- Email: [email protected]
Complex assessment of the quality of foodstuffs through the analysis of visual images, spectrophotometric and hyperspectral characteristics
16th IFAC Conference on Technology, Culture and International Stability 16th Culture International 16th IFAC IFAC Conference Conference on on Technology, Technology,Available Culture and and International online at www.sciencedirect.com 16th IFAC Conference Technology, Culture and International September 24-27, 2015.on Sozopol, Bulgaria Stability Stability Stability September September 24-27, 24-27, 2015. 2015. Sozopol, Sozopol, Bulgaria Bulgaria September 24-27, 2015. Sozopol, Bulgaria
ScienceDirect
IFAC-PapersOnLine 48-24 (2015) 060–065
Complex assessment of the quality of foodstuffs through the analysis of visual Complex assessment of the quality of foodstuffs through the analysis of visual Complex assessment of foodstuffs through the Compleximages, assessment of the the quality quality of ofand foodstuffs through the analysis analysis of of visual visual spectrophotometric hyperspectral characteristics images, spectrophotometric and hyperspectral characteristics images, spectrophotometric and hyperspectral characteristics images, spectrophotometric and hyperspectral characteristics Miroljub Mladenov*, Stanislav Penchev**, Martin Dejanov*** Miroljub Mladenov*, Mladenov*, Stanislav Penchev**, Martin Martin Dejanov*** Penchev**, Miroljub Miroljub Mladenov*, Stanislav Stanislav Penchev**, Martin Dejanov*** Dejanov***
* University of Rousse, 8 Studentska Str., 7017 Rousse, BULGARIA ** University of Rousse, Studentska Str., 7017 Rousse, BULGARIA (e-mail:8 uni-ruse.bg). 8 Studentska 7017 * University University of of Rousse, Rousse, 8 mladenov@ Studentska Str., Str., 7017 Rousse, Rousse, BULGARIA BULGARIA (e-mail: mladenov@ uni-ruse.bg). ** University of Rousse, 8 Studentska Str., 7017 Rousse, (e-mail: mladenov@ uni-ruse.bg). (e-mail: mladenov@ uni-ruse.bg). BULGARIA (e-mail: ** University of Rousse, 88 Studentska Str., 7017 Rousse, BULGARIA (e-mail: [email protected]) ** Str., ** University University of of Rousse, Rousse, 8 Studentska Studentska Str., 7017 7017 Rousse, Rousse, BULGARIA BULGARIA (e-mail: (e-mail: *** University of Rousse, [email protected]) 8 Studentska Str., 7017 Rousse, BULGARIA [email protected]) [email protected]) *** University of Rousse, Studentska Str., 7017 Rousse, BULGARIA (e-mail:8 *** 8 Studentska *** University University of of Rousse, Rousse, 8 [email protected]) Studentska Str., Str., 7017 7017 Rousse, Rousse, BULGARIA BULGARIA (e-mail: [email protected]) (e-mail: (e-mail: [email protected]) [email protected]) Abstract: This paper presents a new approach and a platform for complex, nondestructive, express Abstract: This paper presents presents new approach and a platform platform for complex, nondestructive, express evaluation of quality safety of food approach products based analysisfor of visual images, spectrophotometric Abstract: This paper aaa new and complex, nondestructive, express Abstract: This paperand presents new approach and aaon platform for complex, nondestructive, express evaluation of quality quality and safety safety of of food products products based on analysis analysis of visual visual images, spectrophotometric characteristics and hyperspectral images, followed by fusion the results of these analyzes with the aim of evaluation of and food based on of images, spectrophotometric evaluation of quality and safety of food products based on analysis of visual images, spectrophotometric characteristics and hyperspectral images, followed by fusion fusion the Within results of of these analyzes with the aim aim of of categorization of the investigated products in quality groups. the context of the problem, characteristics and hyperspectral images, followed by the results these analyzes with the characteristics and hyperspectral images, followed by fusion the results of these analyzes with the aimthe of categorization of the the includes investigated products in quality quality groups. Within thevisual context of the the problem, problem, the complex evaluation an assessment of the appearance and characteristics of categorization of investigated products in groups. Within the context of the categorization of the investigated products in quality groups. Within the context of the problem, the complex evaluation includes an assessment of the appearance appearance and visual ofcharacteristics characteristics of the investigated product, includes evaluationan features associated with the composition the product and complex evaluation of and of complex evaluation includes anofassessment assessment of the the appearance and visual visual characteristics of the the investigated product, evaluation of features associated with the composition of the product and the distribution of the features on itsof The problem with of thethe complex evaluation of product the investigated investigated product, evaluation features associated composition of and investigated product, evaluation ofsurface. features associated with the composition of the the product and the the distribution of the features on its surface. The problem of the complex evaluation of the investigated products is represented by two main tasks: 1) The formal description of the investigated objects by fusion distribution distribution of of the the features features on on its its surface. surface. The The problem problem of of the the complex complex evaluation evaluation of of the the investigated investigated products is represented represented by two main main tasks: 1) 1) The formal description of the investigated investigated objects byoffusion fusion the data is color images, spectral hyperspectral characteristics and 2) Evaluation data products by tasks: formal of objects products isfrom represented by two two main tasks:and 1) The The formal description description of the the investigated objects by by fusion the data from color images, spectral and hyperspectral characteristics and 2) Evaluation of data separability in the following two aspects: separability of individual areas (e.g. areas with meat, fat and the the data data from from color color images, images, spectral spectral and and hyperspectral hyperspectral characteristics characteristics and and 2) 2) Evaluation Evaluation of of data data separability inonthe thea certain following two aspects: storage separability of individual individual areas (e.g.theareas areas with meat, fat and and bone tissues) day of product and separability of data for same area on different separability in following two aspects: separability of areas (e.g. with meat, fat separability in the following two aspects: separability of individual areas (e.g. areas with meat, fat and bone tissues) on aa certain certain day day of of product product storage storage and and separability separability of of data data for for the the same same area area on on different different days of storage. bone tissues) on bone tissues) on a certain day of product storage and separability of data for the same area on different days of of storage. days days of storage. storage. © 2015, IFAC (International Federation of Automatic Control) Elsevier Ltd. All analyses rights reserved. Keywords: food products, quality and safety assessment, image,Hosting spectraby and hyperspectral . food products, quality and safety assessment, image, spectra and hyperspectral analyses . Keywords: food products, quality and safety assessment, image, spectra and hyperspectral analyses Keywords: Keywords: food products, quality and safety assessment, image, spectra and hyperspectral analyses ..
1. INTRODUCTION 1. INTRODUCTION 1. 1. INTRODUCTION INTRODUCTION Food quality and safety is a problem that is been engaged a Food quality and safety is and a problem that is been engaged number of famous people scientists ancient timesaaa Food quality and safety that is engaged Food quality and safety is is aa problem problem thatfrom is been been engaged number of famous people and scientists from ancient times till now. And this is not accidental. According to the World number number of of famous famous people people and and scientists scientists from from ancient ancient times times till now. And this is not accidental. According to the World Health Organization one of the key measures of the quality of till till now. now. And And this this is is not not accidental. accidental. According According to to the the World World Health Organization one offood. the key key measures of the the qualityfor of life is the quality of the The traditional methods Health Organization one of the measures of quality Health Organization one of the key measures of the quality of of life is the thethe quality of and the food. food. The traditional methods are: for assessing Quality Safety (QS) of food products life is quality of the The traditional methods for life is the quality of the food. The traditional methods for assessing the Qualitychemical and Safety of food products are: sensory and(QS) microbiological analysis. assessing the (QS) of are: assessingevaluation, the Quality Quality and and Safety Safety (QS) of food food products products are: sensory evaluation, chemical and microbiological analysis. They are not suitable for "on-line" monitoring as well as for sensory evaluation, chemical and microbiological analysis. sensory evaluation, chemical and microbiological analysis. They are not suitable for "on-line" monitoring as well as for express evaluation of food QS "on the ground" in stores, They are not suitable for "on-line" monitoring as well as They are not suitable for "on-line" monitoring as well as for for express evaluation of food QS "on the ground" in stores, warehouses, catering, etc.,"on where they are --not express of food the in stores, express evaluation evaluation of home, food QS QS "on the ground" ground" in always stores, warehouses, catering, home, etc., whereconditions. they are not always stored at regulated by the manufacturer warehouses, catering, home, etc., warehouses, catering, home, etc., where where they they are are not not always always stored at regulated by the manufacturer conditions. stored at regulated by the manufacturer conditions. stored at regulated by the manufacturer conditions. As an alternative to traditional methods, methods for express As an an alternativeevaluation to traditional traditional methods, methods for express non-destructive of food QS methods are morefor andexpress more As methods, As an alternative alternative to to traditional methods, methods for express non-destructive evaluation of food Among QS are are them more and and more widely applied in recent years. the most non-destructive evaluation of food QS more non-destructive evaluation of food QS are more and more more widely applied in recent recent optical years. methods Among based themonthe the most perspective are noncontact analysis widely applied in years. Among them widely applied in recent years. Among them the most most perspective are noncontact optical methods based on analysis of color images, spectrophotometric and hyperspectral perspective are noncontact optical methods based on analysis perspective are noncontact optical methods based on analysis of color images, spectrophotometric and hyperspectral analysis. of of color color images, images, spectrophotometric spectrophotometric and and hyperspectral hyperspectral analysis. analysis. analysis. The purpose of this study is to present a new approach, The this study is aa new methods and of tools complex, non-destructive, express The purpose purpose of this for study is to to present present new approach, approach, The purpose of this study is to present a new approach, methods and tools for complex, non-destructive, evaluation of quality and safety of widespread food products methods and and tools tools for for complex, complex, non-destructive, non-destructive, express express methods express evaluation of quality quality and safety safety of white widespread food products such as meat, structural bacon, brinedfood cheese and evaluation of and of widespread products evaluation of quality and safety of widespread food products such as cheese meat, structural structural bacon, white ofbrined brined cheese and yellow based on analysis visual images, such as meat, bacon, white cheese such as meat, structural bacon, white brined cheese and and yellow cheese based based on analysis analysis of visual visual images, spectrophotometric characteristics and hyperspectral yellow cheese on of yellow cheese based on analysis of visual images, spectrophotometric characteristics and hyperspectral followed by fusion the results of these analyses with images, the aim spectrophotometric characteristics and images, spectrophotometric characteristics and hyperspectral hyperspectral images, followed by fusion the results of these analyses with the aim followed by fusion the results of these analyses with followed by fusion the results of these analyses with the the aim aim
