Decoupled Region of Interest Feature Pooling Diffusion Network for UAV Image Object Detection

Because of the complex background, more small targets, and large scale variation in drone remote sensing images, the existing methods face challenges in effectively extracting image features, resulting in suboptimal detection performance. This work proposes a diffusion model detection framework based on decoupled region of interest pooling(DRIP) and a sequence of large convolutional kernel networks. Firstly, the diffusion model with randomly generating noisy bounding boxes is applied to replace the Region Proposal Network (RPN) in the two-stage detection algorithm framework. Secondly, a sequence of large convolutional kernel backbone is employed to extract distinctive foreground features of the remote image. Lastly, a more rational DRIP feature extraction module has been designed, which decouples the feature extraction process between different scales, to make full use of the features at each level of the pyramid. Our approach eliminates the negative effects of unreasonable bounding box assignments, thereby enhancing the overall performance. The experimental results demonstrate that the proposed method has achieved a maximum improvement of 3% in mean average precision (mAP) compared with the original method in the UAV image dataset VisDrone-2019. While the parameter count has a significant decreased.