【作者】 张刚；

【导师】 汪西原；

【作者基本信息】 宁夏大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 人脸检测及其识别,带来了刷脸支付,让追捕犯罪分子、寻找走失儿童与老人等应用更加便利。深度学习理论与算法的实现更让人脸识别技术日臻成熟。但是在实际生活场景中,特别是近年的疫情背景下,人们常常佩戴口罩、眼镜(墨镜)等使得人脸部分遮挡,从而造成特征缺失导致人脸识别算法的识别准确率大幅下降。消除遮挡区域对人脸识别的影响,提高人脸识别准确率,增强算法的鲁棒性成为了亟待解决的问题。本文从两方面提出并研究改进了遮挡人脸识别算法:一是基于已有人脸特征,通过调整提升未遮挡区域的特征权值、相应的降低遮挡区域的特征权值,来改善或消除遮挡区域对人脸识别的影响;二是先对遮挡人脸图像的遮挡区域进行修复,再将修复图像送入人脸识别网络,实现遮挡人脸可靠识别。主要工作围绕以下两方面展开。1、为解决基于特征提取的人脸识别方法在遮挡区域较大的情况下准确率低的问题,本文在对Inception网络结构改进的基础上结合空间特征和通道特征提出一种融合特征的局部遮挡人脸识别算法,该算法能够将图像的空间特征和通道特征作为重点,进行提取。通过实验测试四种相关算法和本文算法的性能,在无遮挡时,本文算法识别精度比ArcFace低0.11%。眼镜遮挡情况下,本文算法识别精度比MobileFaceNet低2.91%。但是该算法模型仅有3.5M,对于口罩遮挡识别精度比VGGFace提高了 7.99%,围巾遮挡比ArcFace提高了 10.42%。2、在遮挡人脸图像修复识别方面,对比分析了具有代表性的图像修复算法GAN、DCGAN、WGAN的性能。经过综合评测选取WGAN,在整个算法中加入全局判别器和局部判别器,在损失函数中引入了对称损失、内容损失和结构损失,并且生成器网络结构采用U-Net模型,保证了修复人脸图像整体的完整性和语义的连贯性。为此,本文采用结构相似性(Structual Similarity,SSIM)和峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)对修复后的图像质量进行评估,实验表明本文算法相比其他对比算法(如VGG16,ArcFace等)识别准确率分别提高了-0.63%,1.85%,0.77%,1.82%,3.19%,据此可以看出本文算法在无遮挡时识别准确率稍低于ArcFace,但是随着遮挡面积的增加,本文算法识别准确率相对高于实验中对比的其他算法。实验结果表明,本文提出的算法优化策略对解决由于遮挡造成人脸特征缺失而导致的人脸识别准确率降低的问题效果显著,具有一定的优势和较好的稳定性,可进一步拓宽人脸识别技术的应用场景,具有较强的应用价值。更多还原

【Abstract】 Face detection and recognition has led to face payment,making it easier to track down criminals and find lost children and the elderly.Deep learning theory and algorithm realization make face recognition technology mature day by day.However,in real life scenes,especially in the context of this year’s epidemic,people often wear masks,glasses(sunglasses)and other parts of the face cover,resulting in feature loss and resulting in a sharp decline in the recognition accuracy of face recognition algorithms.How to eliminate the influence of occlusion area on face recognition,improve the accuracy of face recognition,enhance the robustness of the algorithm has become an urgent problem to be solved!This paper proposes and studies the improved face occlusion recognition algorithm from two aspects:first,based on the existing face features,the influence of occlusion area on face recognition can be eliminated by adjusting and improving the feature weight of unoccluded area and reducing the feature weight of occluded area accordingly;Second,the occluded area of occluded face image is repaired first,and then the repaired image is sent to face recognition network to realize reliable occluded face recognition.The main work revolves around the following two aspects.1.To solve the face recognition method based on feature extraction in the case of keep out area is larger the low accuracy problem in Inception improvement on the basis of the network structure of the space and channel characteristics put forward a kind of fusion of partial shade face recognition algorithm,the algorithm can image spatial characteristics and channel characteristic,as the key is extracted.The performance of the four related algorithms and the proposed algorithm is tested experimentally.In the case of no occlusion,the recognition accuracy of the proposed algorithm is only 0.11%lower than ArcFace.In the case of glasses occlusion,the recognition accuracy of the proposed algorithm is only 2.91%lower than MobileFaceNet.However,the algorithm model is only 3.5m,and the mask occlusion recognition accuracy is 7.99%higher than VGGFace,and the scarf occlusion accuracy is 10.42%higher than ArcFace.2.In terms of occluded face image repair,the performance of representative image repair algorithms GAN,DCGAN and WGAN is compared and analyzed.After comprehensive evaluation,WGAN was selected,global discriminator and local discriminator were added into the whole algorithm,symmetric loss,content loss and structure loss were introduced into the loss function,and u-NET model was used in the generator network structure,which ensured the integrity of the whole face image and semantic continuity.Therefore,this paper adopts Structual Similarity(SSIM)and Peak signal-to-noise Ratio(PSNR)to evaluate the image quality after restoration.Experimental results show that the proposed algorithm has the best recognition performance compared with other algorithms(VGG16,ArcFace,etc.)improved their recognition accuracy by-0.63%,1.85%,0.77%,1.82%and 3.19%,respectively.Therefore,it can be seen that the recognition accuracy of the proposed algorithm is slightly lower than ArcFace without occlusion,but with the increase of occlusion area,the recognition accuracy of the proposed algorithm keeps improving.The experimental results show that the algorithm optimization strategy proposed in this paper has certain advantages and good stability to solve the problem of face recognition accuracy reduction caused by the loss of face features caused by occlusion.It can further broaden the application scenarios of face recognition technology and has strong practical value.更多还原