feature is insignificant for discriminating different quality fingerprint images and thus, removed from F. Similarly, from Table 3, we can observe that p-values of RPS, Gabor, and Gabor-shen features are more than the significance level (a) for different quality pairs. Similar feature value distribution is the rationale behind high probability (p) between some pairs of quality classes. None of the other features have probability value (p) larger than a = 0.05 for any pair of quality classes as given in Table 3. Therefore, final feature vector Fs after feature selection contains 7 features namely, uniformity, contrast, mean, moisture, variance, RVAU, and RVU for quality clustering of fingerprint images. 4.3. Experimental results of FQA This section provides the experimental results of the FQA phase on FVC 2004 database. The fingerprint images of each data set (DB1 to DB4) are classified by three human experts into dry, normal dry, good, normal wet, and wet quality classes. The final decision regarding the visual quality of a fingerprint image is rendered by a majority based voting approach. An objective method is used to test the efficacy of the FQA phase. Fingerprint images classified by Fuzzy C-means clustering in different classes Qc are compared with the manually clustered data sets by human experts in Qc classes. Fig. 6 shows the fingerprint images of the different quality classes obtained using Fuzzy C-means clustering. The clustering error for each quality class Qc on FVC 2004 DB1 data set is reported in Table 4. Error rates for the individual quality class are determined as number of wrongly classified fingerprint images divided by total images in that quality cluster. Fuzzy C-means clusters 142 fingerprint images into the dry quality class, out of which 139 are predicted correctly and 3 (2 — normal dry, 1 — good) are predicted into the wrong quality classes. Hence, error in clustering of dry images is 3/142 × 100 = 2.11%. Similarly, number of wrongly clustered images for normal dry, good, normal wet, and wet class are 10 (good), 12 (4 — normal dry, 7 — normal wet, and 1 — wet), 11 (10 — good and 1 — wet), and 0 which give error rates of 4.95%, 8.16%, 6.79%, and 0% respectively. The overall error rate is 4.50% for DB1 data set of FVC 2004 database. Experimental results exhibit that error in the prediction of the correct quality class for dry and wet fingerprint images is very less as well as none of the dry images are classified Table 2 Probability value (p) for individual feature ANOVA test. Feature

Probability value (p)

Uniformity Contrast Mean Moisture Variance Ridge line count Ridge valley area uniformity Radial power spectrum Ridge valley Uniformity Gabor Gabor-shen

1.38e−41 4.49e−71 1.50e−78 7.40e−76 6.23e−60 2.54e−14 5.93e−42 1.38e−07 1.48e−43 2.61e−15 2.88e−15

10

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16 100 200

Contrast

Uniformity

80

60

150

40

100 Dry

Wet

Good

Normal Dy

Normal Wet

Dry

Wet

Clusters

Good

Normal Dry

Normal Wet

Normal Dry

Normal Wet

Normal Dry

Normal Wet

Normal Dry

Normal Wet

Normal Dry

Normal Wet

Clusters

(a)

(b) 0.2 0.1 0

Moisture

Mean

200 150

−0.1 −0.2 −0.3

100

−0.4

Dry

Wet

Good

Normal Dry

Dry

Normal Wet

Wet

Good Clusters

Clusters

(c)

(d) Ridge Line Count

100

Variance

80 60 40

1.6 1.4 1.2

20 Dry

Wet

Good

Normal Dry

Normal Wet

Dry

Wet

Clusters

(f) Radial Power Spectrum

Ridge Valley Area Uniformity

(e) 6 4 2 0 Dry

Wet

Good

Normal Dry

7600 7400 7200 7000 6800 6600 Dry

Normal Wet

Wet

Good Clusters

Clusters

(g)

(h) 0.6

1.5

Gabor

Ridge Valley Uniformity

Good Clusters

1

0.4

0.2 0.5 Dry

Wet

Good

Normal Dry

Dry

Normal Wet

Wet

Good Clusters

Clusters

(i)

(j)

Gabor Shen

0.9 0.8 0.7 0.6 Dry

Wet

Good

Normal Dry

Normal Wet

Clusters

(k) Fig. 5. One-way ANOVA test for individual features: (a) Uniformity, (b) Contrast, (c) Mean, (d) Moisture, (e) Variance, (f) Ridge line count, (g) Ridge valley area uniformity, (h) Radial power spectrum, (i) Ridge valley uniformity, (j) Gabor, and (k) Gabor-shen.

as wet or vice-versa. On the contrary, error rates for normal dry and normal wet images are little higher as good quality images contain both the normal dry and normal wet regions. It is also observed that the images which are not clustered correctly are assigned to quality classes which are nearer to the original quality class. As an instance of this observation, it can be seen in Table 4 that some of the images which are originally in wet quality class are clustered wrongly into the normal wet (1) and good (1) quality classes (see Table 4). None of the wet quality images (according to subjective clustering) are clustered into normal dry or dry classes. Similar observations can be seen for other quality clusters of fingerprint images.

