基于深度残差网络和YOLO的人物识别系统

Character Recognition System Based on Deep Residual Network and YOLO

通过残差网络和多尺度的图像训练来提高人头识别的精度,人头特征的检测有两种方法,以往是基于事先描绘好的人头特征,还有基于统计训练模型的方法,后者具有很好的鲁棒性,用神经网络去训练,可以得到比较好的效果;基于密集的多人图像存在重叠的人头和远近尺度不一的人头特征,这就需要对人物头像的训练样本做更多特殊的处理;对于不能实时检测的问题,采用另外一种YOLO检测的方法,进行缺陷的弥补,获得相关的应用模型。

Mainly uses the residual network and multi-scale image training to improve the accuracy of human head recognition. There are two meth- ods for the detection of human head features. The past is based on the previously described good human head features, and the current method is based on statistical training models. The current method has good robustness. Training with neural network can obtain better re- sults. For dense multi-person images, there are overlapped heads and human head features with different scales. This requires us to do more special treatment of the training samples of the character portraits.