Small Target Detection with Context Fusion Learning

The detection of targets in UAV images has emerged as a prominent research area in computer vision. Accurate target detection is crucial for various UAV applications. However, challenges arise due to the presence of small targets and complex backgrounds in UAV perspectives, posing difficulties for target detection performance. To address this challenge, this paper proposes a precise small target detection model, ADA_YOLO, specifically designed for the UAV perspective. Firstly, the model incorporates the Global Attention Mechanism (GAM) to enhance sensitivity to small targets. Next, the C2f_ DCN module is designed to provide the model with enhanced adaptability to geometric variations, strengthening the extraction capabilities for small target features and effectively improving the detection performance of the network model. Additionally, the ASPP module is integrated into the model to leverage contextual information in images, enhancing the fusion of deep and surface semantic information, mitigating issues of false positives and false negatives. Finally, improvements are made to the detection head structure, enabling the model to excel in small target detection. The experimental results on the VisDrone2019 dataset demonstrate that the improved model achieves a 7.4% increase in mAP50 compared to the baseline model, showcasing a significant improvement over the contrastive algorithms.