基于小波域多尺度Retinex的复杂光照的人脸识别

摘要

针对Retinex模型在处理光照不均匀图像的不足,提出了组合小波域多尺度Retinex模型(DwT-MSR)和ICA识别方法,并将其用于不同光照下的人脸识别.在二维小波变换后的小波域中,将其低频小波系数变换到对数空间,使用三种不同的高斯滤波系数和对数空间中的小波系数进行卷积运算,将三种标准偏差尺度下得到的结果进行加权平均,采用gain/offset的方法,对输出图像进行灰度值线性拉伸；小波域中其他三种高频系数保持不变,然后再将交换后的低频系数和高频系数作小波反变换,得到的新图像则为小波域多尺度Retinex模型的处理结果,最后使用定点独立分量分析和神经网络进行分类识别.经实验证明,基于该模型的方法在处理不同光照下的人脸图像时,效果明显优于Grey,Hist,SSR,Embossing,MSR,Quotient等常见的光照处理方法.%Multi-Scale Retinex ( MSR) in discrete wavelet transform model is presented to make up disadvantages of Retinex model while dealing with face images with shadows. In the wavelet domain, low frequency wavelet coefficients are translated to log domain, and convolution operation is implemented between it and three types of different Gaussian filters coefficients. A weighted average of multi scale is gained and mapped to gray range of display device by method of gain/offset Other three high frequency wavelet coefficients are remained. The result of MSR in discrete wavelet transform model is gained by IDWT of the weighted average result and other three high frequency wavelet coefficients. Finally, the translated images are recognized by fixed point Independent Component Analysis (ICA) and RBF. Experimental results with hypothesis testing show that face recognition system based on DWT-MSR model is superior to other methods in dealing with face images under various illuminations.