Vision based methodology for evaluation of meat quality characteristics
Abstract
Meat quality is a subject of growing interest. Important meat quality parameters include colour, tenderness, texture, flavour and odour, water holding capacity and juiciness. The quantification of meat quality is a challenge of major importance in the meat industry. The ability to meet consumer expectations by providing quality products and maintaining the consistency of products is the basis for the long-term success of suppliers in the competitive meat industry. Visual assessment has become a main component of meat quality evaluation. Visual assessment is subjective, inconsistent and highly variable. It is not capable of detecting commercially important attributes of carcasses. Objective measurements are needed to ensure consistency and reproducibility of meat quality evaluation.
Vision based methods have been shown to be an effective means for meat quality evaluation. At present there is no online method available in New Zealand to measure meat quality characteristics, but there is a real need to determine these characteristics as the market moves to supply meat to the quality specifications required by modem consumers. This is reflected by large supermarkets internationally dictating quality specifications to processors with the aim of capturing consumers and market share.
The objective of this project was to investigate the possibilities of computer vision based approaches in lamb quality evaluation and to provide an imaging system, which can be used at processor and retail operations, as an interactive system. The information available in the captured images is translated into parameters, which relate to meat quality characteristics. This imaging system should be able to determine the grade, the relative abundance of different components of meat (lean meat, marbling, etc) and, more importantly, some eating quality characteristics.
The individual modules involved in a computer vision system are image acquisition, pre-processing, segmentation, image analysis, feature extraction, feature pre-processing, feature selection and classification. Illumination is a very important aspect in image acquisition. The specular reflections caused by moisture on the meat surface are a major problem in meat imaging. The polariser-analyser filters were used to eliminate most of these reflections.
The next important steps in a vision system are image pre-processing, segmentation and analysis. Image processing is used to pre-process images before computer analysis. The first step in image pre-processing was background removal. Edge enhancement, filtering and/or sharpening techniques were then used to remove any noise present and to enhance the visual quality of the images. Proper segmentation algorithms are very important for a successful vision system. Colour, brightness and edge detection were used to segment images into regions. Image analysis quantified the elements of interest like lean, fat, connective tissue, marbling and bone areas from the meat images.
Feature extraction is an important pre-processing method for classification; perhaps it is the heart of a pattern recognition system. Geometric and texture information are important elements of a computer vision system. Twelve geometric (area and thickness) features were measured from lamb chop images. In addition, 136 texture features were also measured. These include 36 grey-level difference histogram (first-order texture) features, 90 grey-level co-occurrence matrix (second-order texture) features and 10 grey-level run length matrix (higher-order texture) features.
Feature selection is another important aspect of pattern recognition. It simplifies data by eliminating redundancy and isolating the important characteristics of the data, allowing better insight and concise representation of the data structure. Principal component analysis used for dimensionality reduction. During the dimensionality reduction, six geometric, eight co-occurrence, four run length and four grey-level difference histogram features were selected from an original set of features with total variance of 96.1%, 99.2%, 97.9% and 86.7%, respectively. Principal component scores were calculated as linear combinations of measured geometric variables. The other alternative to the principal component score is to select one of the measured variables to represent the principal component.
In classification, we used both multivariate statistical and neural network analyses. The classification was performed in stages using reduced and original feature sets as well as principal component scores. Different combinations of these feature sets were used. In most cases, a reduced feature set produced better classification than principal component scores. The highest classification of 85.6% was achieved with original features (12 geometric + 10 GLRM + 36 GLDM), which was only 1 % higher than that of reduced features (6 geometric + 8 GLCM + 4 GLRM). It indicates that the reduced features have a similar discriminatory power as the original features. In all cases, the classification rates achieved with reduced features were still comparable, because of the simplicity in feature extraction. Analysis of misclassified images was performed to study the behaviour of the misclassifications. The classification using a single reduced feature set produced a few unusual misclassifications. Such misclassifications were minimised by combining the reduced feature sets.
The lamb carcasses were regraded more objectively using fat thickness measurement (fat 11) and carcass weight, revealing that 41.3% of carcasses have previously been graded incorrectly mainly because of error in estimating fat thickness, and also some carcasses were graded incorrectly as higher weight carcasses. The highest classification of 88.8% was achieved using original features (12 geometric + 10 GLRM + 36 GLDM), which was only 2% higher than that of reduced features (6 geometric + 8 GLCM + 4 GLDM). Principal component scores produced a lower classification than the reduced features.
The hot carcass weight was included in the classification using reduced and original feature sets and increased the accuracy of the overall classification in all cases. The highest classification of 89.4% was achieved with original features (12 geometric + 36 GLDM) and carcass weight, which was only 1.3% higher than the classification achieved with reduced features (6 geometric + 4 GLDM) and carcass weight. The data reduction was useful even if we include carcass weight as a variable in the classification.
The neural network approach was attempted to further improve the results and search for possible non-linear relationships. Multi-layer perceptron networks containing back propagation learning algorithms were trained to classify lamb chop images into different grades. Several different neural networks and learning parameters were tested to select the best neural network. The highest classification rate of 96.9% was achieved with reduced features (6 geometric + 8 GLCM + 4 GLRM). The single reduced feature set produced a few unusual misclassifications and combining the reduced feature sets eliminated the unusual misclassifications.
Single hidden layer networks with a number of hidden neurons were trained to classify lamb chop images into new grades. Several different learning parameters were tested to select the best neural network. The highest classification rate of 96.3% was achieved with reduced features (6 geometric + 8 GLCM + 4 GLRM + 4 GLDM). The neural network analysis was also performed with hot carcass weight added as an input variable in order to classify lamb chop images into old grades as well as new grades. The highest classification rate of 98.8% in predicting old grades was achieved using reduced features (6 geometric + 4 GLRM) in combination with carcass weight. The highest classification rate of 98.8% in predicting new grades was achieved using reduced features (6 geometric and 8 GLCM) in combination with carcass weight. The effect of number of hidden neurons and activation functions were also studied.
The relationship between fat measurements using chemical analysis and image analysis was examined using best-fit models of the SPSS curve estimation procedure. Statistically, the relationship was best described by a non-linear regression. The equation obtained for the prediction of crude fat percentage from image analysis measurements was In(C) = e¹•²⁷⁵⁵⁻⁽⁸•⁶²⁹¹/ⁱ ⁾ (R² = 0.81).
Multi-layer perceptron networks containing back-propagation learning algorithms were trained for the prediction of lamb tenderness. Several different neural networks and learning parameters were tested to select the best neural network. The effects of network parameters, architecture, different features and feature selection on prediction have been studied. The highest coefficient of determination of 0.746 in predicting lamb tenderness was achieved using the reduced (6 geometric + 8 GLCM) features.
The primary objective of this research, investigating the possibilities of computer vision based approaches in lamb quality evaluation, was carried out successfully. The result of this work is an imaging system with image and texture analyses based feature extraction, multivariate statistical techniques based feature selection and a neural network based classification and prediction model.
