In order to efficiently detect the internal cracks of corn seeds, a detection system and batch detection method based on convolutional neural network (CNN) were designed, and the cracked and non-cracked corn seeds were collected to make a data set, and four classics of AlexNet, VGG11, InceptionV3 and ResNet18 were constructed. Convolutional neural network, and compared with the traditional algorithm model SVM and BP neural network at the same time. It was found that the convolutional neural network model was better than those two traditional algorithm models. The ResNet18 model had the best comprehensive detection performance. The recognition accuracy of single seeds with cracks was 95.04%, and the recognition accuracy of single seeds without cracks was 95.04% and 98.06%, and the per grain detection time was 4.42s. During the corn seed internal crack recognition system based on ResNet18, the system experiment carried out 10 sets of batch recognition. The average accuracy rate of cracked seeds was 94.25%, and the average recognition accuracy rate of non-cracked seeds was 97.25%. The transmission of light source in batch recognition was not equivalent. Accuracy can be affected by reasons such as the internal cracks of all seeds and the lack of generalization caused by multiple loading of model weights. Finally, an automatic identification device for internal cracks in the seeds was built, and a software control device for identification was designed to complete the internal crack identification system of corn seeds. The deep learning algorithm provided a guarantee for the detection of internal cracks in corn seeds. The research result would lay a foundation for the detection of internal cracks in corn seeds in the assembly line.