of categorization of the investigated products in quality of categorization of of the investigated investigated products in in quality groups. of of categorization categorization of the the investigated products products in quality quality groups. groups. groups. Two main tasks are discussed in the papers: 1) The formal Two main tasks tasks are discussed products in the the papers: 1) The formal description of theare investigated fusing1) theThe dataformal from Two discussed Two main main tasks are discussed in in the papers: papers: 1) The formal description of the investigated products fusing the data from color images, spectral and hyperspectral characteristics and description of the investigated products fusing the data description of the investigated products fusing the data from from color images, spectral and hyperspectral characteristics and 2) Evaluation of separability of data obtained from individual color images, spectral and hyperspectral characteristics color images, spectral and hyperspectral characteristics and and 2) Evaluation of separability of data obtained from individual areas (e.g. areas meat, fat tissues) a certain 2) of separability of data obtained from 2) Evaluation Evaluation of with separability of and databone obtained fromonindividual individual areas (e.g. areas with meat, and boneof tissues) on aa certain day product and fat separability data for same areas areas with fat and on areasof(e.g. (e.g. areasstorage with meat, meat, fat and bone bone tissues) tissues) on the a certain certain day of product storage and separability of data for the same area on different days of storage. day day of of product product storage storage and and separability separability of of data data for for the the same same area on different days of storage. area area on on different different days days of of storage. storage. 2. STATE-OF-THE-ART 2. STATE-OF-THE-ART STATE-OF-THE-ART 2. 2. STATE-OF-THE-ART 2.1. Systems for color image analysis. 2.1. Systems Systems for color color image analysis. analysis. 2.1. 2.1. Systems for for color image image analysis. In the dairy products industry Computer Vision Systems In the dairy products industry Computer Vision Systems (CVS) are mainly evaluation color In products industry Computer Vision In the the dairy dairy products utilized industry for Computer Visionof:Systems Systems (CVS) are mainly utilized for evaluation of: color characteristics and texture features of cheese and cheese (CVS) are mainly utilized for evaluation of: (CVS) are mainly utilized for evaluation of: color color characteristics and texture features of cheese and cheese melting (Sun, 2012), degreasing control 2004), characteristics and features of cheese characteristics and texture texture features of (Wang&Sun, cheese and and cheese cheese melting (Sun, 2012), degreasing control 2004), to determine the distribution the (Wang&Sun, amount of spices, melting (Sun, degreasing control (Wang&Sun, 2004), melting (Sun, 2012), 2012), degreasingand control (Wang&Sun, 2004), to determine the distribution and the amount of spices, vegetables and other components (Fornal et al., 2007), for to to determine determine the the distribution distribution and and the the amount amount of of spices, spices, vegetables and other components (Fornal et al., 2007), for predicting the moisture content and evaluation of fat content vegetables vegetables and and other other components components (Fornal (Fornal et et al., al., 2007), 2007), for for predicting the moisture moisture content and evaluation evaluation of fat fatdetecting content of the cheese (Wang & Sun, 2004), for predicting the content and of content predicting the moisture content and evaluation of fat content of the cheese cheese (Wang (Wang && & Sun, Sun,2010) 2004), for detecting microorganisms and for otherdetecting features of the of the cheese (Kumar (Wang &Mittal, Sun, 2004), 2004), for detecting microorganisms (Kumar & Mittal, 2010) and other features of cheeses and dairy products (Wang & Sun, 2001). microorganisms (Kumar & Mittal, 2010) and other microorganisms (Kumar & Mittal, 2010) and other features features of cheeses and dairy products (Wang & Sun, 2001). of cheeses and dairy products (Wang & Sun, 2001). of cheeses and dairy products (Wang & Sun, 2001). Computer vision is a fast growing and useful alternative to Computer vision isofaa meat fast growing useful alternative to expert assessment and meatand products. By CVS and Computer vision and useful to Computer vision is is a fast fast growing growing and useful alternative alternative to expert assessment of meat and meat products. By CVS and expert assessment assessment of of meat meat and and meat meat products. products. By By CVS CVS and and expert
2405-8963 © 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Copyright © IFAC 2015 60 Peer review under responsibility of International Federation of Automatic Control. Copyright © IFAC 2015 60 Copyright © IFAC 2015 60 10.1016/j.ifacol.2015.12.057 Copyright © IFAC 2015 60
IFAC TECIS 2015 September 24-27, 2015. Sozopol, Bulgaria Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065
statistical modelling key attributes of fresh meat and processed meat products, related to their QS, can be extracted and evaluated. This approach has proved its relevance in a series of assessments of the characteristics of the products in the visible and infrared region of the spectrum (Jackman et al., 2011). Basic features of meat and meat products that may be determined by image analysis are: appearance of pork and veal (Fagan et al., 2007), color characteristics of the meat tissue and areas of fat in fresh meat (Jackman et al., 2011), the fat content of pork and veal pieces (Chmielet et al. 2011), defects in fresh meat and detection of microorganisms in the composition of the meat (Ziadi et al., 2010), determining the contents of the pieces of muscle tissue forming the ham (Jia & Schinckel, 1992), determining the texture of the surface (Tan, 2008).
61
different components, such as proteins, nitrogens, water content, dry matter, etc., the microbiological composition, water activity, the freshness of the product and other (Pu et al., 2014); determining the microbiological composition (Tao et al., 2012) determination of leanness (Govindarajan et al., 2008), etc. 3. OBJECTS OF STUDY AND CHARACTERISTICS EVALUATED Main objects of the study are widespread food products such as meat, structural bacon, white brined cheese and yellow cheese from cow's milk, in abnormal storage conditions (at 20ºC and a lack of illumination), different from those covered by the manufacturer. Such conditions occurred in Bulgaria this winter, when hundreds of small settlements were left without electric power supply for a long period of time.
2.2 Spectral analysis systems.
The main evaluated characteristics of the investigated objects are the following:
The near infrared spectroscopy (NIRS) is a non-destructive technology, which is mainly used for determining the composition of a variety of dairy products such as milk (Tsenkova et al., 2000) and cheese (Fagan et al., 2007) as well as evaluation of major QS indicators of these products. Some typical examples of the application of NIRS analysis for assessment of various features associated with QS of dairy products may be indicated: determining the sensory features and age of cheese (Fagan et al., 2007); determining the composition of cheese (Růžičková & Šustová 2006); the composition of cow's milk (Tsenkova et al., 2000); of moisture, fat, and inorganic salts in the processed cheese, analysis and prediction of maturity and sensory features of the cheddar cheese (Fagan et al., 2007), etc.
- For white brined cheese and yellow cheese: surface color characteristics and its change during storage; appearance of colonies of mold, fungi and yeasts; acid degree ° T and its change during storage. - For pork meat and structural bacon: surface color characteristics, water content and acid degree °T and its change during storage. 4. MATERIALS AND METHOD 4.1. Formal description of the investigated objects. Representation by weighted features.
Main features of the meat and meat products, which can be determined by analysis of the spectral characteristics, are related to: determination of color and surface texture of meat (Andrés et al., 2007) active acidity of the meat (pH) (Prieto et al., 2009); determination of leanness (Ripoll et al., 2008); determining the content of fat, protein, water content and dry matter (Ripoll et al., 2008); determining the microbiological composition (Ellis et al., 2002); determining the freshness of meat (Horváth et al., 2008); determining the content of water, salt and the water activity of ham, sausages (Valous et al., 2010), etc.
Within the frames of the investigation the main formal description of the investigated objects is based on a vector of features of the form: t
X m1 x1 , m 2 x 2 , m 3 x 3 , , m n x n
(1)
where the components xi provide essential physical and chemical characteristics of the object related to its quality and safety, while mi are weight coefficients, related to each of the characteristics. In summary, the features xi can be presented in the following three groups: - Characteristics associated with visible features and derived from color images of the products examined;
2.3. Hyperspectral analysis systems. Hyperspectral analysis systems (HSA) found relatively few applications in solving various tasks related to the assessment of the QS of dairy products. Published analyses are related to the possibilities of defining the content and distribution of fat and protein, casein, lactose, to identify the type of dairy product, the presence of foreign fat and other.
- Characteristics associated with visible features and the composition derived from spectral characteristics of the products examined; - Characteristics associated with the distribution of features on the surface of the investigated objects derived from hyperspectral images.
