天津工业大学学报
天津工業大學學報
천진공업대학학보
JOURNAL OF TIANJIN POLYTECHNIC UNIVERSITY
2015年
2期
69-74,79
,共7页
肖志涛%伊靓%李月龙%张芳%耿磊%吴骏
肖誌濤%伊靚%李月龍%張芳%耿磊%吳駿
초지도%이정%리월룡%장방%경뢰%오준
非正面人脸合成%筛选评估准则%人脸识别
非正麵人臉閤成%篩選評估準則%人臉識彆
비정면인검합성%사선평고준칙%인검식별
non-frontal face image synthesis%source screening criteria%face recognition
针对传统方法合成的正面人脸图像中信息丢失和变形的问题,提出了一种基于筛选评估准则的非正面人脸图像合成方法。人脸筛选评估准则融合了脸部对称性、正脸差异水平和人脸水平扭转角度3方面信息,其中人脸水平扭转角度利用细节上的眼部信息来评价人脸的正面水平,而脸部对称性和正脸差异水平分别对人脸的左右和垂直方向进行整体评价,综合这三方面信息可有效地排除低质量侧脸图像对合成正脸图像的干扰。首先进行标记点检测跟踪,然后基于此对同一人的多幅侧脸图像进行筛选,最后进行插值运算合成正面人脸,并在FERET图像库中对该方法进行实验验证。结果表明:通过本文筛选准则可有效滤除合成中低质量、强干扰的侧脸图像,可降低姿态问题对人脸识别精度的干扰,最终合成精确逼近真实正面人脸的合成图像。
針對傳統方法閤成的正麵人臉圖像中信息丟失和變形的問題,提齣瞭一種基于篩選評估準則的非正麵人臉圖像閤成方法。人臉篩選評估準則融閤瞭臉部對稱性、正臉差異水平和人臉水平扭轉角度3方麵信息,其中人臉水平扭轉角度利用細節上的眼部信息來評價人臉的正麵水平,而臉部對稱性和正臉差異水平分彆對人臉的左右和垂直方嚮進行整體評價,綜閤這三方麵信息可有效地排除低質量側臉圖像對閤成正臉圖像的榦擾。首先進行標記點檢測跟蹤,然後基于此對同一人的多幅側臉圖像進行篩選,最後進行插值運算閤成正麵人臉,併在FERET圖像庫中對該方法進行實驗驗證。結果錶明:通過本文篩選準則可有效濾除閤成中低質量、彊榦擾的側臉圖像,可降低姿態問題對人臉識彆精度的榦擾,最終閤成精確逼近真實正麵人臉的閤成圖像。
침대전통방법합성적정면인검도상중신식주실화변형적문제,제출료일충기우사선평고준칙적비정면인검도상합성방법。인검사선평고준칙융합료검부대칭성、정검차이수평화인검수평뉴전각도3방면신식,기중인검수평뉴전각도이용세절상적안부신식래평개인검적정면수평,이검부대칭성화정검차이수평분별대인검적좌우화수직방향진행정체평개,종합저삼방면신식가유효지배제저질량측검도상대합성정검도상적간우。수선진행표기점검측근종,연후기우차대동일인적다폭측검도상진행사선,최후진행삽치운산합성정면인검,병재FERET도상고중대해방법진행실험험증。결과표명:통과본문사선준칙가유효려제합성중저질량、강간우적측검도상,가강저자태문제대인검식별정도적간우,최종합성정학핍근진실정면인검적합성도상。
To solve the problem of information missing and deformation of the traditional synthesized frontal face image, a non-frontal face image synthesis method based on source screening criteria is proposed. The source screening criteria combines the facial symmetry, the frontal face image differences level and the horizontal rotating angle. The horizontal rotating angle uses the details of the eye information to assess the level of the front face. The left-to-right and vertical direction is separately assessed according to the facial symmetry and the frontal face image differences level. Combining these three aspects can availably exclude interference of low-quality non-frontal face image in synthesizing the frontal face. Firstly, marking points are detected and tracked. Then, multi-view face images of same person are screened to obtain the high-quality, low-interference non-frontal face images as the best synthetic input source images. Finally, frontal face images are synthesized by calculated interpolation. Experimental findings on the FERET databases demonstrate that the source screening criteria can availably screen out low-quality, strong-interference non-frontal face images and reduce the impact of pose questions on the precision of face recognition algorithm effectively. The synthesized frontal view face can approximate the ground truth frontal view face.