Similar experiments are performed on DB2 to DB4 data sets of FVC 2004 to cluster the fingerprint images of these data sets into appropriate quality class. Results for these data sets are shown in Table 5, 6, and 7, respectively. Prediction of fingerprint images in different quality classes for DB2 are done with error rates of 0.47%, 1.42%, 5.40%, 8.75%, and 3.57% for dry, normal dry, good, normal wet, and wet classes, respectively (see Table 5). Overall error rate is 3.50% for FVC 2004 DB2 data set. FVC2004 DB3 data set is clustered into dry, normal dry, good, normal wet, and wet quality classes with error rates of 4.68%, 6.06%, 4.86%, 7.59%, and 1.56%, respectively which constitute overall error rate of 4.62% (see Table 6). Error rates for clustering FVC

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

11

Table 3 Probability value (p) corresponding to each possible pair of dry (D), wet (W), good (G), normal dry (ND), and normal wet (NW) cluster for individual features. Feature

(D, W)

(D, G)

(D, ND)

(D, NW)

(W, G)

(W, ND)

(W, NW)

(G, ND)

(G, NW)

(ND, NW)

Uniformity Contrast Mean Moisture Variance Ridge line count Ridge valley area uniformity Radial power spectrum Ridge valley uniformity Gabor Gabor-shen

0.0000 0.0000 0.0000 0.0000 0.0000 0.0012 0.0000 0.020 0.0000 0.0000 0.9923

0.0000 0.0000 0.0000 0.0000 0.0000 0.0002 0.0000 0.0000 0.0000 0.0000 0.0000

0.0000 0.0000 0.0000 0.0000 0.0000 0.0048 0.0456 0.0000 0.0117 0.0000 0.0000

0.0000 0.0000 0.0000 0.0000 0.0000 0.0282 0.0000 0.0001 0.0000 0.0000 0.0000

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0482 0.0000 0.0196 0.0000

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.4460 0.0000 1.0000 0.0000

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.5984 0.0000 0.0195 0.0000

0.0128 0.0016 0.0000 0.0000 0.0000 0.9182 0.0473 0.8233 0.0005 0.0173 0.3963

0.0031 0.0000 0.0000 0.0000 0.0000 0.6188 0.0463 0.6883 0.0027 1.0000 0.6619

0.0000 0.0000 0.0000 0.0000 0.0000 0.9781 0.0000 0.9993 0.0000 0.0172 0.9933

2004 DB4 data set into appropriate quality classes are 2.58%, 0.43%, 9.52%, 0.69%, and 11.62% for dry, normal dry, good, normal wet, and wet quality classes, respectively. Overall error rate is 4.12% for FVC 2004 DB4 data set. A comparative study of the FQA phase is done with the existing quality nature prediction methods presented by Munir et al. [38] and Yun et al. [51]. Munir et al. [38] have classified the fingerprint images into dry, wet, good, and normal quality classes and Yun et al. [51] have classified the fingerprint images into oily (wet), dry, and neutral (good) classes. In order to make the proposed method comparable with these methods, the five quality classes considered in our work are reduced to three quality classes by combining the dry and normal dry class to dry class, and wet and normal wet class to wet class, and the good class is termed as normal quality class. The four quality classes considered by Munir et al. methods are also reduced to three classes by combining the good quality class with normal quality class. Table 8 reports the comparative results in terms of error rates of the quality class prediction of fingerprint images. The error in prediction of each quality class is computed using Eq (18).

Error rate =

Number of wrong classified images Total number of fingerprint images in that class (18)

For example, the error rate in prediction of dry fingerprint images using the proposed method on the DB1 data set is obtained as 3.77% (13/344 × 100). Similarly, the error rates of other quality classes are obtained for all methods. Results reported in Table 8 indicates that the proposed method predicts the quality nature of fingerprint images more accurately as compared to the existing methods. These

classification results affirm that the features selected using the statistical analysis method are able to discriminate the fingerprint images of different qualities more accurately as compared to the features present in the existing works. 4.4. Experimental results of two-stage FQE The proposed fingerprint enhancement scheme comprises of two consecutive stages of enhancement as shown in Fig. 3. In the first stage, fingerprint images undergo through QAP to clear the ridgevalley pattern of fingerprints in low-quality (dry, normal dry, normal wet, and wet) images. Further, fingerprint images are enhanced with some well-known enhancement algorithms in the second stage of enhancement. Experimental results of the two-stage enhancement are reported in the following sections. 4.4.1. Quality adaptive preprocessing (QAP) After fingerprint quality assessment using Fuzzy C-means in the first phase, QAP of fingerprint images in each quality cluster Qc is performed. Original and enhanced images using QAP from each quality class are shown in Fig. 7. To assess the efficiency of the proposed QAP method, seven-dimensional feature vector (Fs ) is extracted from subjectively clustered 200 random fingerprint images of each quality class (Qc ) (5 × 200 = 1000 fingerprint images) from DB1 to DB4 data sets of FVC 2004. The feature vectors of these 1000 fingerprint images are fed as input to train a decision tree model for predicting the fingerprint quality class. Decision Tree (DT) is a supervised learning method used for classification problems. The goal of DT is to create a classification model that predicts the target variable or class by learning simple decision rules inferred from the feature

Fig. 6. Fingerprint images obtained in different clusters using Fuzzy C-means: (a) Dry (b) Wet (c) Normal Dry (d) Normal Wet (e) Good.

