MHD-Fine: a multidimensional heterogeneous feature extraction network for UAV small object detection based on D-Fine

To overcome the issues of high computational complexity and poor small object detection performance in UAV images, this paper proposes a Multi-dimensional Heterogeneous feature extraction network based on D-Fine (MHD-Fine). Firstly, a Multi-dimensional Heterogeneous Convolution (MHConv) method is proposed. Through multi-range coverage of spatial scales and differential feature extraction of channel subsets, this method reduces computational redundancy while providing more comprehensive and diverse feature support. Secondly, the Shuffle Attention mechanism (SA) and MHConv are integrated into the neck network. By leveraging SA’s ability to adaptively focus on key feature regions and fusing it with MHConv’s multi-dimensional feature extraction capability, this integration enhances the network’s attention to small object details while suppressing background interference. Finally, the Enhanced Intersection over Union (EIoU) loss function is introduced. By simultaneously optimizing the overlap degree, center distance, and aspect ratio of bounding boxes during calculation, this function further improves the regression accuracy of UAV small object bounding boxes. Experimental results show that on the VisDrone UAV aerial photography dataset, the proposed model improves the mAP50 metric by 1.3% compared with the baseline model, while reducing the number of parameters by 35%.