To realize the automatic recognition of high simulation photosensitive seals and explore the influence of training sample size and network model on the recognition accuracy, two high imitation photosensitive seals were prepared by scanning, printing and rubbing design forgery. 3 000 seals were stamped as training samples and 30 seals were stamped as test samples. Four models of convolutional neural network were utilized to identify the high simulation photosensitive seals. According to the results, all the four network models can get 100% recognition accuracy. The convolutional neural network can be used as a feasible method to provide auxiliary reference for the test. By comprehensive analyzing the four network models, Resnet50 is the best choice for the task of high-simulation photosensitive seal printing.