Hyperspectral analysis systems have found wide application in non-destructive analysis and evaluation of QS of meat and meat products. Published studies are related to the determination of: color characteristics, surface texture and active acid (El Маsry et al., 2012); the composition of the meat and meat products (Pu et al., 2014), the contents of the
Why to use this kind of formal presentation of the investigated objects? The main idea is the features xi to retain its physical nature, to present not mathematical abstractions, but certain characteristics of objects which are understandable to the experts. Furthermore, such description 61
IFAC TECIS 2015 62 Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065 September 24-27, 2015. Sozopol, Bulgaria
Option 3. The value of mi is determined depending on the inaccuracy εi of the measurement/determination of the corresponding feature:
provides a comprehensive presentation of objects through a variety of characteristics, visible, and related to their composition, concerning both specific individual areas and the entire surface of the object studied. This kind of formal representation allows individual features to be ranked and weighted, depending on the context of the problem.
m i 1
- Larger values of mi will determine a stronger presence of the component in the formal description of the object, i.e. the relevant feature will have greater significance for solving task. The suitable choice of the weights of the features may increase or reduce the significance/impact of certain features over the formal description of the investigated objects, and therefore the outcome of categorization;
4.2. Assessment of the separability of data classes. Separability of data for different areas (e.g. areas with meat, fat and bone tissue in pieces of meat) on a certain day of the storage of the product, and for the same area on different days of storage of the product is a major criterion for the correct classification both with regard to composition of the product and of its freshness. The separability between the data of the investigated areas was quantitatively assessed by the overlap error εpr% (the ratio of incorrectly classified examples to the total number of examples). The separability determination was made by two classifiers: LDA (Linear Discriminate Analysis), which implements linear separability, and kernel version of SVM (Support Vector Machines), which satisfies the conditions for linear separability of the classes of data.
- Arranging the components of the vector descriptions allows reducing the impact of certain features, whose values change significantly within a given category and are measured/determined with great uncertainty. How the weights of the weighted features can be defined? Several options for selecting the values of the weighting coefficients mi are proposed: Option 1. The value of mi is determined depending on how the change of corresponding feature influence to the sensitivity of the description to the change of that feature. It is measured by the change in position (distance Dxp) of the sample in relation to the prototype (the formal description of class category) in the feature space of the samples. Quantitatively this relationship can be represented by the expression: D xp max
Quality classes in both cases are set by prototypes (points) in the feature space of the features. The categorization is done by association of the assessed sample to the nearest prototype (Mladenov, 2015).
(2)
where Dxpmax represents the maximum possible deviation of the position of the individual sample X from the prototype due to the change of feature. Option 2. The value of mi is determined depending on the dispersion/accuracy of the features within the category. It is natural to assume that the features with lower accuracy (higher dispersion values) need to be involved with a smaller weighting in the description of the object, i.e., the size of the weighting factor must be inversely proportional to the variance or mean square deviation si: m is 1
1 k s s in
(4)
The choice of the option for determination the weights of the features depends primarily on the specific conditions, under which the problem for the categorization of the investigated products is solved. If the product characteristics are amended in a relatively wide range, Option 2 is more appropriate. If the product characteristics are measured/determined by a relatively large error, it is more appropriate to use Option 3 for determination of the weights mi.
- When solving similar tasks a man intuitively separates essential from non-essential details, i.e. he ranks the characteristics of an object by their importance;
D xp
in
where kε is a factor determining the sensitivity of mi to εi and εin is a value of εi normalized to the maximum value of ε.
Why weighted features, rather than equivalent features? Few considerations that support the idea of weighing the features xi will be given:
m iD k D
1 k
The separability of the data is examined in two variants: 1. When the data for the features derived from the spectral characteristics of the entire spectral range of the spectrophotometer is used, and 2. When using the data for the features, derived from the selected frequency bands, of which the spectral characteristic of the device is divided. 4.3. Description of the system, involving CVS and spectral analysis system. Formation of hyperspectral characteristics.
(3)
where ks is a factor determining the sensitivity of mi to si and sin is a value of si normalized to the maximum value of s.
A system for forming and analysis of visual images, spectral and hyperspectral characteristics is developed at the University of Ruse "A. Kanchev" by a team from the Department "Automation and Mechatronics". The system is presented in Fig. 1. What are the main characteristics of the system? 1. It can form color images of the investigated objects, the spectral characteristics of the diffuse reflection in a small area
62
IFAC TECIS 2015 September 24-27, 2015. Sozopol, Bulgaria Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065
of the object and hyperspectral images in separate small areas (pixels) in lines of pixels as well as in the plane of the object. This is performed using XY motorized scanning stage (5) equipped with a controller (6). The investigated object is placed over the stage. By controlling the motion of the stage the spectrophotometer probe (3) could be placed over individual pixels, separate important object areas or the whole object surface could be scanned.
63
- Cheese: areas with and without colonies of molds and yeasts; - Yellow cheese: areas with and without colonies of mold; - Meat: areas with meat, fat and bone tissues; - Bacon: areas with meat and fat tissues. For each of the characteristics, presented in Chapter 3, the change of the area in different days of storage of the product is examined.
1
5.1. Assessment of area separability using image analysis. In the framework of this study, an attempt was made to separate the above-mentioned fields, by: 1. Image binarization by moving from color image to grayscale image and from it to a binary image with a given threshold T, and 2. Direct binarization by setting a threshold for color using the HSI color model. For both methods, the error in determining the area of individual areas reached significant levels (tens of percent), which is not acceptable even for objects with significantly changing color characteristics. Color characteristics, even for one area, for example the area of meat tissue in pieces of meat or bacon, vary widely, and depend on a number of factors. These are for example the age of the animal on whether the meat is normal or is pale, soft, watery, the duration of storage, even the geographical area in which the animal is kept. Color characteristics of areas with colonies of fungi, yeast and mold in dairy products also vary considerably.
6 3
4
2 5
Unacceptably large errors for separation of the investigated areas were received as for the identification of various areas on a certain day of the storage of the product, and in assessing data separability for the same area on different days of storage.
Fig. 1. Hyperspectral imaging system based on point scanning: 1 – RGBcamera DFK 31AU03; 2 – QE65000 spectrophotometer; 3 – spectrophotometer probe; 4 – illumination system; 5 – 8MTF-102LS05 XY motorized scanning stage; 6 – 8SMC4-USBhF stage motion controller.
5.2. Investigation of areas separability using spectral characteristics analysis across the overall spectral range of the instrument.
2. The second specific feature is related to the way hyperspectral characteristics of individual pixels are obtained. The spectrophotometer forms spectral response across the overall spectral range of the instrument in each pixel. After that the neighbouring points of the spectral characteristic are aggregated together, in order to obtain hyperspectral characteristic, consisting of various non-overlapping spectral bands. The number of frequency bands (frequency ranges Δλi) can be set in advance, or the minimum required number of bands can be formed, based on a certain criterion, which within this study is related to the separability of the data classes (linear or non-linear separability, achieved respectively by LDA and kernel SVM classifiers). The results presented in section 5.3 refer to the frequency band for which the best data separability is achieved.
The possibility for separation of the spectral data for different areas, presented with the first three principal components will be illustrated by one of the investigated objects: structural bacon with meat and fat tissues. To extract the features of spectral characteristics and to reduce the dimensionality of the spectral data, PCA method is used, where the number of principal components ranged from 3 to 10. The study was conducted in VIS (visible) and NIR (near infra red) spectral ranges. Table 1 present the overlap errors εpr% for data classes, described by principal components, obtained from the spectral characteristics of areas with meat and fat tissues in overall spectral range of the spectrophotometers. The data separability for a same area in different days of storage is investigated. 40 samples of bacon are analyzed, where for every area spectral characteristics in 3 different points are measured.
5. RESULTS FOR SEPARABILITY ASSESSMENT, BASED ON IMAGE ANALYSIS, SPECTRAL AND HYPERSPECTRAL ANALYSIS.
The investigation of class separability by empirical data for VIS characteristics from the overall spectral range shows, that the average error values for meat tissues in bacon vary between 7.4% and 45.4%, while for fat tissue – between
The possibility of separation of the following areas of the investigated products, for a particular day of storage, is examined: 63
IFAC TECIS 2015 64 Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065 September 24-27, 2015. Sozopol, Bulgaria
2.1% and 48.9%. The investigation of class separability by empirical data for NIR characteristics from the overall spectral range shows, that the average error values for meat tissues in bacon vary between 0 and 34.9%, while for fat tissues – from 0 to 47%.
objects is denoted with n, while N denotes the number of the spectral band with the best separability. Table 3. Class separability for bacon in a particular day of storage: cl1 – meat tissues, cl2 – fat tissues
Table 1. Overlap errors for bacon in two different days Day 1vs2 2vs3 3vs4 4vs5 5vs6 6vs7 1vs3 3vs5 5vs7
Error value, εpr% VIS NIR Meat tissues Fat tissues Meat tissues Fat tissues 7.37 2.11 24.21 8.42 17.34 44.48 5.618 2.13 44.56 44.56 4.49 2.25 45.35 48.87 0 0 36.14 38.37 0 0 39.14 40.96 34.94 46.99 16.84 22.11 1.05 1.053 30.34 41.57 0 0 40.24 43.37 0 0
The results for overlap error examination during the investigation of data separability for different areas on a same day of storage are similar.
