Filter pruning by quantifying feature similarity and entropy of feature maps

Filter pruning can effectively reduce the time cost and computing resources of convolutional neural net-works (CNNs), and is well applied to lightweight edge devices. However, most of the current pruning methods focus on the inherent properties of the filters themselves to prune the network, and pay less attention to the connection between the filters and the feature maps. Feature similarity (FSIM) utilizes the fact that the human visual system is more sensitive to the underlying features of the images to more accurately assess image quality. We discover that FSIM is also suitable for evaluating feature maps of CNNs. In addition, the information richness in the feature maps reflects the degree of importance of the filters. Based on the above research, we propose to quantify the importance of feature maps with FSIM and two-dimensional entropy (2D Entropy) indicator to further guide filter pruning (FSIM-E). The FSIM-E is executed on CIFAR-10 and ILSVRC-2012 to demonstrate that FSIM-E can effectively compress and accelerate the network model. For example, for ResNet-110 on CIFAR-10, FSIM-E prunes 71.1% of the FLOPs and 66.5% of the parameters, while improving the accuracy by 0.1%. With ResNet-50, FSIM-E can achieve 57.2% pruning rate of FLOPs and 53.1% pruning rate of parameters on ILSVRC-2012 with loss of only 0.42% of Top-5 accuracy. (c) 2023 Elsevier B.V. All rights reserved.