Abstract: Corn seed purity is closely related to corn yield, so seed selection plays an important role in improving grain yield product. The automatic seed selection procedure based on the machine vision is usually divided into three steps: image segmentation, feature extraction and classification. Variational model for image segmentation and corresponding numerical technique of Split Bregman method were introduced into the identification procedure, which had advantages of feature extraction such as high accuracy and closed continuous border. In addition, the adaptive wavelet collocation method was employed to solve the optimality conditions in Bregman split method. Based on the improved method, the corn geometric features can be extracted more precisely. Nongda108 and Ludan981 were taken as examples to test the new method. Based on a classifier designed with SVM, results showed the identification accuracy of Nongda108 and Ludan981 were 97.3% and 98%, respectively, better than 95% in previous research.