YOLOX同步特征融合网络及其遥感目标检测OA

To address the issues of poor localization and low detection accuracy in optical remote sensing image-based object detection,an improved lightweight object detection model and algorithm based on YOLOX is proposed.In the structural network design,building upon a Biased Texture Feature Puamid Networks(BTFPN),a synchronous feature fusion network is proposed to fully exploit shallow-layer detailed information,focus on important channel features,and efficiently transfer location information and edge features from shallow-layer networks.In the detection head,by combining the advantages of high-resolution detection heads for small object detection with the requirements of the detection head for enhanced bounding box regression tasks,the detection head is improved into a regression-enhanced and feature-enhanced detection head.This modification addresses the problem of missed detections caused by the loss of semantic information in small objects and improves the inference capability of bounding box regression.Additionally,by establishing an improved SIoU loss function that focuses on the distance and shape differences between bounding boxes,the object localization accuracy is improved.Comparative experimental results on the remote sensing datasets DIOR and RSOD show that the proposed model achieves small regression loss with fewer parameters and demonstrates high detection accuracy for objects of various sizes.