3 4 7
Classifier Day
n
0 3 6
28 3 43
1. The analysis of color images of the investigated objects gives the most incorrect results when evaluating separability of different areas from the analysed objects. This is due to the fact that the color characteristics of the different areas vary widely and depend on a number of factors that are not related to the conditions of the experiment. However, this method is most suitable for the evaluation of quantitative characteristics of the regions under assessment, for example, their size, which is one of the important quantitative characteristics of products quality.
Table 2. Overlap errors for bacon in two different days
1vs2 2vs3 3vs4 4vs5 5vs6 6vs7 1vs3 3vs5 5vs7
0 3 6
Based on the investigations and the results obtained the following main conclusions can be made:
The overlap errors εpr% for data classes of principal components, obtained from the hyperspectral characteristics of meat and fat tissues in bacon are presented in Table 2. The data separability for a same area in different days of storage is investigated.
2. The investigation of class separability using empirical data in the overall spectral range of the device show, that the overlap error for the two classes of spectral characteristics for bacon vary from several percent to about 49% for visible range and to about 47% for near infrared range. The results for the overlap errors when the separability of the data for different regions on a same day of storage is investigated, are similar.
Error value, εpr% VIS Meat tissues Fat tissues min max min max 0 0.06 0 0 0 0.04 0 0.40 0 0.14 0.13 0.24 0.05 0.17 0.14 0.44 0.25 0.45 0.23 0.38 0.10 0.32 0.07 0.41 0.02 0.20 0.01 0.20 0.03 0.28 0.03 0.35 0.30 0.47 0.26 0.45
n
VIS spectral range LDA SVM-К N εpr% n N εpr % cl1 vs cl2 3 0-0.1 6 6 0-0.06 4 0-0.16 3 3 0-0.12 7 0-0.14 20 9 0-0.12 NIR spectral range LDA SVM-К N εpr% n N εpr% cl1 vs cl2 15 0.05-0.15 55 30 0.08-0.16 3 0.03-0.1 35 9 0.02-0.1 11 0.02=0.12 22 7 0.06-0.13
6. CONCLUSIONS
5.3. Investigation of areas separability using data analysis in selected frequency bands of the hyperspectral characteristics.
Day
Classifier Day
NIR Meat tissues Fat tissues min max min max 0.12 0.26 0.02 0.11 0 0.02 0 0.12 0.01 0.09 0.01 0.09 0 0.00 0 0 0 0.00 0 0 0.24 0.39 0.25 0.43 0 0.012 0 0.011 0 0 0 0 0 0 0 0
3. The investigation of class separability using empirical data from selected frequency bands show, that the maximum overlap error for the two classes of hyperspectral characteristics does not exceed 0.47% for visible range and 0.43% for near infrared range. The results for the overlap errors when the separability of the data for different regions on a same day of storage is investigated are similar.
The investigation of class separability by empirical data for VIS characteristics in selected frequency band shows, that the average error values for meat tissues in bacon vary between 0 and 0.47%, while for fat tissue – between 0 and 0.45%. The investigation of class separability by empirical data for NIR characteristics in selected frequency band shows, that the average error values for meat tissues in bacon vary between 0 and 0.39%, while for fat tissues – from 0 to 0.43%.
4. There is a significant difference between mean values of errors obtained on the basis of spectral data for the entire spectral range and errors obtained using spectral data of the selected frequency bands. The difference in the classification accuracy is two orders of magnitude. These very low values of the overlap errors for data classes from hyperspectral characteristics are received due to the fact that the class separability is accessed in a narrow spectral band, instead of in the entire frequency band of the device. This is perhaps
Table 3 presents the overlap errors εpr%, obtained during the investigation of the data separability for different areas in a same day of storage. The minimal number of frequency bands in hyperspectral characteristics of the investigated 64
IFAC TECIS 2015 September 24-27, 2015. Sozopol, Bulgaria Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065
one of the most important advantages of the hyperspectral analysis to classic spectral analysis.
65
quality assessment of fresh meats, Trends in Food Science & Technology, Vol. 22, Iss. 4, pp. 185–197. Jia J., Schinckel A. P (1992). Automatic measurement of ham composition using image processing. Proceedings of SPIE: The International Society for Optical Engineering, 1778, pp. 128-134. Kumar, S., G. S. Mittal (2010). Rapid detection of microorganisms using image processing parameters and neural network, Food and Bioprocess Technology, vol. 3(5), pp. 741-751. Mladenov, M. Complex assessment of food products quality using color images, spectrophotometric and hyperspectral characteristics analyses. Monography. Publishing center of Ruse University “A. Kanchev”, Ruse, 2015 (in Bulgarian). Prieto, N., Roehe, R., Lavin, P (2009). Application of near infrared reflectance spectroscopy to predict meat and meat products quality: A review. Meat Science, 83, 2, pp175-186. Pu, H., Da-Wen Sun, J. Ma, D. Liu and M. Kamruzzaman (2014), Hierarchical Variable Selection for Predicting Chemical Constituents in Lamb Meats Using Hyperspectral Imaging, Journal of Food Engineering, 143, pp. 44-52. Ripoll G., Alberti P. Panea B., Olleta JL., Sanudo C (2008). Near infrared reflectance spectroscopy for predicting chemical, instrumental and sensoru quality of beef. Meat Science, 80, pp. 697 – 702. Růžičková, J., K. Šustová (2006). Determination of selected parameters of quality of the dairy products by NIR spectroscopy. Czech J. Food Sci., 24: 255–260. Tan J (2008). Meat quality evaluation by computer vision, Journal of Food Engineering, 61, pp.27–35. Tao, F. F., Y. K. Peng, Y. Y. Li, K. L. Chao, and S. Dhakal (2012). Simultaneous determination of tenderness and Escheric hia coli contamination of pork using hyperspectral scattering technique, Meat Science, vol. 90(3), pp. 851–857. Tsenkova, R., Atanassova, S, Itoh, K., Ozaki, Y., Toyoda, K (2000). Near infrared spectroscopy for biomonitoring. Cow milk composition measurements in a spectral region from 1,100 to 2,400 nanometers. J Anim Sci, 78, pp. 515-522. Valous N.A., Fernando Mendoza, Da-Wen Sun (2010), Emerging non-contact imaging, spectroscopic and colorimetric technologies for quality evaluation and control of hams: a review, Trends in Food Science & Technology, Vol 21, Iss. 1, p26–43. Wang, H.H., Sun, D.W (2001). Evaluation of the functional features of cheddar cheese using a computer vision method. J. Food Eng., 49, pp. 49-53. Ziadi, A., X. Maldague, and L. Saucier (2010). Image analysis in computer vision: A high level means for NonDestructive evaluation of marbling in beef meat. In 10th International Conference on Quantitative Infra Red Thermography.
