In this paper, we provide a method to recognize weed seeds based on computer vision. According to the stability and genetic characteristics described in phytotaxonomy for weed seeds. Image processing method encompasses threshold segmentation and smooth processing etc, nine features parameters are extracted by image processing, which keep RST invariance. The principal components analysis method is adopted to reduce dimensions of weed seeds shape data, the first three principle components as the parameters of the weed seeds shape are put input the BP neural network. The results indicate that the precision ratio of recognition is improved while training time is reduced compared with the traditional simple method. This method can be well applied in weed seeds recognition.