In order to identify accurately and fast the unsound kernels of wheat by image processing technology, a novel detection method was studied based on image features of unsound kernels and BP neural network. The images of unsound kernels were captured and some image processings (median filtering, morphological operations and image segmentation etc.) were performed to extract 54 parameters from three characteristic categories (shape, color and texture). 8 principal components vectors were extracted as the inputs of pattern recognition by principal component analysis. The neural network model was established for identifying unsound kernels of wheat. The results showed that the recognition rate of sound kernels, broken kernels, spotted kernels, sprouted kernels and insect damaged kernels was 93%,98%,100%,90% and 85%, respectively, and the average recognition rate was 93%. It is concluded that this method is an effective way to identity unsound kernels of wheat.