Image recognition algorithm based on spatial transform convolutional neural network

Abstract

Aiming at the problem of classification and recognition of noisy handwritten digits, a connection method is proposed to add a spatial transformation network to a convolutional neural network. The spatial transformation network can not only obtain the output results, but also understand the parts of the input data that have the greatest influence on the results and perform feature extraction, which can improve the performance of the model and enhance the interpretability of the model. The experiment was performed on the Cluttered MNIST dataset, and the effects of the conventional convolutional neural network and the improved neural network with spatial transformation network were compared. The experimental results show that the prediction accuracy of convolutional neural networks based on spatial transformation can reach 97.65%, and the results obtained by this method are better than those of conventional convolutional neural networks.