人脸识别是近年来图像处理、模式识别等领域的研究热点,但其面临的小样本空间、图像旋转、尺度变化等难题一直制约着人脸识别技术的发展与应用。为此,提出一种针对单样本人脸识别,提取尺度不变特征和样本脸进行匹配的识别方法:通过提取并匹配人脸图像的尺度不变特征得到特征点匹配关系,用以估计和校正图像间的旋转、尺度变化,并得到相对输入人脸图像尺度、旋转不变的样本脸,再使用子区域归一化互相关算法快速计算样本脸与输入图像之间的匹配程度,最后基于联合判决器综合特征点、样本脸两级匹配置信度综合判定识别结果。该方法有效结合了SIFT尺度不变特征对于旋转、尺度和光照变化稳定不变的优势,对于弱纹理图像也能够有效识别,具有较强的鲁棒性和适应性。

Face recognition has been always a hot spot in image processing and pattern recognition. However,problems,such as small samples,rotation,and scale changes,always restrict its development. In this paper,a scale invariant feature and template based method is proposed to achieve recognition upon a single face sample. First,scale invariant features are extracted and matched to obtain feature-level correspondence between faces,upon which rotation and scale variances between faces can be estimated and rectified for invariant templates. Then a sub-region based zero mean Nermaczed Cross-Correlation (ZNCC) method will be adopted to acquire region-level correspondences between the templates and rectified faces. Finally,a joint judgment will be made according to both feature-level and region-level correspondences between faces. This method is feasible and robust not only to rotation,scale and illumination variance,but also to low texture.