REFERENCES Andrés, S., I. Murray, E. A. Navajas, A. V. Fisher, N. R. Lambe, and L. Bünger (2007). Prediction of sensory characteristics of lamb meat samples by near infrared reflectance spectroscopy. Meat Science, vol. 76, no. 3, pp. 509–516. Chmiel M., M. Slowinski, K. Dasiewicz (2011), Lightness of the color measured by computer image analysis as a factor for assessing the quality of pork meat, Meat Science, Vol. 88, Issue 3, pp. 566–570. Da-Wen Sun (2008), Computer Vision Technology for Food Quality Evaluation, Academic Press, Elsevier London, UK. Da-Wen Sun (2012), Computer Vision Technology in the Food and Beverage Industries, Woodhead Publishing Limited, Cambridge, UK. El Barbri, N., A. Halimi, K. Rhofir (2008). A Nondestructive Method Based on an Artificial Vision for Beef Meat Quality Assesement. Food Science and Technology, pp. 111–138. El Маsry, G., D. F. Barbin, D. W. Sun, and P. Allen (2012). Meat quality evaluation by hyperspectral imaging technique: An overview, Critical Review in Food Science and Nutrition, vol. 52(8), pp. 689-711. Ellis, D. I., D. Broadhurst, D. B. Kell, J. J. Rowland, and R. Goodacre (2002). Rapid and quantitative detection of the microbial spoilage of meat by fourier transform infrared spectroscopy and machine learning, Applied and Environmental Microbiology, vol. 68, no. 6, pp. 2822– 2828. Fagan, C. C., O'Donnell, C. P., O'Callaghan, D. J., Downey, G., Sheehan, E. M., Delahunty, C. M., Everard, C., Guinee, T. P. and Howard, V (2007). Application of mid-infrared spectroscopy to the prediction of maturity and sensory texture attributes of cheddar cheese. Journal of Food Science, vol.72 (3), pp. 130-137. Fornal, J., T. Jeliński, Cheng Jin Du, Da Wen Sun (2007). Inspection of the distribution and amount of ingredients in pasteurized cheese by computer vision. Journal of Food Engineering, Vol. 83, Iss. 1, pp. 3–9. Govindarajan Konda Naganathan, Lauren M. Grimes, Jeyamkondan Subbiah, Chris R. Calkins, Ashok Samal, George E. Meyer (2008). Visible/near-infrared hyperspectral imaging for beef tenderness prediction. Computers and Electronics in Agriculture, Vol. 64, Iss. 2, pp. 225–233. Hai-Hong Wang, Da-Wen Sun (2004). Evaluation of the oiling off property of cheese with computer vision: Correlation with fat ring test. Journal of Food Engineering, Vol. 61. Horváth, K., Z. Seregély, É. Andrássy, I. Dalmadi, and J. Farkas (2008). A preliminary study using near infrared spectroscopy to evaluate freshness and detect spoilage in sliced pork meat, Acta Alimentaria, vol. 37, no. 1, pp. 93–102. Jackman P, Da-Wen Sun, Paul Allen (2011), Recent advances in the use of computer vision technology in the 65
ScienceDirect
IFAC-PapersOnLine 48-24 (2015) 060–065
Complex assessment of the quality of foodstuffs through the analysis of visual Complex assessment of the quality of foodstuffs through the analysis of visual Complex assessment of foodstuffs through the Compleximages, assessment of the the quality quality of ofand foodstuffs through the analysis analysis of of visual visual spectrophotometric hyperspectral characteristics images, spectrophotometric and hyperspectral characteristics images, spectrophotometric and hyperspectral characteristics images, spectrophotometric and hyperspectral characteristics Miroljub Mladenov*, Stanislav Penchev**, Martin Dejanov*** Miroljub Mladenov*, Mladenov*, Stanislav Penchev**, Martin Martin Dejanov*** Penchev**, Miroljub Miroljub Mladenov*, Stanislav Stanislav Penchev**, Martin Dejanov*** Dejanov***
* University of Rousse, 8 Studentska Str., 7017 Rousse, BULGARIA ** University of Rousse, Studentska Str., 7017 Rousse, BULGARIA (e-mail:8 uni-ruse.bg). 8 Studentska 7017 * University University of of Rousse, Rousse, 8 mladenov@ Studentska Str., Str., 7017 Rousse, Rousse, BULGARIA BULGARIA (e-mail: mladenov@ uni-ruse.bg). ** University of Rousse, 8 Studentska Str., 7017 Rousse, (e-mail: mladenov@ uni-ruse.bg). (e-mail: mladenov@ uni-ruse.bg). BULGARIA (e-mail: ** University of Rousse, 88 Studentska Str., 7017 Rousse, BULGARIA (e-mail: [email protected]) ** Str., ** University University of of Rousse, Rousse, 8 Studentska Studentska Str., 7017 7017 Rousse, Rousse, BULGARIA BULGARIA (e-mail: (e-mail: *** University of Rousse, [email protected]) 8 Studentska Str., 7017 Rousse, BULGARIA [email protected]) [email protected]) *** University of Rousse, Studentska Str., 7017 Rousse, BULGARIA (e-mail:8 *** 8 Studentska *** University University of of Rousse, Rousse, 8 [email protected]) Studentska Str., Str., 7017 7017 Rousse, Rousse, BULGARIA BULGARIA (e-mail: [email protected]) (e-mail: (e-mail: [email protected]) [email protected]) Abstract: This paper presents a new approach and a platform for complex, nondestructive, express Abstract: This paper presents presents new approach and a platform platform for complex, nondestructive, express evaluation of quality safety of food approach products based analysisfor of visual images, spectrophotometric Abstract: This paper aaa new and complex, nondestructive, express Abstract: This paperand presents new approach and aaon platform for complex, nondestructive, express evaluation of quality quality and safety safety of of food products products based on analysis analysis of visual visual images, spectrophotometric characteristics and hyperspectral images, followed by fusion the results of these analyzes with the aim of evaluation of and food based on of images, spectrophotometric evaluation of quality and safety of food products based on analysis of visual images, spectrophotometric characteristics and hyperspectral images, followed by fusion fusion the Within results of of these analyzes with the aim aim of of categorization of the investigated products in quality groups. the context of the problem, characteristics and hyperspectral images, followed by the results these analyzes with the characteristics and hyperspectral images, followed by fusion the results of these analyzes with the aimthe of categorization of the the includes investigated products in quality quality groups. Within thevisual context of the the problem, problem, the complex evaluation an assessment of the appearance and characteristics of categorization of investigated products in groups. Within the context of the categorization of the investigated products in quality groups. Within the context of the problem, the complex evaluation includes an assessment of the appearance appearance and visual ofcharacteristics characteristics of the investigated product, includes evaluationan features associated with the composition the product and complex evaluation of and of complex evaluation includes anofassessment assessment of the the appearance and visual visual characteristics of the the investigated product, evaluation of features associated with the composition of the product and the distribution of the features on itsof The problem with of thethe complex evaluation of product the investigated investigated product, evaluation features associated composition of and investigated product, evaluation ofsurface. features associated with the composition of the the product and the the distribution of the features on its surface. The problem of the complex evaluation of the investigated products is represented by two main tasks: 1) The formal description of the investigated objects by fusion distribution distribution of of the the features features on on its its surface. surface. The The problem problem of of the the complex complex evaluation evaluation of of the the investigated investigated products is represented represented by two main main tasks: 1) 1) The formal description of the investigated investigated objects byoffusion fusion the data is color images, spectral hyperspectral characteristics and 2) Evaluation data products by tasks: formal of objects products isfrom represented by two two main tasks:and 1) The The formal description description of the the investigated objects by by fusion the data from color images, spectral and hyperspectral characteristics and 2) Evaluation of data separability in the following two aspects: separability of individual areas (e.g. areas with meat, fat and the the data data from from color color images, images, spectral spectral and and hyperspectral hyperspectral characteristics characteristics and and 2) 2) Evaluation Evaluation of of data data separability inonthe thea certain following two aspects: storage separability of individual individual areas (e.g.theareas areas with meat, fat and and bone tissues) day of product and separability of data for same area on different separability in following two aspects: separability of areas (e.g. with meat, fat separability in the following two aspects: separability of individual areas (e.g. areas with meat, fat and bone tissues) on aa certain certain day day of of product product storage storage and and separability separability of of data data for for the the same same area area on on different different days of storage. bone tissues) on bone tissues) on a certain day of product storage and separability of data for the same area on different days of of storage. days days of storage. storage. © 2015, IFAC (International Federation of Automatic Control) Elsevier Ltd. All analyses rights reserved. Keywords: food products, quality and safety assessment, image,Hosting spectraby and hyperspectral . food products, quality and safety assessment, image, spectra and hyperspectral analyses . Keywords: food products, quality and safety assessment, image, spectra and hyperspectral analyses Keywords: Keywords: food products, quality and safety assessment, image, spectra and hyperspectral analyses ..
1. INTRODUCTION 1. INTRODUCTION 1. 1. INTRODUCTION INTRODUCTION Food quality and safety is a problem that is been engaged a Food quality and safety is and a problem that is been engaged number of famous people scientists ancient timesaaa Food quality and safety that is engaged Food quality and safety is is aa problem problem thatfrom is been been engaged number of famous people and scientists from ancient times till now. And this is not accidental. According to the World number number of of famous famous people people and and scientists scientists from from ancient ancient times times till now. And this is not accidental. According to the World Health Organization one of the key measures of the quality of till till now. now. And And this this is is not not accidental. accidental. According According to to the the World World Health Organization one offood. the key key measures of the the qualityfor of life is the quality of the The traditional methods Health Organization one of the measures of quality Health Organization one of the key measures of the quality of of life is the thethe quality of and the food. food. The traditional methods are: for assessing Quality Safety (QS) of food products life is quality of the The traditional methods for life is the quality of the food. The traditional methods for assessing the Qualitychemical and Safety of food products are: sensory and(QS) microbiological analysis. assessing the (QS) of are: assessingevaluation, the Quality Quality and and Safety Safety (QS) of food food products products are: sensory evaluation, chemical and microbiological analysis. They are not suitable for "on-line" monitoring as well as for sensory evaluation, chemical and microbiological analysis. sensory evaluation, chemical and microbiological analysis. They are not suitable for "on-line" monitoring as well as for express evaluation of food QS "on the ground" in stores, They are not suitable for "on-line" monitoring as well as They are not suitable for "on-line" monitoring as well as for for express evaluation of food QS "on the ground" in stores, warehouses, catering, etc.,"on where they are --not express of food the in stores, express evaluation evaluation of home, food QS QS "on the ground" ground" in always stores, warehouses, catering, home, etc., whereconditions. they are not always stored at regulated by the manufacturer warehouses, catering, home, etc., warehouses, catering, home, etc., where where they they are are not not always always stored at regulated by the manufacturer conditions. stored at regulated by the manufacturer conditions. stored at regulated by the manufacturer conditions. As an alternative to traditional methods, methods for express As an an alternativeevaluation to traditional traditional methods, methods for express non-destructive of food QS methods are morefor andexpress more As methods, As