12

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

Table 4 Results of fingerprint quality clustering for FVC2004 DB1 database. Fuzzy C-means clustering

Subjective clustering

Quality clusters

Dry

Normal Dry

Good

Normal Wet

Wet

Total

Dry Normal Dry Good Normal Wet Wet Total Error rates

139 2 1 0 0 142 2.11%

0 192 10 0 0 202 4.95%

0 4 135 7 1 147 8.16%

0 0 10 151 1 162 6.79%

0 0 0 0 147 147 0%

139 198 156 158 149 800 4.50%

Table 5 Results of fingerprint quality clustering for FVC2004 DB2 database. Fuzzy C-means clustering

Subjective clustering

Quality clusters

Dry

Normal Dry

Good

Normal Wet

Wet

Total

Dry Normal Dry Good Normal Wet Wet Total Error rates

211 1 0 0 0 212 0.47%

1 207 2 0 0 210 1.42%

0 3 175 6 1 185 5.40%

0 1 9 125 2 137 8.75%

0 0 0 2 54 56 3.57%

212 212 186 133 57 800 3.50%

vectors. DT has an advantage that they can handle multi-class classification problems. Further, another 50 fingerprint images (different from 1000 fingerprint images used for training) from each of the quality cluster are selected to verify the improvement in the fingerprint quality. Quality class of these 50 fingerprint images of each class is predicted using the trained decision tree model. Tables 9 and 10 provide the fingerprint quality class prediction before and after the QAP, respectively. Results of the DT quality prediction before QAP in Table 9 show that there are 46 dry and 4 normal dry fingerprint images out of the 50 dry fingerprint images marked by human experts. After processing the fingerprint images with QAP, DT predicts 47 dry images into normal dry class while 3 images remains in dry class as given in Table 10. Therefore, it is evident that after QAP, quality of most of the fingerprint images in the dry class are improved, and they are moved to normal dry class. A similar improvement can be observed

for wet fingerprint images in Table 9, which are moved to normal wet (48) and good (2) quality classes after QAP as shown in Table 10. The quality of normal dry and normal wet fingerprint images are also improved significantly as fingerprint images from these classes are moved to good quality class after QAP. None of the good quality images are moved to any lower quality classes as good quality images are preprocessed with optimal parameter values in QAP. These results exhibit significant improvement in the fingerprint images of each quality class while the quality of good fingerprint images remains the same. 4.4.2. Gabor/STFT/ODF enhancement In order to validate that the proposed first stage QAP process is beneficial for the second stage Gabor/STFT/ODF enhancement algorithms, experiments with QAP as front-end are conducted. Therefore, QAP processed fingerprint images of FVC 2004 data sets are further

Table 6 Results of fingerprint quality clustering for FVC2004 DB3 database. Fuzzy C-means clustering

Subjective clustering

Quality clusters

Dry

Normal Dry

Good

Normal Wet

Wet

Total

Dry Normal Dry Good Normal Wet Wet Total Error rates

61 2 1 0 0 64 4.68%

2 186 10 0 0 198 6.06%

1 9 254 2 1 267 4.86%

0 0 5 73 1 79 7.59%

0 0 0 3 189 192 1.56%

64 197 270 78 191 800 4.62%

Table 7 Results of fingerprint quality clustering for FVC2004 DB4 database. Fuzzy C-means clustering

Subjective clustering

Quality clusters

Dry

Normal Dry

Good

Normal Wet

Wet

Total

Dry Normal Dry Good Normal Wet Wet Total Error rates

151 4 0 0 0 155 2.58%

0 227 1 0 0 228 0.43%

0 0 209 22 0 231 9.52%

0 0 1 142 0 143 0.69%

0 0 1 4 38 43 11.62%

151 231 212 168 38 800 4.12%

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16 Table 8 Comparative analysis of the FQA phase with Munir et al. [38] and Yun et al. [51] methods. Data sets

Quality

Proposed method

Munir et al. [38]

Yun et al. [51]

