The Pruning Method of the CNN Based on the Comprehensive Assessment of the Convolutional Kernel

The pruning method has been widely studied to compress convolutional neural networks (CNN). However, current pruning methods only use a single strategy to evaluate the convolution kernel importance in different convolutional layers of a CNN. They ignored the differences among the convolutional layers and didn't judge whether it is suitable to use the strategy on a convolutional layer. Although some pruning methods have proven to be able to achieve excellent results, their performance will be greatly reduced if they are adopted on the mismatched layers. To solve the above problems, this article proposes the optimal evaluation pruning (OEP) method to evaluate the applicability between pruning methods and a specific convolution layer before slimming. After that, the most suitable pruning method for each layer can be selected to compress the model. When the OEP is used on the target detection model YOLOV3 based on the COCO2014 dataset, it achieves 2.5X compression and 0.7% decline for mAP, and 3.3X compression, and 6.9% decline for mAP, respectively, which proves that the OEP method is quite effective.