MHNet: Multi-scale Hierarchical Extraction Network for Small Object Detection in UAV Images

Unmanned aerial vehicle (UAV) image have the characteristics of small object sizes, dense distributions, and complex backgrounds. Existing detection methods perform well under normal circumstances, but perform poorly when processing UAV images. In this paper, we propose MHNet, a small object detection framework for UAV images, which solves the above problems through multi-scale feature processing and more efficient feature fusion. First, we design a multi-scale hierarchical convolution (MHC) module that extracts features at different scales, layer by layer, providing finer-grained feature information and a larger receptive field. Second, we designed the SPPFC module to capture the multi-scale features extracted by the backbone. We introduce contextual anchor attention (CAA) in the SPPFC module to bolster contextual dependency and fortify feature information across various scales, thereby augmenting the semantic information of high-level features. At the same time, this paper uses an auxiliary detection head, combined with a new feature fusion architecture to improve the prediction ability of small objects. The CAA module downsample the input features of the auxiliary detection head to enhance the feature information of the other two detection heads. This design effectively promotes the fusion of high-level and low-level information. Multiple experiments on VisDrone2019 and UAVDT have demonstrated the effectiveness of MHNet. On VisDrone2019, with mAP and mAP50 reaching 28.2% and 45.8%, respectively. Compared with the benchmark, our MHNet improves mAP and mAP50 by 4.8% and 6.7%, respectively.