Multi-Branch Parallel Networks for Object Detection in High-Resolution UAV Remote Sensing Images

Uncrewed Aerial Vehicles (UAVs) are instrumental in advancing the field of remote sensing. Nevertheless, complexity background and dense distribution objects both present considerable challenges for object detection UAV sensing images. This paper proposes a Multi-Branch Parallel Network (MBPN) based on ViTDet (Visual Transformer Object Detection) model, which aims to improve accuracy Initially, discriminative ability input feature map Feature Pyramid (FPN) is improved by incorporating Receptive Field Enhancement (RFE) Convolutional Self-Attention (CSA) modules. Subsequently, mitigate loss semantic information, sampling process FPN replaced Upsampling (MBUS) Downsampling (MBDS) Lastly, Feature-Concatenating Fusion (FCF) module employed merge maps varying levels, thereby addressing issue misalignment. evaluates performance proposed model custom UAV-captured WCH dataset publicly available NWPU VHR10 dataset. The experimental results demonstrate that achieves an increase APL 2.4% 0.7% datasets, respectively, compared baseline ViTDet-B.