双域特征融合和重校准小目标检测网络OA

To address the challenges in small target detection,such as limited detection regions,sparse feature pixels,and poor recognition performance,this paper proposes an improved small target detection algorithm DFFR-DETR based on RT-DETR.Firstly,a spatial-frequency dual-domain feature fusion module is proposed to improve the backbone network,named SDFF-Net,which replaces the original backbone network to extract the structural features of small targets in the image and enhance the ability to capture information about small targets.Secondly,the introduction of the recalibration attention unit and the RepConv design of the Boundary Aggregation Reparameterization(BAR)module enhances the network's ability to fuse and extract multi-scale features.Finally,based on the BAR module,a multi-scale feature recalibration model based on convolutional neural net-works is designed to enhance the extraction ability of global features and further improve the performance of small target detec-tion.The experimental results on the VisDrone2019 dataset show that the DFFR-DETR model achieves mAP50 scores of 52.6% and 41.3% on the validation set and test set respectively,which are 5.1 percentage points and 4.0 percentage points higher than the baseline model.Additionally,precision and recall rates also demonstrate notable enhancements.The generalization experi-ments are conducted on the TinyPerson dataset.Compared to the baseline model,the DFFR-DETR model achieved an increase of 3.9 percentage points in recall rate and 2.3 percentage points in mAP50,verifying the effectiveness and generalization ability of the improved model.