Recurrent Convolutional Neural Networks With Applications

Convolution neural network combines the artificial neural network model with the depth learning technique effectively.The network is characterized by local receptive field,hierarchical structure,global training for feature extraction and classification process,and so on.In recent years,it has been applied in many fields such as image classification,target detection,character recognition and achieved a lot of outstanding research results.Due to its strong feature learning and widely used classification ability,it is of great significance to study on convolution neural networks.Based on the analysis of the characteristics of traditional neural networks,a recurrent convolutional neural network is presented in this thesis,in which the Elman-Jordan recurrent model is embedded in the full connection layer of it.This network is named as Elman-Jordan recurrent convolution neural network(EJRCNN).Since the structure of convolutional neural networks is optimized and better feature learning ability of the recurrent network is also combined to decrease the error rate of the network.Experiments show that the proposed networks can achieve better performance with lower error rates of the recognition results than traditional convolutional neural networks.Based on the analysis of the characteristics of EJRCNN,in order to make full use of the low-level and high-level features of the network,we employ the principle of the crossconnect,and then the input of the fully connected layer is the output of CNN's sampling layers,we present a recurrent cross-connect convolutional neural network.Then,more characteristics can be obtained,and thus the correct rate achieved by the network can be further improved.To verify the performance of the developed recurrent convolutional neural networks,some experiments are accomplished on the datasets of Chinese car plates,MNIST,Cifar-10 and a two class dataset.The experimental results show that,the EJRCNN model guarantees a smaller error rate compared with the traditional CNN and Elman convolutional neural network.What's more,the proposed cross-connect recurrent convolution neural network further ensures a much smaller error rate.