Lightweight and Hough transform collaborative advancement detection algorithm for small tennis targets

Aiming at the excessive parameter volume and severe missed detection defects of tennis small-object detection algorithms in intelligent sports visual perception, this paper constructs a lightweight multi-scale detection network guided by Hough prior knowledge. Built upon the YOLOv8s baseline, the network adds a brand-new high-resolution feature detection branch at the bottom layer. Its backbone adopts two-stage phantom-based feature transformation, embeds a parallel channel-spatial hybrid attention module and a focus modulation unit for adaptive feature recalibration. Furthermore, the Hough transform is integrated to extract regular circular geometric features of tennis balls. In post-processing, a flexible NMS strategy with distance-IoU metric and center spatial distance penalty term is introduced. Comparative experiments reveal that the proposed network achieves a 1.8% higher mAP@0.5 than the baseline and reaches 106 FPS. On the self-built dataset, the mAP of tiny tennis targets hits 78.6%, representing a 16.2% promotion over the original model. For heavily occluded dense scenes, the improved NMS maintains mAP at 84.5%, with single-image end-to-end post-processing latency merely 1.5 ms. The above results verify that the proposed method efficiently eliminates redundant deep features and strikes a favorable trade-off among lightweight architecture, low computational overhead and high detection precision. This work provides an efficient machine vision cooperation scheme, and provides a promising lightweight solution for the potential edge deployment of intelligent edge computing devices for mobile small object detection tasks.