基于Lucas-Kanade算法的最大Gabor相似度大姿态人脸识别

Pose Invariant Face Recognition Using Maximum Gabor Similarity Based on Lucas-Kanade Algorithm

在人脸识别科学研究和实际应用领域中,大角度姿态是影响人脸识别结果的主要因素之一,成为限制人脸识别技术进步的难点,而姿态的校正归一化是解决该问题的常用手段。首先通过加权的LK(Lucas-Kanade)算法得到侧脸块和对应正脸块的仿射变换参数,基于最大Gabor相似度寻找校正人脸姿态的最优参数。然后,以每一人脸块最优参数得到的平均Gabor相似度作为这一块人脸的识别权重,可以增加大姿态人脸识别的精度和稳健性。在FERET人脸数据库中进行了实验,当水平偏转角度为45°时,准确率达到97.3%,证明本文提出的以最大Gabor相似度作为加权LK算法参数提取的依据是有效的,得到的最优参数具有较好的光照无关性,而将平均Gabor相似度作为识别权重,有助于使算法的应用更加稳健和有效。

In the field of face recognition, pose variation is one of the significant challenges that affects the recognition performance and has been one of the major obstacles hindering the improvement of the face recognition technology. In this study, affine transformation parameters of side face and full-frontal face patches are obtained by applying the weighted Lucas-Kanade(LK) algorithm. We further propose that an optimal parameter for correcting face pose can be obtained based on the maximum Gabor similarity. Furthermore, the average Gabor similarity acquired from the optimal parameter of each face patch can be considered to be the face recognition weight, improving the recognition rate and enhancing the robustness of the pose invariant face recognition. Finally, the experimental results obtained based on the FERET face database denote that the recognition rate for the image with a pose of 45° can reach up to 97.3%, indicating that the usage of the maximum Gabor similarity as a basis for parameter extraction of the weighted LK algorithm is valid. This method can also handle illumination variations. Considering the average Gabor similarity as the recognition weight will ensure the robust and effective application of this algorithm.