Study On The Automatic Grading For Beef Marbling Based On Computer Vision & Artificial Neural Network & Image Processing

Although an official beef grading system has not been put into practice in China, the beef grading standards were worked out and published in 2002.Quality grades of beef are primarily based on visual appraisal of the longissimus dorsi muscle. the marbling level is the dominant parameter in deciding the beef quality, which is usually determined by authorized experts called graders according to the marbling abundance. Since the grading of beef marbling is largely determined by the subjective experience of the graders, there are inconsistencies and errors in judgment. So computer vision has been recognized as the most promising approach to objective assessment of beef quality at present.In this study, it was proposed the method of automatic control the brightness value of target in image (ACBT) in RGB color space. The brightness value of target in image can be adjusted to set value by ACBT before images segmentation. Probabilistic neural network( PNN) trained by collected pixel samples in this study and image processing methods were used to divide beef marbling from the beef image and the beef marbling includes the little branch of subcutaneous fat. We used the methods to divide the beef marbling from 112 images of beef and extracted the data of 7 characteristic parameters of the beef marbling. The residuals analysis was adopted to deal with the outliers and get the correct motion parameters from the data. By using the method of PCA (Principal Components Analysis), two main components were chosen as the comprehensive evaluation index from7 characteristic parameters. Probabilistic neural network (PNN) models were created based on the two main components to evaluate automatically grade of beef marbling.Analysis of the limitation of Otsu method in images segmentation shows that the method are affected by the background so that the method are disagree with field imaging. But the advantages of the method are rapid operation.In this study, RGB and I₁I₂I₃ color space have been adopted because it is computational complexity that RGB color space be converted into HSI. Artificial neural network have been adopted for images segmentation. Through the methods comparison between probabilistic neural network and BP neural network, the results showed that the BP neural network is unstable and the probabilistic neural network is very stable and more precise and it is in RGB color space better than in I₁I₂I₃ color space. But they are all sensitive to brightness of image.In this study, the two methods of automatic control brightness value of target in image (ACBT) in the RGB color space were proposed. The first method adjusts brightness value of target in image, through brightness value of image is multiplied by the ratio of brightness value of central zone of image and set value. The second method adjusts brightness value of target in image, through RGB sorting of image are multiplied by the ratio of average of brightness value of RGB sorting of central zone of image and set value. The results showed that the first method is better than the second method.In this study, inefficient longissimus dorsi muscle were eliminated by image processing methods and the beef marbling includes the little branch of subcutaneous fat were divided accurately from the beef image. The divided area of little branch of subcutaneous fat can be adjusted by set value.The image processing methods includes median filter, image filling, image corrosion, labelling connected component, marker selection , image dilation, edge-extracting, and operation, not operation, or operation and image add.In this study, the adopted methods of probabilistic neural network(PNN) and image processing can be used to divide accurately beef marbling from the beef image in almost any backgrounds.But the method also has limitation. This paper has analyzed the limitation and put forward the simple and effective solutions.The extracted 7 characteristic parameters of beef marbling include fat particle,big fat particle,small fat particle,big fat particle ratio,small fat particle ratio, equilibrium of distribution of fat. By using the method of PCA (Principal Components Analysis), two main components were chosen as the comprehensive evaluation index from 7 characteristic parameters. The total of variance contribution of two main components was 97. 2569%. Probabilistic neural network, BP neural network and genetic BP neural network were trained and verified by the data of 7 characteristic parameters and two main components respectively.The results showed that the R of linear regression analysis of experimental results and simulated values of neural network model were higher than the R of multiple regression analysis model (R of multiple regression analysis model is 0.767 and 0.761 respectively) .But the BP neural network model and genetic BP neural network model are unstable. The probabilistic neural network model is very stable and more precise . The R of linear regression analysis of simulated values and experimental results is 0.932 and 0.994 respectively. Obviously, the probabilistic neural network model base on the two main components is better than base on the 7 characteristic parameters.This study has important reference value for the developing new and high technical products based on computer vision and neural networks to grade for beef marbling of our country. It has important economic meaning for the extending and application beef grade standard, raising the grading technology of meat quality in China.All the experiments in this study have been done on Matlab7.0 platform by programming.