基于迹变换和旋转增量调制特征的模糊人脸识别  

Rotational Delta Modulation Feature and Its Application in Blurry Face Recognition Based on Trace Transform

作　　者：汪宇玲[1] 陈立 黎明 钟国韵[1] 何月顺[1] 常玉祥 宋伟宁 WANGYu-ling;CHENLi;LIMing;ZHONGGuo-yun;HEYue-shun;CHANGYu-xiang;SONGWei-ning(Jiangxi Engineering Technology Research Center of Nuclear Geoscience Data Science and System,East China University of Technology,Nanchang,Jiangxi 330013,China;Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition,Nanchang Hangkong University,Nanchang,Jiangxi 330063,China)

机构地区：[1]东华理工大学江西省核地学数据科学与系统工程技术研究中心,江西南昌330013 [2]南昌航空大学江西省图像处理与模式识别重点实验室,江西南昌330063

出　　处：《电子学报》2021年第12期2437-2448,共12页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62066003,No.61866025,No.D0213);国家重点研发计划(No.2018YFB1702702);江西省核地学数据科学与系统工程技术研究中心开放基金(No.JETRCNGDSS202006)。

摘　　要：为提高各种不同特殊场景下的模糊人脸识别精确性和鲁棒性,本文提出一种基于迹变换和旋转增量调制编码的特征提取方法.该方法将通信语音编码技术与图像变换技术相结合,首先通过迹线旋转扫描整幅图像,并对迹线上的采样信息进行增量调制编码,从而获得多角度的全局有序结构特征,然后用支持向量机对样本图像的这些特征进行训练以分类并识别图像.实验结果表明,在各种不同模糊级的低质量人脸数据库上,本文方法对不同光照变化、不同拍摄角度、不同遮挡等不同场景的人脸图像均能取得较好的识别效果,与一些传统方法相比识别性能大幅提升,相对于VGGNet和Sphereface两种先进方法在三组不同模糊度测试图像集的平均识别率分别提高2.18%和2.20%,具有更高的识别精度和更好的鲁棒性.For the accuracy and robustness of the blurry face image recognition in various special scenarios;this paper proposes a novel rotation delta modulation texture extraction method based on trace transform and delta modulation encod⁃ing.The method combines the communication speech coding with the image transformation technology.Firstly,the whole image is scanned by the rotating trace line based on trace transformation theory.Secondly,the sampled pixels on the trace line are encoded used delta modulation technology.The global ordered structure features are obtained in various angles.Fi⁃nally,the features are trained by support vector machine for blurry face image classification and recognition.The experimen⁃tal results indicate the proposed method has the better performance of blurry face image recognition under illumination fluc⁃tuations,various camera angles and image occlusions,and so on,on a variety of low-quality face image sets with different blur levels.The average recognition ratio are increased 2.18%and 2.20%compared to VGG and Sphereface methods on the face image sets with three blur levels,which indicates the proposed method has the higher accuracy and better robustness.

关 键 词：模糊人脸识别 特征提取 迹变换 旋转增量调制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN911.73[自动化与计算机技术—计算机科学与技术]