An Improved Lightweight Small Target Detection Algorithm for Air-to-Ground Images

作     者：Cao, Lijia Song, Pinde Wang, Yongchao Geng, Chuang Peng, Baoyu 

作者机构：Sichuan University of Science & Engineering Yibin644000 China Technical University of Munich Munich80333 Germany Artificial Intelligence Key Laboratory of Sichuan Province Yibin644000 China Sichuan Province University Key Laboratory of Bridge Non-destruction Detecting and Engineering Computing Yibin644000 China 

出 版 物：《SSRN》 

年 卷 期：2022年

核心收录：

主　　题：Signal detection 

摘      要：A lightweight target detection algorithm, MobileNetv3 extended neck YOLOv5 (MEN-YOLOv5) is proposed in this paper to simplify the complex structure of the YOLOv5 target detection network, reduce the number of model parameters, and improve detection accuracy of small targets. Firstly, to reduce the number of parameters, the backbone network YOLOv5 is re-constructed using the MobileNetv3 depthwise separable convolution. Secondly, in accordance with the characteristics of the air-to-ground detected targets and the YOLOv5 algorithm, the layers are extended to the shallow-layer on the basis of the Neck network structure and more shallow feature information are fused, which makes the developed algorithm have a smaller receptive field and improves accuracy of detection of small targets. Finally, channel parameters are re-designed and the Neck network structure is rebuilt to avoid the dramatic increase in the number of parameters. As a result, the number of parameters in the entire network structure are compressed to the maximum extent possible to produce lightweight target detection models. The effectiveness of the proposed algorithm is verified by experiments with the VisDrone-DET2019 dataset and the experimental results show that, compared with YOLOv5l and YOLOv5x, the numbers of parameters of MEN-YOLOv5-L reduce by 86.8% and 92.98%, and floating point operations (FLOPs) decrease by 46.18% and 71.58%, and mAP50 values increase by 7.18% and 5.78%, respectively. MEN-YOLOv5-S has 52.9% and 30.8% lower number of parameters and FLOPs, respectively, and 2.5% higher mAP50 than YOLOv5s. © 2022, The Authors. All rights reserved.