梯度-LBP优化深度图像分析的性别人脸识别

Depth image analysis optimized by gradient-LBP for gender face recognition

针对目前最先进的3DLBP人脸识别算法中仍存在特征长度大、编码不稳定等固有缺陷,提出了一种基于梯度LBP的深度图像分析算法。从各种不同方向视觉化LBP算子,计算相邻像素的深度差,产生多个有导向的深度差图像,串联合并各个深度差直方图信息,形成唯一有导向的深度差直方图。在Kinect和范围扫描仪数据库图像上的所有实验均证明了所提描述符优于3DLBP。此外,还加权合并所提描述符和灰度图像的LBPU2,在高质量3D范围扫描仪数据库图像(Texas 3DFR)和Kinect设备采集的低质量图像(EURECOM Kinect人脸数据库)上的总体平均识别率可高达96.70%。

According to the present that many inherent defects such as characteristic length and coding instability exist in the most advanced 3DLBP face recognition algorithm,this paper proposed a depth image analysis optimized by gradient-LBP.First-ly,it visualized the LBP operator from different directions.Then,it calculated the adjacent pixels’depth differences,which would produce multiple guided images of depth differences.Finally,it tandem combined histogram information of each depth difference,formed a unique oriented depth difference histogram.All experiments on the images of Kinect and range scanner im-age database prove that the proposed descriptor is better than 3DLBP.In addition,this paper also proposed the weighted combi-nation of the proposed descriptor and gray image LBPU2.Recognition accuracy of proposed descriptor on the high quality 3D range image scanner database （Texas 3DFR）and the low quality images collected by Kinect （EURECOM Kinect face data-base）experiment equipments can achieve 96.70%.