Tennis Ball Recognition in Complex Scenes Based on Improved YOLOv5

In order to solve the problems of low accuracy of tennis recognition, oversized model and low detection frame rate in complex environment, this paper proposes a method based on improving YOLOv5s network. The method improves tennis recognition in complex scenes by improving the loss function, adding a convolutional attention mechanism, and designing a new feature extraction network to extract features. The experimental results show that the mean average precision (mAP) of the improved tennis ball detection model reaches 87.73% under the same experimental conditions, which is 9.28% better than the original model of YOLOv5s. And the detection speed of the model can reach 69 frames/s, and the size of the model is only 12.1MB, which can meet the requirements of real-time and light weight for tennis ball recognition while greatly improving the accuracy.