Pruning rate-controlled filter order–information structure similarity graph clustering for DCNN structure optimization methods

Filter pruning is a compression and acceleration method for deep convolutional neural network models that operates at a large scale. Many researchers have studied this approach and made significant progress, but the problem remains an open research topic. In this paper, we propose a DCNN structure optimization method for filter order–information structure similarity graph clustering with global pruning rate control, which considers the mapping strength distribution of convolution kernels in filters and the influence of equivalent convolution kernels on filter similarity. In this method, the relative strengths of the mappings between different convolution kernels in a filter determine the overall type of information combined in the output channel through the superposition filtering of information extracted from different input channels. The structural differences between the equivalent convolution kernels of different filters reflect the differences between the types of information extracted from the same input channel in the convolutional layer. By combining these two factors, we construct a measure of filter order–information structure similarity and then construct a filter similarity graph for the convolutional layer. For the pruning strategy, we establish a convolutional layer filter number allocation model with global pruning rate control using the scaling factors of batch normalization (BN) layers in sparse network. Then, in the filter similarity graph, we perform filter pruning by clustering subgraphs according to the given filter number allocation model for each convolutional layer. This yields an optimized structure for the pruned DCNN model. The experimental results and analysis demonstrate that our proposed method achieves effective pruning. In particular, on the ImageNet dataset, when pruning ResNet-50, the acceleration ratio and compression ratio of the model are 5.31x and 3.78x, respectively, while the model’s classification accuracy decreases only slightly. Our method outperforms several state-of-the-art methods.