DB1

Dry Normal Wet Dry Normal Wet Dry Normal Wet Dry Normal Wet

3.77 8.16 3.55 0.71 5.40 7.25 5.72 4.86 3.32 1.30 9.52 3.22

4.94 10.32 4.67 1.89 6.11 9.73 7.89 5.18 5.26 1.89 9.89 4.57

5.15 9.65 5.32 1.26 5.93 8.79 6.43 5.74 5.87 2.13 10.61 5.13

DB2

DB3

DB4

enhanced with Gabor enhancement [12] using morphological segmentation [57], STFT enhancement [16], and ODF enhancement [14] with gradient based orientation estimation. The implementation of the ODF method [14] is made publicly available by the authors of the paper but this implementation doesn’t contain orientation estimation method used by the ODF method. Therefore, the results reported in this manuscript are evaluated using gradient based orientation estimation method. Use of different orientation estimation method causes a slight degradation in the performance of the ODF method. Fingerprint images of different qualities enhanced by QAP and without QAP based methods are shown in Fig. 8. It can be observed from Fig. 8 that fingerprint images enhanced with QAP based methods (columns 3, 5, and 7 of Fig. 8) show better visual quality as compared to the traditional Gabor/STFT/ODF enhancement methods (columns 2, 4, and 6 of Fig. 8). To relate the results of the proposed two-stage QAP based enhancement methods with earlier enhancement methods [12,14,16], EERs for NBIS matcher are compared in Table 11. Table 11 shows the EER for all four data sets of the FVC 2004 database. EERs presented for original images, Gabor enhancement, and STFT enhancement methods are cited from Ref. [26], and for ODF enhancement [14], EERs are obtained using gradient-based orientation estimation technique. Generally, it is assumed that enhancement techniques improve the recognition performance of the system. However, the performance can be degraded if the images are noisy. This can be seen in case of traditional Gabor enhancement and STFT enhancement methods as EER of the verification system is increased for them as compared to EER of the original fingerprint images. These techniques leads to lower EERs for all four data sets when combined with our proposed first stage QAP method. This improvement is clearly visible for the traditional Gabor, STFT, and ODF enhancement techniques where relative improvements are between 1.54% to 50.62%. The proposed

13

Table 9 Fingerprint quality class prediction before QAP. Decision tree classification

Subjective quality

Quality clusters

Dry

Normal Dry

Good

Normal Wet

Wet

Dry Normal Dry Good Normal Wet Wet

46 0 0 0 0

4 50 2 0 0

0 0 48 3 0

0 0 0 46 2

0 0 0 1 48

QAP based enhancement method improves EER up to 50% relative to the traditional Gabor/STFT/ODF enhancement methods [12,14,16]. No improvement in EERs for DB3 and DB4 data sets are observed when using only QAP relative to the original fingerprint images. Therefore, it is evident that using only QAP may not be enough to improve the performance for thermal sweeping sensor data set(DB3) and SFinGE v3.0 sensor (synthetic) data set(DB4). However, QAP combined with Gabor/STFT/ODF enhancement methods improves the performance for these data sets. The lowest EER of 9.12% (QAP + Gabor), 8.27% (QAP + STFT), and 5.40% (QAP + STFT) for DB1, DB2, and DB3 data sets, respectively are achieved when QAP is applied before enhancement. Further, the lowest EER of 7.30% for DB4 data set obtained for original images without any enhancement. The second lowest EER of 9.32% is achieved using the QAP + Gabor enhancement method for FVC 2004 DB4 data set. Furthermore, comparison of EERs for VeriFinger matcher is listed in Table 12 which shows that the QAP process improves the recognition performance for all data sets. EERs for the original fingerprint images are cited from Ref. [52]. Significant performance improvement in EERs for DB1, DB2, DB3, and DB4 data sets is observed when QAP based enhancement methods are used over the traditional methods of enhancement. Further, the lowest EER of 1.90% (QAP + STFT), 1.49% (QAP + STFT), 0.59% (QAP + ODF), and 0.68% (QAP + STFT) for DB1, DB2, DB3, and DB4 data sets are achieved for QAP based enhancement methods. Thus, it is evident that quality based preprocessing helps in improving the performance of different fingerprint matching algorithms. The efficiency and robustness of our proposed method is also evaluated on FVC 2002 database which contains four data sets namely, DB1, DB2, DB3, and DB4. The results obtained using the NBIS matcher are reported in Table 13. From the reported results, it is observed that the least EER for DB1, DB2, DB3, and DB4 data sets are 7.89% (QAP + STFT), 6.89% (QAP + ODF), 11.35% (QAP + STFT), and 5.93 (QAP + Gabor), respectively. All these results are obtained using QAP based preprocessing before Gabor/STFT/ODF enhancement. EERs obtained using Verifinger matcher are reported in Table 14. The Verifinger matcher achieves the lowest EER of 0.20% (QAP + ODF), 0.14% (QAP +

Fig. 7. Quality adaptive preprocessing of images from FVC2004 DB1 database: (a) Dry, (b) Normal dry, (c) Good, (d) Normal wet, and (e) Wet, where the columns correspond to 1) the original image, and 2) preprocessed image.

