RT-DETR-iRMB: A Lightweight Real-Time Small Object Detection Method
Zhuofan Yu · 2024
In order to solve the difficulty of small object detection, this paper takes RT-DETR (Real-Time object DEtection model based on TRansformer) as the baseline model, and optimates its backbone network and IoU loss to improve the detection effect of this model on small targets. Specific improvements are as follows: In this paper, the backbone network of RT-DETR is lightweight and improved, and iRMB (Inverted Residual Mobile Block) is used to replace the original module. Then, in order to solve the problem of poor effect and slow convergence in the classification task of fine-grained images in the data set, an improvement is proposed for the original loss function of RT-DETR model. In this paper, MPDIoU loss function is proposed to replace GIoU loss function to calculate the regression loss of prediction frame. Compared with 57GFLOPs of the benchmark model, the number of parameters of the improved RT-DETR-iRMB decreases to 50.2GFLOPs, which is reduced by 11.9%, and the confusion between similar categories is reduced, mAP value is increased by 0.2%, FPS is increased by 3.7, which proves the effectiveness of the improved model.