FMIP: Feature Map Importance Pruning for Efficient CNN Optimization

Deep convolutional neural networks (CNNs) have reformed numerous fields, comprising computer vision and natural language processing. Nevertheless, CNNs' computational complexity frequently raises obstacles to utilization on resource-constrained devices. Different optimization strategies have been proposed in response to this issue, comprising quantization, knowledge distillation, and pruning. This paper familiarizes a novel pruning method, Feature Map Importance Pruning (FMIP), designed to optimize the performance of deep CNNs while reducing computational costs. The FMIP computes the importance of feature maps within convolutional layers according to the area of their activation values, facilitating a systematic approach to pruning decisions. By detecting and eliminating the redundant feature maps, FMIP can considerably cut the number of parameters and FLOPs in a CNN model without conceding accuracy. This is chiefly advantageous for using CNNs on devices with limited memory and computational power. We evaluated FMIP on various CNN architectures, including VGG16 and ResNet50, using datasets such as CIFAR-10 and ImageNet. Our experiments demonstrated significant model compression while preserving high accuracy. Specifically, FMIP achieved the following results: VGG16 with CIFAR-10 demonstrated an 81.9% reduction in parameters and a 48.6% reduction in FLOPs, with an accuracy of 93.33%. ResNet50 with ImageNet achieved an 83.11% reduction in parameters and a 71.55% reduction in FLOPs, with an accuracy of 92.03%.