Convolutional neural network abbreviated CNN has been widely used in pattern recognition field, and the CNN network structure optimization is one of the important factors affect recognition accuracy and efficiency. In this paper，the network structure is systematically analyzed, and the key parameters of the network structure and the way to assume value are given. In the process of recognition based on CNN, the results have been deeply influenced by the number of hidden layer characteristics figures. Meanwhile, recognition accuracy of the system is affected by some small relational samples between a layer features figure and the next layer features figure. Grey relational analysis excavates internal relational with data. The paper based on GRA in the process of network training is automatically selected effective features of hidden layers and the network structure is optimized, and traffic signs are served as object recognition and validation. Experiment results show that the proposed method adaptively determines the number of characteristics figure, and realizes optimization of CNN network structure. It improves the efficiency of confirming the network that compared with the experimental method.
展开▼
