Progressive kernel pruning with saliency mapping of input-output channels

As the smallest structural unit of feature mapping, the convolution kernel in a deep convolution neural networks (DCNN) convolutional layer is responsible for the input channel features to output channel features. A specific convolution kernel belongs to a specific group from the perspective of the input channel, and it belongs to a specific filter from the perspective of the output channel. If the input and output channels are simultaneously considered in the pruning process, the performance of the pruning model can be further improved. This paper proposes progressive kernel pruning with salient mapping of input-output channels, introduces the concept of input-output channel saliency and defines single-port salient mapping channels and dual-port salient mapping channels. This study demonstrates that single-port salient mapping channels can ensure that each input channel signal has a relatively strong convolution kernel mapped to the output channel, and vice versa. The dual-port salient mapping channel is a channel with high feature mapping abilities from both the input and output directions. Additionally, the average mapping ability measure index is defined, which is used to control the kernel pruning process of the single-port salient mapping channel to switch to the kernel pruning process of the dual-port salient mapping channel. The experimental results and analysis show that the method proposed in this paper can be used to effectively prune a heavyweight model and a lightweight model and can obtain a better accuracy under the conditions of higher compression ratio and acceleration ratios. For example, when VGG-16 is pruned on CIFAR-10, the compression ratio and acceleration ration are 91.00 × and 15.24 × , respectively, and the classification accuracy of the model decreased slightly by 0.22%. When ResNet-101 is pruned on ImageNet, the compression ratio and acceleration ratio are 3.90 × and 3.38 × , respectively, and the classification accuracy of the model decreased slightly by 0.48%. The proposed method is significantly better than state-of-the-art methods.