14

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

Table 10 Fingerprint quality class prediction after QAP. Decision tree classification

Subjective quality

Quality clusters

Dry

Normal Dry

Good

Normal Wet

Wet

Dry Normal Dry Good Normal Wet Wet

3 0 0 0 0

47 22 2 17 0

0 27 48 21 2

0 1 0 12 48

0 0 0

Gabor), 0.57% (QAP + STFT), and 0.16 (QAP) for DB1, DB2, DB3, and DB4 data sets, respectively. The results obtained using both NBIS and Verifinger matchers show that the QAP process improves the performances of Gabor/STFT/ODF enhancement algorithms. The majority of fingerprint images in FVC 2002 database are already of good quality. Therefore, due to the lack of low-quality fingerprint images (dry and wet), the relative improvement is not much for these data sets. Further, results obtained with the Verifinger matcher show that the QAP achieves better performance than the traditional Gabor/STFT/ODF enhancement techniques for DB1, DB2, and DB4 data sets. It is evident that for data sets containing the fingerprint images of good quality, a normal enhancement process such as QAP is better as compared to other well-known Gabor/STFT/ODF enhancement techniques. 5. Conclusion In this paper, a two-stage quality adaptive fingerprint enhancement method is proposed which processes fingerprint images

0

based on their quality characteristics. The two-stage fingerprint enhancement scheme is preceded by Fuzzy C-means clustering based FQA method to cluster fingerprint images into dry, normal dry, good, normal wet, and wet classes. Emphasizing the enhancement of fingerprint images of different qualities (Qc ), a QAP method is designed in the first stage which controls the unsharp mask filter parameters according to the quality (Qc ) of the fingerprint images. In the second stage, the enhanced fingerprint images of the first stage QAP method are further enhanced with Gabor, STFT, and ODF enhancement techniques. The experimental results show that the proposed two-stage enhancement using FQA can successfully handle various fingerprint quality contexts, and achieves better performance with the combination of Gabor/STFT/ODF enhancement algorithms for both NBIS and VeriFinger matching algorithms. The improved results for traditional enhancement techniques can be termed as a paradigm for future studies of quality based fingerprint enhancement. A future direction of this study can consider few other frequency and spatial domain features to improve FQA performance. Furthermore, the proposed FQA

Fig. 8. Enhancement results for different quality fingerprint images from FVC2004 DB1 database. (a) Dry, (b) Normal dry, (c) Good, (d) Normal wet, and (e) Wet, where the columns correspond to 1) the original image, 2) Gabor enhancement, 3) QAP + Gabor enhancement, 4) STFT enhancement, 5) QAP + STFT enhancement, 6) ODF enhancement, and 7) QAP + ODF enhancement.

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

15

Table 11 EER’s obtained using NBIS package (MINDTCT and BOZORTH3) for QAP and without QAP based enhancement methods on FVC 2004 database. NBIS matcher FVC 2004

DB1 DB2 DB3 DB4

FVC 2004

∗

DB1 DB2 DB3 DB4

Original images

QAP

Relative improvement

Gabor [12]

QAP + Gabor

Relative improvement

14.50% 9.50% 6.20% 7.30%

10.74% 9.31% 7.53% 9.95%

25.93% 2.00% −21.45% −36.30%

16.90% 14.40% 7.10% 9.80%

9.12% 8.41% 5.65% 9.32%

46.03% 41.59% 20.42% 4.89%

STFT [16]

QAP + STFT

Relative improvement

ODF* [14]

QAP + ODF*

Relative improvement

19.10% 11.90% 7.60% 10.90%

9.43% 8.27% 5.40% 9.55%

50.62% 30.50% 28.94% 12.38%

9.55% 9.06% 6.69% 9.98%

8.55% 8.92% 6.16% 9.57%

10.47% 1.54% 7.92% 4.10%

ODF with gradient based orientation estimation.

Table 12 EER’s obtained using VeriFinger matcher for QAP and without QAP based enhancement methods on FVC 2004 database. VeriFinger matcher

Original images

QAP

Relative improvement

Gabor [12]

QAP + Gabor

Relative improvement

FVC 2004

3.91% 3.62% 4.03% 1.40%

2.10% 1.93% 0.85% 0.71%

46.29% 46.68% 78.90% 49.28%

2.03% 1.87% 0.71% 0.79%

1.92% 1.83% 0.66% 0.77%

5.41% 2.13% 7.04% 2.53%

STFT [16]

QAP + STFT

Relative improvement

ODF* [14]

QAP + ODF*

Relative improvement

1.95% 1.58% 0.67% 0.70%

1.90% 1.49% 0.61% 0.68%

2.56% 5.69% 8.95% 2.85%

2.24% 1.56% 0.60% 1.25%

2.07% 1.53% 0.59% 1.21%

7.58% 1.92% 1.66% 3.20%

based QAP can be combined with other enhancement techniques for better results.

