RTPR-DETR : A Real Time Small Object Detection Network Based on Partial-Convolution with structural re-parameterization

To address the issue of high parameter and computational requirements in current drone target detection, a lightweight small object detection network for drone aerial images is proposed. Firstly, through simple experiments, the differences in parameter and computational requirements of various convolution operations are analyzed. The PConv (Partial Convolution) is then introduced to replace the convolution operations in RT-DETR. This replacement reduces the computational and parameter requirements of the model while maintaining its accuracy. Secondly, structural re-parameterization and multi-scale attention are incorporated into the backbone network to enhance feature extraction, thereby improving the model’s accuracy in detecting small objects. Lastly, considering the dominance of small objects in drone images, the IoU loss function is modified to Inner-MPDIoU loss, which increases the contribution of loss for small objects and further improves small object detection performance. According to the experimental results on the VisDrone dataset, there is a significant improvement in comparison to the original RT-DETR-r18 model, the proposed model achieves a reduction of 29.4% in model parameters, a decrease of 28.5% in model size, a decrease of 24.6% in GFLOPs, an increase of 3.1% in mAP
50
, an increase of 2.5% in mAP
50:95
, and an FPS (Frames Per Second) of up to 65.7.