A Pruning Method Based on Feature Abstraction Capability of Filters

Abstract

With the wide application of convolutional neural network, the optimization of CNN has received ever-increasing research focus. This paper proposes a new pruning strategy, which aims to accelerate and compress off-the-shelf CNN models. Firstly, we propose the pruning criteria for the feature abstraction capability of the filter, which is evaluated by combining the kernel sparsity of the filter with the dispersion of the feature maps activated by the filter. Then, the filter with weak Feature Abstraction Capability (FAC) is pruned to obtain a compact CNN model. Finally, fine-tuning is used to restore the generalization ability. And Compared with other pruning methods which use filters of the same layer for contrast, Our method normalizes each layer, the proposed criterion can be applied to the filters between cross-layer of CNN. Experiments on CAFAR-10 and CUB-200-2011 datasets verify the effectiveness of our method. The FAC-based method achieves better performance than previous filter importance evaluation criteria.