一种改进YOLOv3的手势识别算法

为了解决YOLOv3算法在手势识别中存在识别精度低及易受光照条件影响的问题,提出了一种改进的YOLOv3手势识别算法。首先,在原来3个检测尺度上新增加1个更小的检测尺度,提高对小目标的检测能力;其次,以DIoU代替原来的均方差损失函数作为坐标误差损失函数,用改进后的Focal损失函数作为边界框置信度损失函数,目标分类损失函数以交叉熵作为损失函数。结果表明,将改进的YOLOv3手势识别算法用于手势检测中,mAP指标达到90.38%,较改进前提升了6.62%,FPS也提升了近2倍。采用改进的YOLOv3方法训练得到的新模型,识别手势精度更高,检测速度更快,整体识别效率大幅提升,平衡了简单样本和困难样本的损失权重,有效提高了模型的训练质量和泛化能力。

A gesture recognition algorithm based on improved YOLOv3

In order to solve the problems of low recognition accuracy and easily affected by illumination conditions in the gesture recognition, an improved YOLOv3 gesture recognition algorithm was proposed. Firstly, a smaller detection scale was added to the original three detection scales to improve the detection ability of small targets; secondly, DIoU was used instead of the original mean square error loss function as the coordinate error loss function, the improved focal loss function was used as the confidence loss function of the boundary frame, and the cross entropy was used as the loss function of the target classification loss function. The results show that when the improved YOLOv3 gesture recognition algorithm is applied to gesture detection, the map index reaches 90.38%, which is 6.62% higher than that before the improvement, and FPS is nearly twice as high as before. After the new model is trained by the improved YOLOv3 method, the gesture recognition accuracy is higher, the detection speed is faster, the overall recognition efficiency is greatly improved, the loss weights of simple samples and difficult samples are balanced, and the training quality and generalization ability of the model are effectively improved.