Research on Image Classification Algorithm Based on CSNet
Changkun Wang, Lei Liu, Qing Li · 2023
This paper proposes a convolutional neural network (CSNet) based on channel separation and re-fusion for the problem of parameter redundancy with the single convolution core of traditional convolutional neural network and the large stack of network structures. First, to solve the problems of single feature extraction and insufficient feature extraction caused by using a single convolution kernel to extract features, as well as the problem of large parameters commonly encountered in traditional convolution neural networks, based on the separation of convolution structures, a feature extraction method based on convolutional structure separation was proposed, which reduces the number of parameters and can extract increasingly rich features. Secondly, by introducing dense residual structure, the network can ensure better feature fusion under low parameters, maximize the utilization of features, improve the generalization ability of the network, speed up network training, and avoid the problem of gradient disappearance during training. Then, the attention separation mechanism is introduced to enable the network to focus on the key and important information among the numerous information without introducing parameters, so as to strengthen the network's attention to features and its ability to extract features. Finally, taking MINIST, CIRFA10, CIFAR100, and Mini ImageNet-100 datasets as experimental objects, the experimental results show that the CSNet network can guarantee high performance while ensuring low parameter quantities. The Top1 accuracy on the test sets of the four datasets is 99.42%, 90.88%, 69.19%, and 68.68% respectively.