Abstract:
Aiming at the problems of the traditional capsule neural network feature extraction structure is single, the model parameter is large, and the similarity measurement in the dynamic routing algorithm is rough, an improved capsule neural network is proposed. Using the Fire Module module, the number of feature map channels in the network is first compressed, and then the feature information is extracted through the multi-scale convolution kernel, thereby improving the feature extraction capability of the network and reducing the parameters of the network model. The Dropout idea is introduced into the capsule neural network to increase the diversity of the model, and the Tanimoto coefficient is used in the dynamic routing structure to improve the performance of the dynamic routing algorithm, speed up the model convergence and improve the accuracy. In order to verify the effectiveness of the improved capsule neural network, the improved capsule neural network is compared with the double convolutional capsule neural network, the traditional convolutional neural network and the VGG network model. Experimental results show that the improved model has higher accuracy and faster training speed.