Research On The Methods Of Meat Quality Grading Based On Image Processing

With the increase of the scale of animal production, inspection and evaluation on the quality of meat become more and more strict and standardized in China. Usually, the evaluation method of meat quality needs a large number of physical and chemical tests, and it's not suitable for fast and automatic production line. The application of image processing technology in meat quality evaluation system is expected to realize the automation of nondestructive testing for meat quality. The methods of meat quality grading by image processing technology are studied in depth, including denoising, segmentation, feature extraction and classification of meat image, and are described as follows:Firstly, a denoising method of meat image based on Harris corner detection and spatially adaptive iterative singular-value thresholding(SAIST) is researched. The Harris operator is used to detect and filter out the impulse noise; and SAIST method is used to further remove the residual Gauss noise. The experimental results show that, compared with the method based on wavelet transform, the method based on pulse coupled neural network and median filtering and the method based on joint statistical model, meat images obtained by the proposed method are better in the matter of subjective visual effect and objective evaluation indexes.Secondly, a segmentation method of two-dimensional Arimoto gray entropy for meat image is proposed. considering the uniformity of within-class gray scale directly, the formula for one-dimensional Arimoto gray entropy threshold selection is constructed. Then the formula for two-dimensional Arimoto gray entropy threshold selection based on gray scale-gradient two-dimensional histogram is derived. Artificial bee colony algorithm, improved by the chaotic sequence based on Tent mapping, can improve the precision and speed while searching the optimal threshold. The experimental results show that, the overall effect of proposed method is better than those of two-dimensional Shannon entropy thresholding method, two-dimensional Tsallis gray entropy thresholding method and two-dimensional Arimoto entropy thresholding method.Then, two segmentation methods of meat image are proposed using fuzzy local information C-means clustering based on generalized kernel function(Kernel Fuzzy Local Information C-means-Universal Gaussian, KFLICM_UG) or hybrid kernel function(Kernel Fuzzy Local Information C-means-Mixed Gaussian, KFLICM_MG). Firstly, the generalized kernel functions or hybrid kernel functions are used to strike a good balance between the learning ability and the generalization ability. Pixels which are mapped into a high dimensional feature space can be clustered more easily in the high dimensional feature space. Then, based on the combination of its local space information with grayscale information, the meat image segmentation is completed via the fuzzy local information C-means clustering. The experimental results show that, compared with the existing fuzzy C-meanssegmentation method, kernel fuzzy C-means segmentation method and fuzzy local information C-means segmentation method, the proposed method(KFLICM_UG, KFLICM_MG) can achieve stronger adaptiveness and robustness against noise.Next, A beef marbling grading method based on invariant moments, gray level co-occurrence matrix and mixed kernel support vector machine(SVM) optimized by chaotic bee colony is proposed. Firstly, invariant moments and the statistical quantity of gray level co-occurrence matrix of the beef marbling image are computed in order to construct a feature vector. To attain the optimal recognition performance of mixed kernel support vector machine, the penalty factor and kernel parameters of the mixed kernel SVM are optimized by chaotic bee colony algorithm. Finally, the samples to be graded are input to SVM for classification and recognition. Compared with the method based on gray moment and SVM, and the method based on gray level co-occurrence matrix and BP neural network, the proposed method attains the highest grading accuracy.Finally, a beef marbling grading method based on completed local binary pattern(CLBP), kernel principal component analysis, and random forests(RF) is given. Firstly, CLBP-Center, CLBP-Sign, and CLBP-Magnitude are used to form completecode, which represents the feature vector of beef marbling image. Finally, RF is used for classification. A large number of experimental results show that, compared with the method based on fractal dimension and image features, method based on gray level co-occurrence matrix, and the method based on back propagation neural network, the proposed method attains the highest recognition rate.