Abstract: Aiming at the problem that traditional convolutional neural network relied heavily on the amount of data, we proposed an image recognition algorithm based on mean iterative threshold segmentation method and convolutional neural network. The image background was filtered by means iterative threshold segmentation method, and a new convolutional neural network was constructed based on AlexNet. Compared with other commonly used convolutional neural networks, the experimental results show that the recognition effect of the proposed algorithm is ideal in image recognition tasks with insufficient samples, and it has higher recognition accuracy, lower recognition error and faster convergence rate than that of other convolutional neural networks.

Key words: image recognition, threshold segmentation method, convolutional neural network