动态时间规整优化局部分块匹配的人脸识别

Local partition matching optimized by DTW for face recognition

针对现实人脸识别中姿势、光照、表情变化及遮挡等严重影响识别性能的问题,提出了一种基于动态时间规整优化局部分块匹配的户外人脸识别算法。将人脸图像划分成若干大小相等且互不重叠的局部小块;借助于光栅扫描顺序将各个小块按照前额、眼睛、鼻子、嘴巴和下巴的顺序连接成一个单一序列;计算查询人脸与注册人脸之间图像到类的距离,利用动态时间规整的设计思想寻找查询序列与所有注册序列之间的最佳对齐方式。在三个公开人脸数据库LFW、AR及YouTube上的实验验证了该方法的有效性及可靠性,实验结果表明,相比其他几种较为先进的人脸识别方法,该方法取得了更高的识别率,此外,该方法无需任何训练过程,计算成本低。

Large variation of pose, illustration, expression and occlusion in truly face recognition will seriously impact recognition performance, for which Local Partition Matching（LPM） algorithm optimized by Dynamic Time Warping （DTW）is proposed. Face image is divided into many non-overlapping patches with same size. All patches are combined to be a unique sequence sorting by forehead, eyes, nose, mouth and chin by using raster scan sequence. Distance from im-age to class between query face and register faces is calculated, and idea of DTW is used to find the best alignment be-tween query sequence and all register sequences. The effectiveness and reliability of proposed method have been verified by experiments on the three common databases LFW, AR and YouTube. Experimental results show that proposed method has higher recognition accuracy than several advanced face recognition methods. Besides, it has lower cost without any training process.