Improved Faster R-CNN Detection Algorithm for Small Unmanned Aerial Vehicle Targets

The small size of Unmanned Aerial Vehicles (UAVs) targets make the detection model challenging to extract texture features, resulting in low detection accuracy when using deep learning methods. To address this issue, this work proposes a Spatial attention Deformable Pyramid networks (SDPnet) based on Faster R-CNN framework. Combined with the ResNet-50 residual network and the Feature Pyramid Network(FPN) in the proposed model, the spatial expression ability of the bottom-layer features is improved. We also design a deformable convolution mechanism to learn richer feature representations with stronger generalization ability by adding offsets. Additionally, the spatial attention mechanism is added to suppress background interference and highlight important features. The experimental results demonstrate that, compared with the Faster R-CNN model, the SDPnet achieved a precision rate increase of 57.34% and a total error rate decrease of 40.51% on the self-collected dataset. Furthermore, when compared to YOLOv5, SDPnet also exhibits relative superiority.