Basketball Small Object Detection Algorithm Based on Improved YOLO v7

Intelligent sports have emerged as a focal area of computer vision research. With a broad public audience around the world, basketball has gained significant attention as one of the most attractive application scenarios. However, achieving a balance between speed and accuracy in basketball detection poses great challenges due to the small size and rapid movement of the basketball target, as well as the occlusion caused by players. To address this issue, we propose a small object detection algorithm based on an improved YOLO v7. Firstly, to enhance the feature extraction capability for small targets, a Multi-scale Dilated Attention Feature Extraction Module (MDAF) is introduced. Subsequently, a Feature Denoising Enhancement Module (FDEM) is designed to obtain enhanced feature representations for small targets. Finally, the NWD is used instead of IoU to reduce the sensitivity of small target detection boxes to positional deviations. Together with the strategies above, our method achieves an improvement of 7.7 % and 8.6 % in mAP@ 0.5 on a self-customized basketball dataset and the public Tsinghua-Tencent 100 K dataset, respectively, while also meeting real-time requirements with a low latency, which demonstrates the effectiveness of our method for small object detection.