人脸识别中光照补偿方法的研究及FPGA实现

发布时间：2018-07-13 18:43

【摘要】：人类主要依赖视觉系统获取外界信息,一旦视觉通路发生病变或者损伤就会导致人类视力的衰弱甚至失明,作为一种新型的辅助设备,视觉假体能够通过电刺激视觉神经系统使患者产生光幻视,帮助盲人重见光明。在视觉假体的辅助下,盲人是可以完成日常生活中的人脸识别任务的,然而在实际应用中复杂的光照变化会导致人脸识别率的下降,因此对光照变化下的人脸图像进行光照补偿具有很重要的意义。本文针对单尺度Retinex算法的光晕缺陷,从硬件实现的角度上提出了改进的基于双边滤波的单尺度Retinex算法。在用双边滤波对图像的光照分量进行估计的基础上,对对数域中图像反射分量进行自适应亮度调整。通过将经典的单尺度Retinex算法(SSR)、多尺度Retinex算法(MSR)、基于双边滤波的SSR算法和本文改进算法进行MATLAB仿真。试验结果表明改进算法能够有效地消除各种条件下的光照影响,具有较好的光照鲁棒性,并且与经典的MSR算法相比亮度提升了 26%,对比度提升了 23%,在ExtendedYaleB人脸库上的平均人脸识别率较经典的Retinex算法提升了 13%,在人脸识别系统的应用中具有很强的适用性。本文完成了光照补偿算法的FPGA硬件实现,并利用ModelSim对各个模块进行了功能仿真。然后在DE2-115开发板上搭建相关硬件平台对算法进行了功能性验证。为了直观地验证算法对不同图像的处理效果,本文通过串口输入数据到算法处理模块,并用VGA显示处理后的图像。同时通过SignalTap Ⅱ捕获算法处理的图像数据与matlab的软件结果对比。最终硬件验证结果表明,本文提出的算法在硬件上取得了与软件相同的处理效果,能够在消除图像光照影响的同时,显现增强人脸暗部区域的轮廓细节。
[Abstract]:People rely mainly on the visual system to obtain external information. Once the visual pathway changes or damages, it will lead to human vision decline or even blindness, as a new type of auxiliary equipment. Visual prostheses can help blind people rediscover light by electrical stimulation of the visual nervous system. With the aid of visual prosthesis, blind people can complete the task of face recognition in daily life. However, in practical applications, complicated illumination changes will lead to the decline of face recognition rate. Therefore, it is very important to compensate face images with illumination. Aiming at the halo defect of single-scale Retinex algorithm, an improved single-scale Retinex algorithm based on bilateral filtering is proposed from the point of view of hardware implementation. On the basis of estimating the illumination component of the image by using bilateral filtering, the image reflection component in logarithmic domain is adjusted adaptively. The classical single-scale Retinex algorithm (SSR), the multi-scale Retinex algorithm (MSR), the two-sided filter based SSR algorithm and the improved algorithm are simulated in MATLAB. The experimental results show that the improved algorithm can effectively eliminate the illumination effect under various conditions, and has better illumination robustness. Compared with the classical MSR algorithm, the brightness and contrast are improved by 26% and 23% respectively. The average face recognition rate in extended YaleB face database is 13% higher than that in the classical Retinex algorithm. It has strong applicability in the application of face recognition system. In this paper, FPGA hardware implementation of illumination compensation algorithm is completed, and each module is simulated by ModelSim. Then build the related hardware platform on the DE2-115 development board to verify the function of the algorithm. In order to intuitively verify the processing effect of the algorithm to different images, this paper input data into the algorithm processing module through serial port, and display the processed image with VGA. At the same time, the image data processed by signal Tap II acquisition algorithm is compared with the results of matlab software. Finally, the hardware verification results show that the proposed algorithm achieves the same processing effect as the software in hardware, which can eliminate the influence of image illumination and show the contour details of the enhanced dark area of the face at the same time.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】