Improved RT-DETR for remote sensing image object detection

Abstract

Aiming at the problems such as complex background, large target scale, dense distribution of small targets, and unstable bounding box regression in remote sensing image target detection, a remote sensing image target detection algorithm based on improved RT-DETR is proposed. Firstly, the convolution and attention fusion module is introduced in the neck network to replace the multi-head self-attention mechanism, forming the AIFI-CAFM module, thereby enhancing the model's ability to capture the global dependency and local detail information of the image; Secondly, a more efficient DySample is adopted as the upsampling operator to reduce the loss of feature information and enhance the ability to retain and recognize the features of small targets. Finally, the Focaler-IoU and MPDIoU loss functions are combined to construct the Focaler-MPDIoU loss function, which improves the regression accuracy of the bounding box and thereby reduces the missed detection rate of the model. The experimental results on the RSOD dataset show that the mAP of the improved RT-DETR algorithm reaches 94.4%, proving that this method can effectively improve the detection accuracy of targets in remote sensing images.