An acceleration and optimization method of edge computing using filter level pruning

Abstract

In order to reduce the complexity of the network and the amount of floating-point computation, this paper proposes an efficient target detection algorithm. Firstly, we use the feature of deep separable convolution to reduce the amount of calculation and design a target detection network structure. Based on the new network structure, the filter is pruned by using the size of the L1 norm value as the threshold to further compress the network model. Then it is deployed on the embedded GPU platform of the edge device for verification, and excellent algorithm performance is achieved. The combination of deep separable convolution and filter-level pruning greatly improves the computational performance of the target detection algorithm on the edge device.