Abstract: Pork color grading is an important task of nondestructive detection of meat quality. With analyzing the color feature of pork image and combining its color component value of red, green and blue and its luminance and saturation values converted by RGB color format, a new color feature value was created for the purpose of better expression of pork color. Then, 1070 collected pork samples were graded and marked by experts, and the color feature values of each pork image with special grade mark was obtained after image processing. A series of experiments were conducted to test the pork color grading accuracy rate based on support vector machine. The results showed that along with the increment of training sample number, the pork color grading accuracy rate increased. When all samples were selected as training set, the accuracy reached up to 96.5%. In order to improve the prediction accuracy of SVM (about 37%), locally liner embedding (LLE), a manifold learning method, was applied to reduce or increase the dimensions of pork color feature. The results showed that the prediction accuracy rate of the grading procedure based on LLE and SVM was nearly twice as high as that of the procedure only relying on SVM. Hence, combining the LLE manifold learning method and SVM, pork color grading accuracy will be improved. This method can provide a reference for study and application of nondestructive detection of pork quality.