【摘要】 本文提出一种快速人脸特征描述(FFD,fast facial descriptor)算法和基于权重的人脸图像相似度分数匹配策略,以解决加速鲁棒性特征(SURF,speed up robust features)在描述单样本人脸特征时出现的特征点分布不均匀和光照变化鲁棒性差的问题。首先通过重构积分图来增加位于细长边缘的特征点的数量;为了减少冗余特征,提出两幅训练图像对应特征点间的区别度概念,对训练样本中的特征点进行稀疏化;然后,根据人脸各区域对识别结果贡献度的不同对人脸各区域赋予不同权重,并根据加权计算人脸图像的相似度分数得出识别结果。在AR、Yale B和CMU PIE标准人脸数据库及真实身份证人脸库上进行了单样本人脸识别实验。结果表明,本文算法对具有光照、遮挡和表情变化的单样本人脸识别有很好的鲁棒性,耗时仅为0.042s;与目前典型的特征描述算法相比,本文算法的识别率最高可提升65%;虽然真实身份证人脸库中人脸图像受实际环境因素影响较大,但本文方法也可提高30%的识别率。更多还原

【Abstract】 In this paper,a fast facial descriptor(FFD)algorithm and a face images similarity score matching strategy are proposed to improve the uneven distribution of features and poor illumination robustness of speed up robust features(SURF)in describing features for single image face recognition.The integral image is reconstructed to increase the number of features on the lathy edge of face image to alleviate the problems of SURF.In order to reduce the number of redundant features,the conception of distinction between the corresponding features is introduced to extract sparse features on training images.According to the contribution to recognition rate,face image regions are weighted to calculate the similarity score between face images to conduct the final result.Experiments of single image based face recognition are conducted on AR,Yale B,CMU PIE databases and real-life ID card image database.The results indicate that the proposed method is robust to single image based face recognition under illumination,occlusion and expression changes,and the running time is only 0.042s.Compared with some other recent feature description methods,the proposed method can increase the recognition rate by up to 65%.Although the images in real-life ID card image database are greatly affected by actual environmental factors,the recognition rate is increased by 30%.更多还原