Idesf: An Edge Computing-Oriented Filter Pruning Method Based on Indirect and Direct Evaluation Space Fusion

At present, edge computing has attracted widespread attention because of its potential to overcome the problems of high latency and high network occupancy in cloud computing, but it faces the constraints of limited storage space and computing resources in the actual process. To this end, this paper proposed an edge computing-oriented filter pruning method based on indirect and direct evaluation space fusion (IDESF) to make the YOLOv5 network lightweight. IDESF has two advantages over existing methods: (1) Novel evaluation strategy. IDESF compresses CNN models by pruning filters with redundancy in the constructed importance evaluation fusion space, rather than those with “direct-based less” or “indirect-based less” importance. (2) Stronger interpretability of the filter pruning process. The directed graph constructed based on the filters in the fusion space makes the distribution and closeness relationship between the filters visible, and the in-degree of nodes (i.e., filters) in its adjacency matrix enables the redundancy of the filter to be quantified. Therefore, the filter pruning process in our method has better interpretability and visualization. IDESF and SOTAs (i.e., Yolov5s-ghostnet, EagleEye, FPGM, SFP) are evaluated on the VOC2007-2012 dataset and the private MM-dataset. Results show that, when the pruning rate is 0.5, compared with the SOTAs, IDESF reduces the most parameters and the required storage resources, as well more than 40% FLOPs on the YOLOv5 with the highest accuracy on both datasets. Notably, IDESF even improves the accuracy by 8.1% than that of the baseline on MM-dataset.