基于DTW的LPM在面部遮挡人脸识别中的应用

针对面部遮挡人脸识别问题中遮挡源变幻莫测、遮挡位置、大小及形状未知而导致从人脸图像中分割遮挡区域困难的问题,提出了基于动态时间规整(DTW)的局部分块匹配(LPM)算法。首先,将人脸图像划分成若干大小相等且互不重叠的局部小块；然后,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；最后,计算查询人脸与注册人之间图像到类的距离,并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。在两个公开人脸数据库FRGC2.0、AR及一个户外人脸数据库上的实验验证了所提算法的有效性及可靠性,实验结果表明,相比其它几种较为先进的人脸识别算法,所提算法取得了更高的识别率,此外,所提算法无需任何训练过程,计算成本低,更适合应用于现实面部遮挡人脸识别中。

Application of LPM Based on DTW in FR with Facial Occlusion

For the issue that occlusion source is variable, occlusion position, size and shape is unknown in face recognition with facial occlusion which causes it is difficult to segment occlusion area from face image, local partition matching algorithm optimized by dynamic time warping (DTW) is proposed. Firstly, face image is divided into many non-overlapping patches with same size. Then, all patches are combined to be a unique sequence sorting by forehead, eyes, nose, mouth and chin by using raster scan sequence. Finally, distance from image to class between query face and register faces is calculated, and idea of DTW is used to find the best alignment between query sequence and all register sequences. The effectiveness and reliability of proposed algorithm has been verified by experiments on the two common databases FRGC2.0, AR and a wild face database. Experimental results show that proposed algorithm has higher recognition accuracy than several advanced face recognition algorithms. Besides, none training process is necessary in proposed algorithm and low cost of which indicates it is more suitable to be applied into real face recognition with facial occlusion.