Design of Convolutional Network Based on Cascade
Feng Yanyan, Yue Fei Zhou, Zeng Shangyou, Pan Bing · 2019
As to the phenomenon that traditional convolutional neural network using a single channel and a single scale convolution kernel to extract image feature leading to low classification accuracy, the cascade convolution module used to extract image feature in this paper. The module uses three different sets of convolution kernels to extract feature. In the process of sampling, the dimension of output is kept consistent. This module can reduces convolution layer parameters. Simultaneously, it also can increase the width and depth of the network. Convolutional neural network is built by using the module. The experimental datasets are Caltech256 and 101_food. Compared with AlexNet, the improved network model parameters of the module are reduced. The recognition accuracy of Caltech256 dataset is increased from 52.83% to 60.01%, and 101_food dataset is increased from 56.76% to 67.39%. The experimental results show that the cascading module can effectively improve network performance while reducing network parameters.