An improved YOLOv5 method for small object detection in high resolution images

Abstract

The dense small objects detection is a challenging task in the scenario of UAV aerial surveillance. This paper proposes an improved YOLOv5 detection method for the dense small objects in high resolution images. To augment the dataset, a 20% overlap crop is used for the UAV aerial photography training set. In order to detect the tiny objects in the aerial photos of UAV, a tiny detection head is added on the basis of YOLOv5. The SPP and CBAM modules are introduced in the head of the model, SPP for feature fusion at different scales and CBAM for adding attention to spatial and channel dimensions. Multiple experiments are conducted on the VisDrone 2019 dataset, the results show that the mAP of 12 classes detected by the model is 30.4%, and 3.1% higher than the original YOLOv5.