An enhanced YOLOv8 with C2f-CFA module for basketball target detection

Basketball target detection is crucial for sports video analysis and intelligent refereeing. However, complex game scenarios, including target size variations, occlusions, lighting interference and background interference, make it difficult for traditional algorithms to balance detection accuracy and speed. While YOLOv8 performs excellently in general scenarios, it has problems such as insufficient information interaction, weak long-range dependency modeling and feature redundancy, which limit its detection performance for fast-moving and small-sized basketballs. This paper proposes the YOLO-CFA model, an enhanced version of YOLOv8. Its core is the introduction of the C2f-CFA module, which replaces the original C2F module in the Backbone and Neck. This module integrates a multi-head Bottleneck structure (used to enhance local features) and a Cross-Dimensional Feature Attention mechanism (used to model cross-channel and cross-position contextual relationships), while retaining the lightweight design of YOLOv8. Experimental results on a customized dataset show that YOLO-CFA outperforms the YOLOv5 to YOLOv12 series models. Its F1- score and mAP50:95 reach 86.0% and 61.0% respectively, which are 5.2% and 8.1% higher than those of the baseline model YOLOv8. In real-scene tests, the model reduces false detections and missed detections while maintaining high real-time performance, providing a scalable solution for target detection tasks in the sports field.