彩色图像边缘特征及其人脸检测性能评价

Edge Features in Color Image and Their Face Detection Performance Evaluation

介绍了一种有效的彩色图像边缘特征提取算法,提出了一种新的边缘方向编码——双轴对称方向编码,利用多重交叉验证的ROC(receiver operating characteristic)曲线对基于颜色及其边缘直方图的SVM(supportvector machine)人脸检测进行平均性能评价.实验结果表明,图像颜色边缘特征比灰度边缘特征具有明显优势.通过分别与RGB三色直方图线性拼接,新的双轴对称边缘方向编码表现出比传统方向编码更好的SVM分类性能.利用颜色及其边缘直方图特征能够明显提高人脸检测性能,分辨出不同光照条件下、不同表情甚至部分遮挡的非深度旋转的彩色人脸.

This paper introduces an effective algorithm for color edge features extraction and proposes a novel edge orientation encoding, biaxial symmetry orientation encoding. The average performance of human face detection system, which is based on the support vector classifier using the histograms of color and color edge features, is evaluated with ROC in multi-fold cross validation. Experimental results show that color edge features outperform gray edge features evidently; the classification accuracy of the novel edge orientation coding outperforms the traditional edge orientation coding when they are linearly combined with color histogram respectively; the face detection accuracy can be significantly improved when both color and color edge histograms are used, non-deep rotated human face can be correctly detected in color image under different illuminations, with different expressions and partial occlusions.