Content area
Channel pruning is a method to compress convolutional neural networks, which can significantly reduce the number of model parameters and the computational amount. Current methods that focus on the internal parameters of a model and feature mapping information rely on artificially set a priori criteria or reflect filter attributes by partial feature mapping, which lack the ability to analyze and discriminate the channel feature extraction and ignore the basic reasons for the similarity of the channels. This study developed a pruning method based on similar structural features of channels, called SSF. This method focuses on analysing the ability to extract similar features between channels and exploring the characteristics of channels producing similar feature mapping. First, adaptive threshold coding was introduced to numerically transform the channel characteristics into structural features, and channels with similar coding results could generate highly similar feature mapping. Secondly, the spatial distance was calculated for the structural features matrix to obtain the similarity between channels. Moreover, in order to keep rich channel classes in the pruned network, different class cuts were made on the basis of similarity to randomly remove some of the channels. Thirdly, considering the differences in the overall similarity of different layers, this study determined the appropriate pruning ratio for different layers on the basis of the channel dispersion degree reflected by the similarity. Finally, extensive experiments were conducted on image classification tasks, and the experimental results demonstrated the superiority of the SSF method over many existing techniques. On ILSVRC-2012, the SSF method reduced the floating-point operations (FLOPs) of the ResNet-50 model by 57.70% while reducing the Top-1 accuracy only by 1.01%.