Acknowledgments

FVC 2004

∗

DB1 DB2 DB3 DB4

DB1 DB2 DB3 DB4

ODF with gradient based orientation estimation.

The authors are thankful to SERB (ECR/2017/000027), Department of Science and Technology, Government of India for providing financial support. Also, We would like to acknowledge Indian Institute of Technology Indore for providing the laboratory support and research facilities to carry out this research work.

Conflict of interest The authors have no potential conflict of interest to disclose.

Table 13 EERs obtained using NBIS package (MINDTCT and BOZORTH3) for QAP and without QAP based enhancement methods on FVC2002 database. NBIS matcher FVC 2002

DB1 DB2 DB3 DB4

FVC 2002

DB1 DB2 DB3 DB4

∗

Original images

QAP

Relative improvement

Gabor [12]

QAP + Gabor

Relative improvement

8.50% 7.51% 14.40% 6.75%

8.35% 7.32% 13.56% 6.29%

1.76% 2.52% 5.83% 6.81%

8.27% 7.23% 12.89% 6.25%

8.13% 7.03% 11.50% 5.93%

1.69% 2.76% 10.78% 5.12%

STFT [16]

QAP + STFT

Relative improvement

ODF* [14]

QAP + ODF*

Relative improvement

8.31% 7.18% 13.26% 6.37%

7.89% 6.93% 11.35% 5.97%

5.05% 3.48% 14.40% 6.27%

8.14% 7.17% 12.87% 6.38%

7.93% 6.89% 11.63% 6.13%

2.57% 3.90% 9.63% 3.91%

ODF with gradient based orientation estimation.

Table 14 EER’s obtained using VeriFinger matcher for QAP and without QAP based enhancement methods on FVC2002 database. VeriFinger matcher

Original images [53]

QAP

Relative improvement

Gabor [12]

QAP + Gabor

Relative improvement

FVC 2002

0.98% 0.52% 1.78% 0.68%

0.23% 0.21% 1.27% 0.16%

76.53% 59.61% 28.65% 76.47%

0.29% 0.21% 0.94% 0.35%

0.24% 0.14% 0.87% 0.32%

17.24% 33.33% 7.44% 8.67%

STFT [16]

QAP + STFT

Relative improvement

ODF* [14]

QAP + ODF*

Relative improvement

0.27% 0.18% 0.60% 0.26%

0.26% 0.16% 0.57% 0.24%

3.70% 11.11% 5.00% 7.69%

0.24% 0.20% 0.75% 0.31%

0.20% 0.17% 0.69% 0.31%

16.66% 17.64% 8.00% 0.00%

FVC 2002

∗

DB1 DB2 DB3 DB4

DB1 DB2 DB3 DB4

ODF with gradient based orientation estimation.

