CLAHE低频DCT系数重变换在人脸识别中的应用

APPLICATION OF LOW FREQUENCY DCT COEFFICIENTS RETRANSFORMING WITH CLAHE IN FACE RECOGNITION

针对光照变化显著影响自动人脸识别系统性能的问题,为了规范化光照变化以提高人脸识别率,提出基于对比度限制自适应直方图均衡化CLAHE(Contrast Limiting Adaptive Histogram Equalization)的低频离散余弦变换DCT(Discrete Cosine Transform)系数重变换算法。首先将图像划分成多个互不重叠的局部小块,使用CLAHE对局部小块进行局部对比拉伸以实现去噪;然后,通过缩减适当数目的低频DCT系数来消除人脸图像中的光照变化;最后,利用核主成分分析进行特征提取,稀疏系数重建和k-近邻分类器完成最终的人脸识别。在扩展Yale B及AR人脸数据库上的实验验证了算法的有效性。实验结果表明,算法在识别非常困难的Yale B子集5上的识别率可高达98.20%,相比其他几种规范化技术,该算法取得了更高的识别率,同时大大降低了识别所耗时间。

For the problem that the illumination variations impact the performance of automated face recognition system significantly, we proposed the low frequency discrete cosine transform （DCT） coefficients retransforming algorithm, it is based on contrast limiting adaptive histogram equalisation （CLAHE） and is for standardising the illumination changes so as to improve face recognition accuracy. Firstly, we divided the original images into multiple non-overlapping local patches and used CLAHE to do local contrast stretching on them so as to reduce noises. Then, we removed the illustration variation on face image by holding down the proper numbers of low frequency DCT coefficients. Finally, we used kernel principle component analysis to extract features, and used sparse coefficient reconstruction and k-nearest neighbour classifier to eventually complete the face recognition. The effectiveness of the proposed algorithm was verified by experiments on extended Yale B and AR face datasets. Experimental results showed that the proposed algorithm reached the recognition accuracy as high as 98.20% on Yale B subset 5 which is extremely difficult in recognition, and it achieved the higher recognition rate and less recognition time than several other standardised technologies as well.