光照变化条件下的人脸识别研究

发布时间：2018-06-26 03:49

本文选题：人脸识别 + 光照归一化　；参考：《宁夏大学》2016年硕士论文

【摘要】：随着信息技术的发展,人脸识别不仅在门禁系统有广泛应用,案件侦破、电子商务等更多领域也应用了人脸识别系统,所以人脸识别在当今社会中的研究价值很高。人脸识别在日常生活中经常会受到一些外界因素的干扰,如：表情、遮挡物、光照和姿态,人脸识别系统在采集人脸信息时通常处于不可控的光源环境中,所以如何更快更简洁的去除光照变化在人脸识别中的影响是未来人脸识别的一个重要研究领域。文章主要研究人脸识别中的光照变化领域,主要研究内容与创新点如下：(1)本文深入研究了几种光照归一化的处理算法,如：对数变换、伽马校正、直方图规定化和直方图均衡化,将这几种算法分别在Yale B人脸库中进行仿真实验。结果证明,光照归一化的处理算法虽然简便,在一定程度上削弱了光照影响,但在光照条件较差的情况时,只采用基于光照归一化的处理算法去除光照效果不明显。(2)本文重点研究了自商图像、多尺度Retinex算法和梯度脸这三种提取光照不变特性的算法。三种算法均在Yale B人脸数据库下进行测试试验,测试表明这三种算法均能很好的提取出入脸图像中的轮廓和纹理等光照不变特征,有明显的消除光照的作用。然而图像经过MSR算法的处理,人脸图像的各个细节保留不完整；SQI算法在低信噪比处扩大了高频噪声,加权的高斯滤波器很难保留完整的边缘信息,算法中的参数选择也比较复杂；而梯度脸算法不仅计算简单,而且提取的光照不变特性信息较完整,且最终的识别率也要好于另外两种光照不变特征提取算法。(3)研究发现人脸图像有噪声的情况下,噪声会严重干扰梯度信息,从而使识别率大大下降。传统的降噪方法在降噪时会模糊图像,降低了图像的梯度信息从而影响识别率。本文将非下采样轮廓波变换(NSCT)与梯度脸提取光照不变特征的方法相结合,提出了NSCT-Gradientfaces算法,在有噪声的人脸图像上有很好的识别结果。将本文算法与高斯滤波、小波变换和Contourlet变换与梯度脸算法结合,并在Yale B人脸库中测试,实验结果表明,本文算法的识别性能比前面几种结合算法更好。
[Abstract]:With the development of information technology, face recognition is not only widely used in access control system, but also applied in many fields, such as case detection, electronic commerce and so on. Therefore, face recognition is of great value in today's society. Face recognition is often disturbed by some external factors in our daily life, such as expression, occlusion, illumination and pose. When collecting face information, face recognition system is usually in an uncontrollable light source environment. So how to remove the influence of illumination change more quickly and succinctly is an important research field of face recognition in the future. The main contents and innovations of this paper are as follows: (1) in this paper, several illumination normalization algorithms, such as logarithmic transformation, gamma correction, are studied in depth. Histogram specification and histogram equalization are simulated in Yale B face database. The results show that the illumination normalization algorithm is simple and weakens the illumination effect to some extent, but in the case of poor illumination conditions, Only using the illumination normalization algorithm to remove illumination is not obvious. (2) this paper focuses on the self-quotient image, multi-scale Retinex algorithm and gradient face to extract the illumination invariant characteristics of the three algorithms. The three algorithms are tested in Yale B face database. The test results show that the three algorithms can extract the illumination invariant features such as contour and texture in the face image, and can eliminate the illumination obviously. However, when the image is processed by MSR algorithm, every detail of the face image remains incomplete and the squi algorithm expands the high frequency noise at low signal-to-noise ratio, so it is difficult for weighted Gao Si filter to retain complete edge information. The selection of parameters in the algorithm is also complex, and the gradient face algorithm is not only simple to calculate, but also the extracted information of illumination invariance is relatively complete. And the final recognition rate is also better than the other two illumination invariant feature extraction algorithms. (3) it is found that when the face image has noise, the noise will seriously interfere with the gradient information, thus the recognition rate will be greatly reduced. The traditional denoising method will blur the image and reduce the gradient information of the image, thus affecting the recognition rate. In this paper, the non-downsampling contour wave transform (NSCT) is combined with the method of extracting illumination invariant features from gradient face, and the NSCT-Gradientfaces algorithm is proposed, which has good recognition results on noisy face images. The algorithm is combined with Gao Si filter, wavelet transform and Contourlet transform with gradient face algorithm, and tested in Yale B face database. The experimental results show that the performance of this algorithm is better than that of the previous algorithms.
【学位授予单位】：宁夏大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】