16

R.P. Sharma and S. Dey / Image and Vision Computing 83-84 (2019) 1–16

References [1] L.M. Seghetti, S.R. Viña, U.S. Visitor and Immigrant Status Indicator Technology Program (US-VISIT), CRS Report for Congress (2004) 1–36. [2] Govt. of India, Unique Identification Authority of India (UIDAI) “https://uidai.gov.in/”, Aadhaar (Targeted Delivery of Financial and Other Subsidies, Benefits and Services) Act, 2016, 2016. [3] B. Schouten, B. Jacobs, Biometrics and their use in e-passports, Image and Vision Computing 27 (3) (2009) 305–312. [4] K.A. Nagaty, An adaptive hybrid energy-based fingerprint matching technique, Image and Vision Computing 23 (5) (2005) 491–500. [5] A. Jin, D. Ling, O. Song, An efficient fingerprint verification system using integrated wavelet and Fourier–Mellin invariant transform, Image and Vision Computing 22 (6) (2004) 503–513. [6] N.K. Ratha, K. Karu, S. Chen, A.K. Jain, A real-time matching system for large fingerprint databases., IEEE Transactions on Pattern Analysis and Machine Intelligence 18 (8) (1996) 799–813. [7] S.J. Xie, J. Yang, H. Gong, S. Yoon, D.S. Park, Intelligent fingerprint quality analysis using online sequential extreme learning machine, Soft Computing 16 (9) (2012) 1555. [8] L. Ghiani, D.A. Yambay, V. Mura, G.L. Marcialis, F. Roli, S.A. Schuckers, Review of the fingerprint liveness detection (LivDet) competition series: 2009 to 2015, Image and Vision Computing 58 (2017) 110–128. [9] D. Maltoni, D. Maio, A.K. Jain, S. Prabhakar, Handbook of Fingerprint Recognition, Springer-Verlag New York Inc. 2009. [10] L. O’Gorman, J.V. Nickerson, An approach to fingerprint filter design, Pattern Recognition 22 (1) (1989) 29–38. [11] F. Turroni, R. Cappelli, D. Maltoni, Fingerprint enhancement using contextual iterative filtering, Proceedings of the 5th IAPR International Conference on Biometrics (ICB), 2012. pp. 152–157. [12] L. Hong, Y. Wan, A.K. Jain, Fingerprint image enhancement: algorithm and performance evaluation, IEEE Transactions on Pattern Analysis and Machine Intelligence 20 (8) (1998) 777–789. [13] C. Gottschlich, Curved-region-based ridge frequency estimation and curved Gabor filters for fingerprint image enhancement, IEEE Transactions on Image Processing 21 (4) (2012) 2220–2227. [14] C. Gottschlich, C.B. Schonlieb, Oriented diffusion filtering for enhancing low-quality fingerprint images, IET Biometrics 1 (2) (2012) 105–113. [15] W. Wang, J. Li, F. Huang, H. Feng, Design and implementation of Log-Gabor filter in fingerprint image enhancement, Pattern Recognition Letters 29 (3) (2008) 301–308. [16] S. Chikkerur, A.N. Cartwright, V. Govindaraju, Fingerprint enhancement using STFT analysis, Pattern Recognition 40 (1) (2007) 198–211. [17] G. Zhu, C. Zhang, A top-down fingerprint image enhancement method based on Fourier analysis, Proceedings of the 5th Chinese Conference on Advances in Biometric Person Authentication, 2004. pp. 439–448. [18] J. Yang, N. Xiong, A.V. Vasilakos, Two-stage enhancement scheme for low-quality fingerprint images by learning from the images, IEEE Transactions on Human-Machine Systems 43 (2) (2013) 235–248. [19] M. Ghafoor, I.A. Taj, W. Ahmad, N.M. Jafri, Efficient 2-fold contextual filtering approach for fingerprint enhancement, IET Image Processing 8 (7) (2014) 417–425. [20] S. Ding, W. Bian, T. Sun, Y. Xue, Fingerprint enhancement rooted in the spectra diffusion by the aid of the 2D adaptive Chebyshev band-pass filter with orientation-selective, Information Sciences 415-416 (2017) 233–246. [21] S. Ding, W. Bian, H. Liao, T. Sun, Y. Xue, Combining Gabor filtering and classification dictionaries learning for fingerprint enhancement, IET Biometrics 6 (6) (2017) 438–447. [22] E.K. Yun, S.B. Cho, Adaptive fingerprint image enhancement with fingerprint image quality analysis, Image and Vision Computing 24 (1) (2006) 101–110. [23] J.S. Bartunek, M. Nilsson, B. Sallberg, I. Claesson, Adaptive fingerprint image enhancement with emphasis on preprocessing of data, IEEE Transactions on Image Processing 22 (2) (2013) 644–656. [24] S. Greenberg, M. Aladjem, D. Kogan, Fingerprint image enhancement using filtering techniques, Real-Time Imaging 8 (3) (2002) 227–236. [25] Y. Hu, X. Jing, B. Zhang, X. Zhu, Low quality fingerprint image enhancement based on Gabor filter, Proceedings of the 2nd International Conference on Advanced Computer Control, 2, 2010. pp. 195–199. [26] H. Fronthaler, K. Kollreider, J. Bigun, Local features for enhancement and minutiae extraction in fingerprints, IEEE Transactions on Image Processing 17 (3) (2008) 354–363. [27] B.G. Sherlock, D.M. Monro, K. Millard, Fingerprint enhancement by directional Fourier filtering, IEEE Proceedings - Vision, Image and Signal Processing 141 (2) (1994) 87–94. [28] A. Willis, L. Myers, A cost-effective fingerprint recognition system for use with low-quality prints and damaged fingertips, Pattern Recognition 34 (2) (2001) 255–270. [29] C.T. Hsieh, E. Lai, Y.C. Wang, An effective algorithm for fingerprint image enhancement based on wavelet transform, Pattern Recognition 36 (2) (2003) 303–312. [30] S. Jirachaweng, V. Areekul, Fingerprint enhancement based on discrete cosine transform, Proceedings of the International Conference on Advances in Biometrics, 2007. pp. 96–105.

