Abstract:

Aiming at the problem that the table tennis object detection method was susceptible to interference from various factors such as environment， light， and speed， resulting in poor accuracy and real-time， a lightweight table tennis object detection algorithm Shuffle-YOLOv5s （SYOLO5） was proposed based on the YOLOv5s framework. Firstly， YOLOv5s backbone network was reconstructed by improved ShuffleNetV2 network unit combination to speed up feature extraction. Secondly， the Efficient Channel Attention （ECA） in the process of feature fusion was introduced to improve the detection performance. Thirdly， the convergence speed and positioning accuracy of the network were improved by using SIoU （S-Intersection over Union） Loss as the positioning loss function. Finally， dual-scale object detection method was adopted to further improve the reasoning speed of the model based on the small size characteristic of table tennis. Experimental results show that compared with YOLO5s， SYOLO5 reduces the parameter amount and calculation amount by 80% and 60% respectively， and increases precision by 1.9 percentage points.

Key words: table tennis detection, YOLOv5s, ShuffleNetV2, Efficient Channel Attention (ECA), SIoU (S-Intersection over Union) Loss

摘要：

针对乒乓球目标检测方法易受环境、光线、速度等多种因素干扰导致精度和实时性不佳的问题，提出了一种基于YOLOv5s框架的轻量化乒乓球目标检测算法——SYOLO5（Shuffle-YOLOv5s）。首先，采用改进的ShuffleNetV2网络单元组合重构YOLOv5s主干网络，提高特征提取速度；其次，在特征融合的过程中引入高效通道注意力（ECA）机制，有效提升模型的检测性能；接着，采用SIoU Loss（S-Intersection over Union）作为定位损失函数提升网络的收敛速度和定位精度；最后，贴合乒乓球小尺寸的特点，采用双尺度目标检测，进一步提高模型推理速度。实验结果表明，所提算法与YOLOv5s相比，参数量和计算量分别减少了80%和60%，精确率提升了1.9个百分点。

关键词: 乒乓球检测, YOLOv5s, ShuffleNetV2, 高效通道注意力, SIoU Loss