FL-YOLO: Lightweight Small Target Detection Algorithm Based on Transformer and CNN Hybrid Structure

Abstract

Although the target detection technology has made amazing progress in the detection of large and medium objects, due to the limited size of small objects, the occlusion between targets and the limitations of the convolutional network itself, small target detection is still as a challenging. In this paper, by improving the YOLOv5s algorithm, a lightweight small target detection FL-YOLO model with a hybrid structure is proposed, and the standard convolution in the original backbone network is replaced by depth-separable convolution and the residual connection method is adjusted. lightweight. A Patch-Attention (PA) module is proposed to extract context information of small objects and insert it into the backbone to enhance the feature extraction ability for tiny target. A new upsampling module named “Transition” is proposed to replace the nearest interpolation in YOLOv5s to minimize information conflict and information redundancy. Finally, the public dataset VisDrone2019 specially constructed for small targets was used to conduct experiments, which verified that the FL-YOLO proposed in this paper is more effective than YOLOv5s.