[31] P. Sutthiwichaiporn, V. Areekul, S. Jirachaweng, Iterative fingerprint enhancement with matched filtering and quality diffusion in spatial-frequency domain, Proceedings of the 20th International Conference on Pattern Recognition, 2010. pp. 1257–1260. [32] C. Dong, C.C. Loy, K. He, X. Tang, Learning a deep convolutional network for image super-resolution, in: D. Fleet, T. Pajdla, B. Schiele, T. Tuytelaars (Eds.), Computer Vision - ECCV 2014, 2014. pp. 184–199. [33] K. Cao, A.K. Jain, Latent orientation field estimation via convolutional neural network, 2015 International Conference on Biometrics (ICB), 2015. pp. 349–356. [34] J. Li, J. Feng, C.-C.J. Kuo, Deep convolutional neural network for latent fingerprint enhancement, Signal Processing: Image Communication 60 (2018) 52–63. [35] P. Schuch, Survey on the impact of fingerprint image enhancement, IET Biometrics 7 (2018) 102–115. [36] H. Hasan, S. Abdul-Kareem, Fingerprint image enhancement and recognition algorithms: a survey, Neural Computing and Applications 23 (6) (2013) 1605–1610. [37] E. Lim, K.A. Toh, P.N. Suganthan, X. Jiang, W.Y. Yau, Fingerprint image quality analysis, Proceedings of the International Conference on Image Processing, 2, 2004. pp. 1241–1244. Vol.2. [38] M.U. Munir, M.Y. Javed, S.A. Khan, A hierarchical k-means clustering based fingerprint quality classification, Neurocomputing 85 (Supplement C) (2012) 62–67. [39] E. Tabassi, C.L. Wilson, A novel approach to fingerprint image quality, Proceedings of the International Conference on Image Processing, 2, 2005. pp. 37–40. [40] R.P. Sharma, S. Dey, Fingerprint image quality assessment and scoring, Proceedings of the 5th International Conference on Mining Intelligence and Knowledge Exploration, 2017. pp. 1–10. [41] F. Alonso-Fernandez, J. Fierrez, J. Ortega-Garcia, J. Gonzalez-Rodriguez, H. Fronthaler, K. Kollreider, J. Bigun, A comparative study of fingerprint image-quality estimation methods, IEEE Transactions on Information Forensics and Security 2 (4) (2007) 734–743. [42] S. Bharadwaj, M. Vatsa, R. Singh, Biometric quality: a review of fingerprint, iris, and face, EURASIP Journal on Image and Video Processing 2014 (1) (2014) 1–34. [43] M.A. Olsen, M. Dusio, C. Busch, Fingerprint skin moisture impact on biometric performance, 3rd International Workshop on Biometrics and Forensics (IWBF 2015), 2015. pp. 1–6. [44] R.D. Labati, A. Genovese, E.M. Ballester, V. Piuri, F. Scotti, G. Sforza, Automatic classification of acquisition problems affecting fingerprint images in automated border controls, 2015 IEEE Symposium Series on Computational Intelligence, 2015. pp. 354–361. [45] P. Krishnasamy, S. Belongie, D. Kriegman, Wet fingerprint recognition: challenges and opportunities, 2011 International Joint Conference on Biometrics (IJCB), 2011. pp. 1–7. [46] M.S. Patil, S.S. Patil, Wet and dry fingerprint enhancement by using multi resolution technique, 2016 International Conference on Global Trends in Signal Processing, Information Computing and Communication (ICGTSPICC), 2016. pp. 188–193. [47] B.M. Mehtre, Fingerprint image analysis for automatic identification, Machine Vision and Applications 6 (2) (1993) 124–139. [48] N. Otsu, A threshold selection method from gray-level histograms, IEEE Transactions on Systems, Man and Cybernetics 9 (1) (1979) 62–66. [49] E. Tabassi, Development of NFIQ 2.0 NIST, Available at. (2016) https://www. nist.gov/services-resources/software/development-nfiq-20/,2016. [50] M.A. Olsen, V. Šmida, C. Busch, Finger image quality assessment features: definitions and evaluation, IET Biometrics 5 (2) (2016) 47–64. [51] E.-K. Yun, S.-B. Cho, Adaptive fingerprint image enhancement with fingerprint image quality analysis, Image and Vision Computing 24 (1) (2006) 101–110. [52] D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, A.K. Jain, FVC2004: Third Fingerprint Verification Competition, Proceedings of the 1st International Conference on Biometric Authentication, 2004. pp. 1–7. [53] D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, A.K. Jain, FVC2002: Second Fingerprint Verification Competition, Object Recognition Supported by User Interaction for Service Robots, 3, 2002. pp. 811–814. [54] R. Cappelli, D. Maio, D. Maltoni, J.L. Wayman, A.K. Jain, Performance evaluation of fingerprint verification systems, IEEE Transactions on Pattern Analysis and Machine Intelligence 28 (1) (2006) 3–18. [55] D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, A.K. Jain, FVC2000: Fingerprint Verification Competition, IEEE Transactions on Pattern Analysis and Machine Intelligence 24 (3) (2002) 402–412. [56] C.I. Watson, M.D. Garris, E. Tabassi, C.L. Wilson, R.M. McCabe, S. Janet, K. Ko, User’s guide to NIST Biometric Image Software (NBIS), National Institute of Standard and Technology (2007). [57] M.F. Fahmy, M.A. Thabet, A fingerprint segmentation technique based on Morphological processing, Proceedings of the IEEE International Symposium on Signal Processing and Information Technology, 2013. pp. 215–220.