an alternative alternative to to traditional methods, methods for express non-destructive evaluation of food Among QS are are them more and and more widely applied in recent years. the most non-destructive evaluation of food QS more non-destructive evaluation of food QS are more and more more widely applied in recent recent optical years. methods Among based themonthe the most perspective are noncontact analysis widely applied in years. Among them widely applied in recent years. Among them the most most perspective are noncontact optical methods based on analysis of color images, spectrophotometric and hyperspectral perspective are noncontact optical methods based on analysis perspective are noncontact optical methods based on analysis of color images, spectrophotometric and hyperspectral analysis. of of color color images, images, spectrophotometric spectrophotometric and and hyperspectral hyperspectral analysis. analysis. analysis. The purpose of this study is to present a new approach, The this study is aa new methods and of tools complex, non-destructive, express The purpose purpose of this for study is to to present present new approach, approach, The purpose of this study is to present a new approach, methods and tools for complex, non-destructive, evaluation of quality and safety of widespread food products methods and and tools tools for for complex, complex, non-destructive, non-destructive, express express methods express evaluation of quality quality and safety safety of white widespread food products such as meat, structural bacon, brinedfood cheese and evaluation of and of widespread products evaluation of quality and safety of widespread food products such as cheese meat, structural structural bacon, white ofbrined brined cheese and yellow based on analysis visual images, such as meat, bacon, white cheese such as meat, structural bacon, white brined cheese and and yellow cheese based based on analysis analysis of visual visual images, spectrophotometric characteristics and hyperspectral yellow cheese on of yellow cheese based on analysis of visual images, spectrophotometric characteristics and hyperspectral followed by fusion the results of these analyses with images, the aim spectrophotometric characteristics and images, spectrophotometric characteristics and hyperspectral hyperspectral images, followed by fusion the results of these analyses with the aim followed by fusion the results of these analyses with followed by fusion the results of these analyses with the the aim aim
of categorization of the investigated products in quality of categorization of of the investigated investigated products in in quality groups. of of categorization categorization of the the investigated products products in quality quality groups. groups. groups. Two main tasks are discussed in the papers: 1) The formal Two main tasks tasks are discussed products in the the papers: 1) The formal description of theare investigated fusing1) theThe dataformal from Two discussed Two main main tasks are discussed in in the papers: papers: 1) The formal description of the investigated products fusing the data from color images, spectral and hyperspectral characteristics and description of the investigated products fusing the data description of the investigated products fusing the data from from color images, spectral and hyperspectral characteristics and 2) Evaluation of separability of data obtained from individual color images, spectral and hyperspectral characteristics color images, spectral and hyperspectral characteristics and and 2) Evaluation of separability of data obtained from individual areas (e.g. areas meat, fat tissues) a certain 2) of separability of data obtained from 2) Evaluation Evaluation of with separability of and databone obtained fromonindividual individual areas (e.g. areas with meat, and boneof tissues) on aa certain day product and fat separability data for same areas areas with fat and on areasof(e.g. (e.g. areasstorage with meat, meat, fat and bone bone tissues) tissues) on the a certain certain day of product storage and separability of data for the same area on different days of storage. day day of of product product storage storage and and separability separability of of data data for for the the same same area on different days of storage. area area on on different different days days of of storage. storage. 2. STATE-OF-THE-ART 2. STATE-OF-THE-ART STATE-OF-THE-ART 2. 2. STATE-OF-THE-ART 2.1. Systems for color image analysis. 2.1. Systems Systems for color color image analysis. analysis. 2.1. 2.1. Systems for for color image image analysis. In the dairy products industry Computer Vision Systems In the dairy products industry Computer Vision Systems (CVS) are mainly evaluation color In products industry Computer Vision In the the dairy dairy products utilized industry for Computer Visionof:Systems Systems (CVS) are mainly utilized for evaluation of: color characteristics and texture features of cheese and cheese (CVS) are mainly utilized for evaluation of: (CVS) are mainly utilized for evaluation of: color color characteristics and texture features of cheese and cheese melting (Sun, 2012), degreasing control 2004), characteristics and features of cheese characteristics and texture texture features of (Wang&Sun, cheese and and cheese cheese melting (Sun, 2012), degreasing control 2004), to determine the distribution the (Wang&Sun, amount of spices, melting (Sun, degreasing control (Wang&Sun, 2004), melting (Sun, 2012), 2012), degreasingand control (Wang&Sun, 2004), to determine the distribution and the amount of spices, vegetables and other components (Fornal et al., 2007), for to to determine determine the the distribution distribution and and the the amount amount of of spices, spices, vegetables and other components (Fornal et al., 2007), for predicting the moisture content and evaluation of fat content vegetables vegetables and and other other components components (Fornal (Fornal et et al., al., 2007), 2007), for for predicting the moisture moisture content and evaluation evaluation of fat fatdetecting content of the cheese (Wang & Sun, 2004), for predicting the content and of content predicting the moisture content and evaluation of fat content of the cheese cheese (Wang (Wang && & Sun, Sun,2010) 2004), for detecting microorganisms and for otherdetecting features of the of the cheese (Kumar (Wang &Mittal, Sun, 2004), 2004), for detecting microorganisms (Kumar & Mittal, 2010) and other features of cheeses and dairy products (Wang & Sun, 2001). microorganisms (Kumar & Mittal, 2010) and other microorganisms (Kumar & Mittal, 2010) and other features features of cheeses and dairy products (Wang & Sun, 2001). of cheeses and dairy products (Wang & Sun, 2001). of cheeses and dairy products (Wang & Sun, 2001). Computer vision is a fast growing and useful alternative to Computer vision isofaa meat fast growing useful alternative to expert assessment and meatand products. By CVS and Computer vision and useful to Computer vision is is a fast fast growing growing and useful alternative alternative to expert assessment of meat and meat products. By CVS and expert assessment assessment of of meat meat and and meat meat products. products. By By CVS CVS and and expert
2405-8963 © 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Copyright © IFAC 2015 60 Peer review under responsibility of International Federation of Automatic Control. Copyright © IFAC 2015 60 Copyright © IFAC 2015 60 10.1016/j.ifacol.2015.12.057 Copyright © IFAC 2015 60
IFAC TECIS 2015 September 24-27, 2015. Sozopol, Bulgaria Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065
statistical modelling key attributes of fresh meat and processed meat products, related to their QS, can be extracted and evaluated. This approach has proved its relevance in a series of assessments of the characteristics of the products in the visible and infrared region of the spectrum (Jackman et al., 2011). Basic features of meat and meat products that may be determined by image analysis are: appearance of pork and veal (Fagan et al., 2007), color characteristics of the meat tissue and areas of fat in fresh meat (Jackman et al., 2011), the fat content of pork and veal pieces (Chmielet et al. 2011), defects in fresh meat and detection of microorganisms in the composition of the meat (Ziadi et al., 2010), determining the contents of the pieces of muscle tissue forming the ham (Jia & Schinckel, 1992), determining the texture of the surface (Tan, 2008).
61
different components, such as proteins, nitrogens, water content, dry matter, etc., the microbiological composition, water activity, the freshness of the product and other (Pu et al., 2014); determining the microbiological composition (Tao et al., 2012) determination of leanness (Govindarajan et al., 2008), etc. 3. OBJECTS OF STUDY AND CHARACTERISTICS EVALUATED Main objects of the study are widespread food products such as meat, structural bacon, white brined cheese and yellow cheese from cow's milk, in abnormal storage conditions (at 20ºC and a lack of illumination), different from those covered by the manufacturer. Such conditions occurred in Bulgaria this winter, when hundreds of small settlements were left without electric power supply for a long period of time.
2.2 Spectral analysis systems.
The main evaluated characteristics of the investigated objects are the following:
The near infrared spectroscopy (NIRS) is a non-destructive technology, which is mainly used for determining the composition of a variety of dairy products such as milk (Tsenkova et al., 2000) and cheese (Fagan et al., 2007) as well as evaluation of major QS indicators of these products. Some typical examples of the application of NIRS analysis for assessment of various features associated with QS of dairy products may be indicated: determining the sensory features and age of cheese (Fagan et al., 2007); determining the composition of cheese (Růžičková & Šustová 2006); the composition of cow's milk (Tsenkova et al., 2000); of moisture, fat, and inorganic salts in the processed cheese, analysis and prediction of maturity and sensory features of the cheddar cheese (Fagan et al., 2007), etc.
- For white brined cheese and yellow cheese: surface color characteristics and its change during storage; appearance of colonies of mold, fungi and yeasts; acid degree ° T and its change during storage. - For pork meat and structural bacon: surface color characteristics, water content and acid degree °T and its change during storage. 4. MATERIALS AND METHOD 4.1. Formal description of the investigated objects. Representation by weighted features.
Main features of the meat and meat products, which can be determined by analysis of the spectral characteristics, are related to: determination of color and surface texture of meat (Andrés et al., 2007) active acidity of the meat (pH) (Prieto et al., 2009); determination of leanness (Ripoll et al., 2008); determining the content of fat, protein, water content and dry matter (Ripoll et al., 2008); determining the microbiological composition (Ellis et al., 2002); determining the freshness of meat (Horváth et al., 2008); determining the content of water, salt and the water activity of ham, sausages (Valous et al., 2010), etc.
Within the frames of the investigation the main formal description of the investigated objects is based on a vector of features of the form: t
X m1 x1 , m 2 x 2 , m 3 x 3 , , m n x n
(1)
where the components xi provide essential physical and chemical characteristics of the object related to its quality and safety, while mi are weight coefficients, related to each of the characteristics. In summary, the features xi can be presented in the following three groups: - Characteristics associated with visible features and derived from color images of the products examined;
2.3. Hyperspectral analysis systems. Hyperspectral analysis systems (HSA) found relatively few applications in solving various tasks related to the assessment of the QS of dairy products. Published analyses are related to the possibilities of defining the content and distribution of fat and protein, casein, lactose, to identify the type of dairy product, the presence of foreign fat and other.
- Characteristics associated with visible features and the composition derived from spectral characteristics of the products examined; - Characteristics associated with the distribution of features on the surface of the investigated objects derived from hyperspectral images.
