Design of Lightweight Convolutional Neural Network Based on Dimensionality Reduction Module

The traditional CNN extracted image features is insufficient, so the classification accuracy is not ideal. It also brought overmuch model parameters and calculation. Based on these problems, this paper proposes a dimension reduction residual module. First, reducing the dimension of the output feature map. Then, feature extraction used two different sets of convolution kernels. It can get more sufficient and different characteristic information. Two sets constitute the cascaded layer. Finally, the cascaded layer act as the input of the next layer. This module can reduce parameters. Meanwhile, it increases the depth of the network and enriches the diversity of feature acquisition. A new convolution neural network is build through this module. The performance of new network and other recognition algorithms is compared on GTSRB and 101_food datasets. The new network model is reduced to about 6.2MB, and the classification accuracy can reach 98.2% on GTSRB, 72.3% on 101_food. The experimental results show that this module can effectively improve network performance and control model size.