基于FPGA的低分辨率人脸识别系统设计

发布时间：2018-05-29 19:10

本文选题：低分辨率 + 人脸识别　；参考：《西安理工大学》2017年硕士论文

【摘要】：人脸识别一直是人工智能和机器视觉领域的热门研究方向,广泛应用在各个方面。对视力健全的人来说,可以很轻易地识别出一副人脸,但是要使盲人准确的识别每一个熟知的人就十分困难。人脸识别系统是帮助盲人提升识别效果、改善生活质量的最好选择,因为面部特征是人与人之间最明显的差异。本文以辅助盲人识别为大背景,并且根据盲人所处环境相对简单和盲人视觉假体项目,设计了适用于盲人的特定环境低分辨率人脸识别系统,整个系统在FPGA平台上进行了验证。全面分析现有的低分辨率人脸识别算法在特定环境下的识别效果后,本文选取了识别效果最佳的主成分分析和线性鉴别分析相组合的算法,实验结果表明该算法在特定环境下的人脸库中可达95%左右的识别率。依据算法原理和训练结果,本文完成了识别模块的硬件电路设计。硬件电路分为三个模块,包括特征提取模块、欧氏距离计算模块、最小欧氏距离提取模块,并且在FPGA平台上进行验证和测试。结果表明,硬件识别模块可以正确识别测试的低分辨率人脸图像,且识别时间为20us左右,远远高于软件识别的速度。为了更好的将硬件识别模块应用到辅助盲人识别的系统中去,本文使用SOPC技术搭建了一个系统,并将硬件识别模块使用Avalon总线挂载到系统。系统根据硬件识别模块的识别结果调取SD卡中相应的高分辨率人脸图像并通过VGA显示。实验结果表明,除去系统第一次启动时间,系统完成一次识别和显示过程大约需要0.16s的时间,即系统帧率可达6fps。本文在全面分析现有的低分辨率人脸识别算法的基础上,通过理论和实验证明主成分分析加线性鉴别分析算法在特定环境下的识别效果具有明显优势,对今后类似的研究有一定的参考意义。
[Abstract]:Face recognition has been a hot research direction in the field of artificial intelligence and machine vision. It is widely used in all aspects. For people with sound vision, a face can be easily identified, but it is very difficult for the blind to identify each well known person accurately. The best choice for good quality of life is that the facial features are the most obvious differences between people. This paper designs a specific environment low resolution face recognition system for blind people based on the blind person recognition and the blind human visual prosthesis. The whole system is on the FPGA platform. After analyzing the recognition effect of the existing low resolution face recognition algorithm in a specific environment, this paper selects the algorithm of combination of the principal component analysis and the linear discriminant analysis which has the best recognition effect. The experimental results show that the algorithm can reach about 95% recognition rate in the face database under a specific environment. The hardware circuit of the recognition module is completed in this paper. The hardware circuit is divided into three modules, including feature extraction module, Euclidean distance calculation module, minimum Euclidean distance extraction module, and the verification and test on the FPGA platform. The results show that the hardware recognition module can correctly identify the low resolution face of the test face. The image, and the recognition time is about 20us, is far higher than the speed of the software recognition. In order to better apply the hardware recognition module to the auxiliary blind recognition system, this paper uses the SOPC technology to build a system and mount the hardware recognition module to the system using the Avalon bus. The system is adjusted according to the recognition result of the hardware recognition module. The corresponding high resolution face images in the SD card are taken and displayed by VGA. The experimental results show that the system completes the process of recognition and display for a time of about 0.16s, that is, the system frame rate is up to 6fps., which is based on the analysis of the existing low resolution face recognition algorithm and through theory and reality. It is proved that the recognition effect of the principal component analysis and the linear discriminant analysis algorithm in the specific environment has obvious advantages, and it has some reference significance for the similar research in the future.
【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】：TN791;TP391.41

【参考文献】