An Attention-Lightweight YOLOv5 Aerial Dense Object Detection Method

Abstract

The paper proposes an improved YOLOv5-CBG (SX-CBAM, Ghostbottleneck) real-time target detection network for the problem of false detection and high missed detection rate in real-time detection caused by a large number of dense small targets in complex aerial images. The network first adds a hybrid pooling layer to the attention mechanism CBAM (Convolutional Block Attention Module), and integrates the residual path into its structure, and then embeds the improved SX-CBAM into the YOLOv5 detection network to improve its detection performance. Secondly, combined with the lightweight Ghost bottleneck module to reduce the computational cost of the network and ensure the real-time performance of target detection. The results of comparison experiments on the dataset VisDrone2019 show that the mAP50 of YOLOv5-CBG is 5.2% higher and the GFLOPs are reduced by 1.4 compared to YOLOv5, demonstrating that the improved detection network reduces the false detection rate and missed detection rate for dense small targets while satisfying real-time target detection.