Lightweight-YOLOv5s for the camouflaged personnel detection in the woodland

Abstract

Unmanned aerial vehicles(UAVs) technology and target detection technology are increasingly integrated. However, the target detection networks are cumbersome and complex, and it is hard to directly deploy these networks on the UAVs. To address this problem, a Lightweight-YOLOv5s which can be applied to the UAVs is proposed to detect the camouflaged personnel in the woodland. Depthwise separable convolutions are added to the YOLOv5s for reducing the parameters and computational effort. Then a new channel pruning method based on Batch Normalization(BN) layer scaling factor with adaptive sparsity training is proposed for network compression, which can balance the compression ratio and accuracy of the detection. The pruned network is then fine-tuned to detect the camouflaged personnel. In addition, a special dataset is built for the camouflaged personnel detection in the woodland. The quantitative results on this dataset prove that the proposed Lightweight-YOLOv5s greatly improve the YOLOv5s. Parameters from 7.06M to 1.86M, floating point operations from 16.4G to 9.9G and inference time from 8.16ms to 7.06ms. Besides, mAP of the Lightweight-YOLOv5s is 0.83, which shows the excellent detection performance of the proposed method.