Hyperspectral analysis systems have found wide application in non-destructive analysis and evaluation of QS of meat and meat products. Published studies are related to the determination of: color characteristics, surface texture and active acid (El Маsry et al., 2012); the composition of the meat and meat products (Pu et al., 2014), the contents of the
Why to use this kind of formal presentation of the investigated objects? The main idea is the features xi to retain its physical nature, to present not mathematical abstractions, but certain characteristics of objects which are understandable to the experts. Furthermore, such description 61
IFAC TECIS 2015 62 Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065 September 24-27, 2015. Sozopol, Bulgaria
Option 3. The value of mi is determined depending on the inaccuracy εi of the measurement/determination of the corresponding feature:
provides a comprehensive presentation of objects through a variety of characteristics, visible, and related to their composition, concerning both specific individual areas and the entire surface of the object studied. This kind of formal representation allows individual features to be ranked and weighted, depending on the context of the problem.
m i 1
- Larger values of mi will determine a stronger presence of the component in the formal description of the object, i.e. the relevant feature will have greater significance for solving task. The suitable choice of the weights of the features may increase or reduce the significance/impact of certain features over the formal description of the investigated objects, and therefore the outcome of categorization;
4.2. Assessment of the separability of data classes. Separability of data for different areas (e.g. areas with meat, fat and bone tissue in pieces of meat) on a certain day of the storage of the product, and for the same area on different days of storage of the product is a major criterion for the correct classification both with regard to composition of the product and of its freshness. The separability between the data of the investigated areas was quantitatively assessed by the overlap error εpr% (the ratio of incorrectly classified examples to the total number of examples). The separability determination was made by two classifiers: LDA (Linear Discriminate Analysis), which implements linear separability, and kernel version of SVM (Support Vector Machines), which satisfies the conditions for linear separability of the classes of data.
- Arranging the components of the vector descriptions allows reducing the impact of certain features, whose values change significantly within a given category and are measured/determined with great uncertainty. How the weights of the weighted features can be defined? Several options for selecting the values of the weighting coefficients mi are proposed: Option 1. The value of mi is determined depending on how the change of corresponding feature influence to the sensitivity of the description to the change of that feature. It is measured by the change in position (distance Dxp) of the sample in relation to the prototype (the formal description of class category) in the feature space of the samples. Quantitatively this relationship can be represented by the expression: D xp max
Quality classes in both cases are set by prototypes (points) in the feature space of the features. The categorization is done by association of the assessed sample to the nearest prototype (Mladenov, 2015).
(2)
where Dxpmax represents the maximum possible deviation of the position of the individual sample X from the prototype due to the change of feature. Option 2. The value of mi is determined depending on the dispersion/accuracy of the features within the category. It is natural to assume that the features with lower accuracy (higher dispersion values) need to be involved with a smaller weighting in the description of the object, i.e., the size of the weighting factor must be inversely proportional to the variance or mean square deviation si: m is 1
1 k s s in
(4)
The choice of the option for determination the weights of the features depends primarily on the specific conditions, under which the problem for the categorization of the investigated products is solved. If the product characteristics are amended in a relatively wide range, Option 2 is more appropriate. If the product characteristics are measured/determined by a relatively large error, it is more appropriate to use Option 3 for determination of the weights mi.
- When solving similar tasks a man intuitively separates essential from non-essential details, i.e. he ranks the characteristics of an object by their importance;
D xp
in
where kε is a factor determining the sensitivity of mi to εi and εin is a value of εi normalized to the maximum value of ε.
Why weighted features, rather than equivalent features? Few considerations that support the idea of weighing the features xi will be given:
m iD k D
1 k
The separability of the data is examined in two variants: 1. When the data for the features derived from the spectral characteristics of the entire spectral range of the spectrophotometer is used, and 2. When using the data for the features, derived from the selected frequency bands, of which the spectral characteristic of the device is divided. 4.3. Description of the system, involving CVS and spectral analysis system. Formation of hyperspectral characteristics.
(3)
where ks is a factor determining the sensitivity of mi to si and sin is a value of si normalized to the maximum value of s.
A system for forming and analysis of visual images, spectral and hyperspectral characteristics is developed at the University of Ruse "A. Kanchev" by a team from the Department "Automation and Mechatronics". The system is presented in Fig. 1. What are the main characteristics of the system? 1. It can form color images of the investigated objects, the spectral characteristics of the diffuse reflection in a small area
62
IFAC TECIS 2015 September 24-27, 2015. Sozopol, Bulgaria Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065
of the object and hyperspectral images in separate small areas (pixels) in lines of pixels as well as in the plane of the object. This is performed using XY motorized scanning stage (5) equipped with a controller (6). The investigated object is placed over the stage. By controlling the motion of the stage the spectrophotometer probe (3) could be placed over individual pixels, separate important object areas or the whole object surface could be scanned.
63
- Cheese: areas with and without colonies of molds and yeasts; - Yellow cheese: areas with and without colonies of mold; - Meat: areas with meat, fat and bone tissues; - Bacon: areas with meat and fat tissues. For each of the characteristics, presented in Chapter 3, the change of the area in different days of storage of the product is examined.
1
5.1. Assessment of area separability using image analysis. In the framework of this study, an attempt was made to separate the above-mentioned fields, by: 1. Image binarization by moving from color image to grayscale image and from it to a binary image with a given threshold T, and 2. Direct binarization by setting a threshold for color using the HSI color model. For both methods, the error in determining the area of individual areas reached significant levels (tens of percent), which is not acceptable even for objects with significantly changing color characteristics. Color characteristics, even for one area, for example the area of meat tissue in pieces of meat or bacon, vary widely, and depend on a number of factors. These are for example the age of the animal on whether the meat is normal or is pale, soft, watery, the duration of storage, even the geographical area in which the animal is kept. Color characteristics of areas with colonies of fungi, yeast and mold in dairy products also vary considerably.
6 3
4
2 5
Unacceptably large errors for separation of the investigated areas were received as for the identification of various areas on a certain day of the storage of the product, and in assessing data separability for the same area on different days of storage.
Fig. 1. Hyperspectral imaging system based on point scanning: 1 – RGBcamera DFK 31AU03; 2 – QE65000 spectrophotometer; 3 – spectrophotometer probe; 4 – illumination system; 5 – 8MTF-102LS05 XY motorized scanning stage; 6 – 8SMC4-USBhF stage motion controller.
5.2. Investigation of areas separability using spectral characteristics analysis across the overall spectral range of the instrument.
2. The second specific feature is related to the way hyperspectral characteristics of individual pixels are obtained. The spectrophotometer forms spectral response across the overall spectral range of the instrument in each pixel. After that the neighbouring points of the spectral characteristic are aggregated together, in order to obtain hyperspectral characteristic, consisting of various non-overlapping spectral bands. The number of frequency bands (frequency ranges Δλi) can be set in advance, or the minimum required number of bands can be formed, based on a certain criterion, which within this study is related to the separability of the data classes (linear or non-linear separability, achieved respectively by LDA and kernel SVM classifiers). The results presented in section 5.3 refer to the frequency band for which the best data separability is achieved.
The possibility for separation of the spectral data for different areas, presented with the first three principal components will be illustrated by one of the investigated objects: structural bacon with meat and fat tissues. To extract the features of spectral characteristics and to reduce the dimensionality of the spectral data, PCA method is used, where the number of principal components ranged from 3 to 10. The study was conducted in VIS (visible) and NIR (near infra red) spectral ranges. Table 1 present the overlap errors εpr% for data classes, described by principal components, obtained from the spectral characteristics of areas with meat and fat tissues in overall spectral range of the spectrophotometers. The data separability for a same area in different days of storage is investigated. 40 samples of bacon are analyzed, where for every area spectral characteristics in 3 different points are measured.
5. RESULTS FOR SEPARABILITY ASSESSMENT, BASED ON IMAGE ANALYSIS, SPECTRAL AND HYPERSPECTRAL ANALYSIS.
The investigation of class separability by empirical data for VIS characteristics from the overall spectral range shows, that the average error values for meat tissues in bacon vary between 7.4% and 45.4%, while for fat tissue – between
The possibility of separation of the following areas of the investigated products, for a particular day of storage, is examined: 63
IFAC TECIS 2015 64 Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065 September 24-27, 2015. Sozopol, Bulgaria
2.1% and 48.9%. The investigation of class separability by empirical data for NIR characteristics from the overall spectral range shows, that the average error values for meat tissues in bacon vary between 0 and 34.9%, while for fat tissues – from 0 to 47%.
objects is denoted with n, while N denotes the number of the spectral band with the best separability. Table 3. Class separability for bacon in a particular day of storage: cl1 – meat tissues, cl2 – fat tissues
Table 1. Overlap errors for bacon in two different days Day 1vs2 2vs3 3vs4 4vs5 5vs6 6vs7 1vs3 3vs5 5vs7
Error value, εpr% VIS NIR Meat tissues Fat tissues Meat tissues Fat tissues 7.37 2.11 24.21 8.42 17.34 44.48 5.618 2.13 44.56 44.56 4.49 2.25 45.35 48.87 0 0 36.14 38.37 0 0 39.14 40.96 34.94 46.99 16.84 22.11 1.05 1.053 30.34 41.57 0 0 40.24 43.37 0 0
The results for overlap error examination during the investigation of data separability for different areas on a same day of storage are similar.
3 4 7
Classifier Day
n
0 3 6
28 3 43
1. The analysis of color images of the investigated objects gives the most incorrect results when evaluating separability of different areas from the analysed objects. This is due to the fact that the color characteristics of the different areas vary widely and depend on a number of factors that are not related to the conditions of the experiment. However, this method is most suitable for the evaluation of quantitative characteristics of the regions under assessment, for example, their size, which is one of the important quantitative characteristics of products quality.
