MSEIGDNet: a muti-scale edge information generation and diffusion network for UAV aerial images

With the rapid development of unmanned aerial vehicle (UAV) technology, object detection in UAV aerial images has attracted increasing attention. However, this task faces numerous challenges, including scale imbalance, complex backgrounds, and dense occlusions. Therefore, this paper proposes a Multi-Scale Edge Information Generation and Diffusion Network (MSEIGDNet). Specifically, a Multi-Scale Edge Information Generation Network (MSEIGNet) is designed to significantly improve the model's ability to perceive small object contours. In addition, a Shared Feature Pyramid Module (SFPM) is introduced, which leverages weight sharing and convolutional structures with different dilation rates to enhance multi-scale feature representation. Furthermore, an Aggregation and Diffusion Feature Fusion Network (ADFFN) is constructed to improve semantic consistency between shallow and deep layers through aggregation and diffusion mechanisms, thereby enhancing semantic expression for small object detection. Moreover, an Adaptive Power Transformation (APT) function is proposed to dynamically adjust IoU weights during the regression phase, reinforcing the learning of low-quality prediction boxes and mitigating localization shifts under dense occlusion conditions. Experimental results on the VisDrone, UAVDT and CoDrone datasets demonstrate that compared to the baseline model, the proposed method improves mAP50 by 7.6%, 3.7% and 3.4%, respectively. These innovations provide an efficient and robust solution for small object detection in complex UAV scenarios, and effectively promote the development and practical deployment of object detection technology in low-altitude visual environments.