Cartoon Figure Recognition with The Deep Residual Network

Abstract

Because of the wide application of deep learning, there are more neural network structures in image recognition technology nowadays, but there are various differences in the accuracy of image recognition because of the various differences in network structures. For this reason, it is especially important to use different neural network structures for different forms of image data. This paper focuses on exploring the differences between LSTM networks, residual networks, and CNN networks in terms of the accuracy of cartoon character recognition.[1]Firstly, the web crawler acquires 14 different cartoon character images and manually screens the original data to remove the duplicate images and obtain the preliminary data. Then data enhancement was performed on the preliminary data, and the form of rotating the images was selected to complete the pre-processing of the data, which solved the problem of using different code forms for different forms of data importing into the neural network; the LSTM network, CNN network and CNN network with added residual function were used to recognize the pre-processed data. The experiments show that the CNN network structure with residual function can achieve higher accuracy compared to LSTM, with the final result of 76.08%.