In this chapter, we study the feasibility of using transfer learning to identify the embryo surface and variety of maize seeds. First, the original image is segmented, and region of interest (ROI) is extracted; second, the extracted image is imported into a model that is pretrained by image-net data set; and finally, the recognition task can be completed efficiently through the local fine-tuning of the model. In order to comprehensively evaluate the productivity of a maize species, we test and identify its embryo surface, non-embryo surface, and longitudinal cut surface. The model we use here has also been applied to and achieved good results in recognition of peanut pod. Experimental results show that the transfer learning can be used for image recognition and has good prospects of application. More maize varieties need to be studied in future to expand the application of transfer learning to image recognition.