Volume weight is the main index of maize qualitygrades according to national standard. In order to discriminate the maizegrades accurately by machine vision, thegrain image of four differentgrades of maize were adopted by the method of image processing with the industrial camera, the kernels and their background were processed, divided by the average filter, Otsu and morphological operation, the characteristic parameters were selected. The number of optimal main factors was determined by principal component analysis (PCA). The 8-21-4 three layers BP neural network model was established for the identification of maizegrades based on volume weight. Results showed that the overall recognition rate was over 90% based on the image of complete kernel and the kernel transverse section by BP neural network. So the model had high feasibility for detecting maizegrades.