YOLOX-TE: Remote Sensing Image Object Detection Based on Improved YOLOX-Tiny

Abstract

Lightweight object detection models have been widely applied in remote sensing image object detection and have achieved good results. However, there are still some unresolved issues in lightweight remote sensing image object detection, such as dense target distribution, small size, and the inability of the network to effectively extract deep feature information of the target, which can easily lead to issues such as false detections, missed detections, and inaccurate boundary box positioning. Therefore, we propose YOLOX-TE. It can effectively extract deep features of small target objects and enhance the modeling ability of target bounding boxes at different scales. Firstly, it optimizes the feature pyramid network FPN through inverse Focus operation and transposed convolution. Secondly, a multi-scale fusion module based on ECA attention mechanism is added to the head network. Finally, an improved SENet attention mechanism is added to the bounding box regression branch. According to the RSOD dataset validation, compared to the YOLOX-Tiny model, the YOLOX-TE model only increased the parameter count by 0.363M, but the accuracy of AP, AP75, APS, and APL increased by 7.1%,6.7%, 6.8%, and 7.3%, respectively. Compared with the most advanced lightweight YOLOV7-Tiny model, the accuracy of AP and AP75 is 0.4% and 2.1% higher, respectively, and YOLOX-TE has a smaller volume and lighter weight.