Table 2. Overlap errors for bacon in two different days
1vs2 2vs3 3vs4 4vs5 5vs6 6vs7 1vs3 3vs5 5vs7
0 3 6
Based on the investigations and the results obtained the following main conclusions can be made:
The overlap errors εpr% for data classes of principal components, obtained from the hyperspectral characteristics of meat and fat tissues in bacon are presented in Table 2. The data separability for a same area in different days of storage is investigated.
2. The investigation of class separability using empirical data in the overall spectral range of the device show, that the overlap error for the two classes of spectral characteristics for bacon vary from several percent to about 49% for visible range and to about 47% for near infrared range. The results for the overlap errors when the separability of the data for different regions on a same day of storage is investigated, are similar.
Error value, εpr% VIS Meat tissues Fat tissues min max min max 0 0.06 0 0 0 0.04 0 0.40 0 0.14 0.13 0.24 0.05 0.17 0.14 0.44 0.25 0.45 0.23 0.38 0.10 0.32 0.07 0.41 0.02 0.20 0.01 0.20 0.03 0.28 0.03 0.35 0.30 0.47 0.26 0.45
n
VIS spectral range LDA SVM-К N εpr% n N εpr % cl1 vs cl2 3 0-0.1 6 6 0-0.06 4 0-0.16 3 3 0-0.12 7 0-0.14 20 9 0-0.12 NIR spectral range LDA SVM-К N εpr% n N εpr% cl1 vs cl2 15 0.05-0.15 55 30 0.08-0.16 3 0.03-0.1 35 9 0.02-0.1 11 0.02=0.12 22 7 0.06-0.13
6. CONCLUSIONS
5.3. Investigation of areas separability using data analysis in selected frequency bands of the hyperspectral characteristics.
Day
Classifier Day
NIR Meat tissues Fat tissues min max min max 0.12 0.26 0.02 0.11 0 0.02 0 0.12 0.01 0.09 0.01 0.09 0 0.00 0 0 0 0.00 0 0 0.24 0.39 0.25 0.43 0 0.012 0 0.011 0 0 0 0 0 0 0 0
3. The investigation of class separability using empirical data from selected frequency bands show, that the maximum overlap error for the two classes of hyperspectral characteristics does not exceed 0.47% for visible range and 0.43% for near infrared range. The results for the overlap errors when the separability of the data for different regions on a same day of storage is investigated are similar.
The investigation of class separability by empirical data for VIS characteristics in selected frequency band shows, that the average error values for meat tissues in bacon vary between 0 and 0.47%, while for fat tissue – between 0 and 0.45%. The investigation of class separability by empirical data for NIR characteristics in selected frequency band shows, that the average error values for meat tissues in bacon vary between 0 and 0.39%, while for fat tissues – from 0 to 0.43%.
4. There is a significant difference between mean values of errors obtained on the basis of spectral data for the entire spectral range and errors obtained using spectral data of the selected frequency bands. The difference in the classification accuracy is two orders of magnitude. These very low values of the overlap errors for data classes from hyperspectral characteristics are received due to the fact that the class separability is accessed in a narrow spectral band, instead of in the entire frequency band of the device. This is perhaps
Table 3 presents the overlap errors εpr%, obtained during the investigation of the data separability for different areas in a same day of storage. The minimal number of frequency bands in hyperspectral characteristics of the investigated 64
IFAC TECIS 2015 September 24-27, 2015. Sozopol, Bulgaria Miroljub Mladenov et al. / IFAC-PapersOnLine 48-24 (2015) 060–065
one of the most important advantages of the hyperspectral analysis to classic spectral analysis.
65
quality assessment of fresh meats, Trends in Food Science & Technology, Vol. 22, Iss. 4, pp. 185–197. Jia J., Schinckel A. P (1992). Automatic measurement of ham composition using image processing. Proceedings of SPIE: The International Society for Optical Engineering, 1778, pp. 128-134. Kumar, S., G. S. Mittal (2010). Rapid detection of microorganisms using image processing parameters and neural network, Food and Bioprocess Technology, vol. 3(5), pp. 741-751. Mladenov, M. Complex assessment of food products quality using color images, spectrophotometric and hyperspectral characteristics analyses. Monography. Publishing center of Ruse University “A. Kanchev”, Ruse, 2015 (in Bulgarian). Prieto, N., Roehe, R., Lavin, P (2009). Application of near infrared reflectance spectroscopy to predict meat and meat products quality: A review. Meat Science, 83, 2, pp175-186. Pu, H., Da-Wen Sun, J. Ma, D. Liu and M. Kamruzzaman (2014), Hierarchical Variable Selection for Predicting Chemical Constituents in Lamb Meats Using Hyperspectral Imaging, Journal of Food Engineering, 143, pp. 44-52. Ripoll G., Alberti P. Panea B., Olleta JL., Sanudo C (2008). Near infrared reflectance spectroscopy for predicting chemical, instrumental and sensoru quality of beef. Meat Science, 80, pp. 697 – 702. Růžičková, J., K. Šustová (2006). Determination of selected parameters of quality of the dairy products by NIR spectroscopy. Czech J. Food Sci., 24: 255–260. Tan J (2008). Meat quality evaluation by computer vision, Journal of Food Engineering, 61, pp.27–35. Tao, F. F., Y. K. Peng, Y. Y. Li, K. L. Chao, and S. Dhakal (2012). Simultaneous determination of tenderness and Escheric hia coli contamination of pork using hyperspectral scattering technique, Meat Science, vol. 90(3), pp. 851–857. Tsenkova, R., Atanassova, S, Itoh, K., Ozaki, Y., Toyoda, K (2000). Near infrared spectroscopy for biomonitoring. Cow milk composition measurements in a spectral region from 1,100 to 2,400 nanometers. J Anim Sci, 78, pp. 515-522. Valous N.A., Fernando Mendoza, Da-Wen Sun (2010), Emerging non-contact imaging, spectroscopic and colorimetric technologies for quality evaluation and control of hams: a review, Trends in Food Science & Technology, Vol 21, Iss. 1, p26–43. Wang, H.H., Sun, D.W (2001). Evaluation of the functional features of cheddar cheese using a computer vision method. J. Food Eng., 49, pp. 49-53. Ziadi, A., X. Maldague, and L. Saucier (2010). Image analysis in computer vision: A high level means for NonDestructive evaluation of marbling in beef meat. In 10th International Conference on Quantitative Infra Red Thermography.
REFERENCES Andrés, S., I. Murray, E. A. Navajas, A. V. Fisher, N. R. Lambe, and L. Bünger (2007). Prediction of sensory characteristics of lamb meat samples by near infrared reflectance spectroscopy. Meat Science, vol. 76, no. 3, pp. 509–516. Chmiel M., M. Slowinski, K. Dasiewicz (2011), Lightness of the color measured by computer image analysis as a factor for assessing the quality of pork meat, Meat Science, Vol. 88, Issue 3, pp. 566–570. Da-Wen Sun (2008), Computer Vision Technology for Food Quality Evaluation, Academic Press, Elsevier London, UK. Da-Wen Sun (2012), Computer Vision Technology in the Food and Beverage Industries, Woodhead Publishing Limited, Cambridge, UK. El Barbri, N., A. Halimi, K. Rhofir (2008). A Nondestructive Method Based on an Artificial Vision for Beef Meat Quality Assesement. Food Science and Technology, pp. 111–138. El Маsry, G., D. F. Barbin, D. W. Sun, and P. Allen (2012). Meat quality evaluation by hyperspectral imaging technique: An overview, Critical Review in Food Science and Nutrition, vol. 52(8), pp. 689-711. Ellis, D. I., D. Broadhurst, D. B. Kell, J. J. Rowland, and R. Goodacre (2002). Rapid and quantitative detection of the microbial spoilage of meat by fourier transform infrared spectroscopy and machine learning, Applied and Environmental Microbiology, vol. 68, no. 6, pp. 2822– 2828. Fagan, C. C., O'Donnell, C. P., O'Callaghan, D. J., Downey, G., Sheehan, E. M., Delahunty, C. M., Everard, C., Guinee, T. P. and Howard, V (2007). Application of mid-infrared spectroscopy to the prediction of maturity and sensory texture attributes of cheddar cheese. Journal of Food Science, vol.72 (3), pp. 130-137. Fornal, J., T. Jeliński, Cheng Jin Du, Da Wen Sun (2007). Inspection of the distribution and amount of ingredients in pasteurized cheese by computer vision. Journal of Food Engineering, Vol. 83, Iss. 1, pp. 3–9. Govindarajan Konda Naganathan, Lauren M. Grimes, Jeyamkondan Subbiah, Chris R. Calkins, Ashok Samal, George E. Meyer (2008). Visible/near-infrared hyperspectral imaging for beef tenderness prediction. Computers and Electronics in Agriculture, Vol. 64, Iss. 2, pp. 225–233. Hai-Hong Wang, Da-Wen Sun (2004). Evaluation of the oiling off property of cheese with computer vision: Correlation with fat ring test. Journal of Food Engineering, Vol. 61. Horváth, K., Z. Seregély, É. Andrássy, I. Dalmadi, and J. Farkas (2008). A preliminary study using near infrared spectroscopy to evaluate freshness and detect spoilage in sliced pork meat, Acta Alimentaria, vol. 37, no. 1, pp. 93–102. Jackman P, Da-Wen Sun, Paul Allen (2011), Recent advances in the use of computer vision technology in the 65