基于YOLOv4改进算法的乒乓球识别

传统的基于颜色分割的乒乓球识别方法易受光线、清晰度影响,鲁棒性较低。为解决这一问题,对YOLOv4进行改进,用K-means聚类设计先验框,增强尺度适应性。针对乒乓球尺寸,裁剪网络分支并压缩卷积层,加快特征提取速度。针对采集数据正负样本不均衡,改进损失函数,提高预测框边界准确度。使用快速NMS算法加速预测过程,提高模型的计算速度。实验结果表明,基于YOLOv4的改进模型在乒乓球识别任务中精度达到94.12%,帧处理速率达到39.34fps。

The traditional table tennis recognition method based on color segmentation is easily affected by light and clarity,and its robustness is low.In order to solve this problem,YOLOv4 is improved and a priori frame is designed by K-means clustering to enhance the scale adaptability.Aiming at the size of table tennis,the network branches are cut and the convolution layer is compressed to speed up the speed of feature extraction.In view of the imbalance between positive and negative samples of collected data,the loss function is improved to enhance the accuracy of the boundary of the prediction box.The fast NMS algorithm is used to accelerate the prediction process and increase the calculation speed of the model.The experimental results show that the accuracy of the improved model based on YOLOv4 in table tennis recognition task is 94.12%,and the frame processing rate is up to 39.34fps.