Image Classification Based on IndCRNN Module
Huaqing Liao · 2020
As a core of computer vision, image classification has become a hot topic in recent years. RNN framework can not only model sequence data, predict the trend of sequence data, but also show good results in image classification. In this paper, an improved model IndCRNN is proposed for the over-fitting phenomenon of Independent Recurrent Neural Network (IndRNN) in the training process. IndCRNN is the Independent Convolutional Recurrent Neural Network, which applies the convolutional idea of Convolutional Neural Network (CNN) to IndRNN. On the one hand, the dimension reduction can be realized. In the process of training iteration, the training time will be reduced due to the reduction of dimension. On the other hand, if the high-dimensional feature is taken as the input of IndRNN, the basis of the final output classification becomes the feature of two dimensions of space-time. Another advantage of IndCRNN is that it selects Leaky ReLU as the activation function, replaces the traditional ReLU activation function, maximizes the historical information of the training data, and reduces the vulnerability in the process of updating iterative training. Finally, we did an experiment with TensorFlow environment, experiment selects the MNIST dataset and permuted MNIST dataset. The IndCRNN model is compared with IRNN, LSTM, RNN and IndRNN on two datasets respectively. The experimental results show the feasibility and superiority of the proposed model.