Author(s)

Guangmin Sun, Zihao Zhang, Zixi Zhu, Shikui Guan, Yu Weng and Chong Shi

Corresponding Author

Guangmin Sun

ABSTRACT

Deep learning makes object detection easy, but the network structure is complex and there is a lot of redundancy, which results in a very slow operation. This paper proposes a pruning method based on the BN scaling factor to prune the redundant structure of the network and apply it to vehicle detection tasks. First, train the network for channel sparsity to make it easier to screen out important channels. While training network weights, the L1-norm is applied to the scale factor and bias of BN. Structures with smaller BN layer parameter values are penalized. Secondly, the network is automatically pruned to obtain the minimum structure. The BN scaling factors in the network structure are sorted, and the structures with smaller scaling factors are pruned according to the proportion. Then, the bias of the pruned structure is transferred to subsequent layers to maintain accuracy. In this paper, the pruning strategy considers all convolutional layers, so a high pruning rate is obtained, and the model after pruning can still maintain high accuracy without any training. For vehicle data, after training and testing, the accuracy of the model after pruning only decreased from 50.4Map to 47.27Map, but the network structure reached 76% compression. GFLOPS decreased from 33.12 to 8.16, and the number of million parameters decreased from 61.95 to 6.95.

KEYWORDS

Network pruning; vehicle detection; BN scaling factor; deep learning; YOLO