一种基于多级梯度能量特征的DCT压缩域人脸检测算法  

An Algorithm of Multilevel Gradient Energy based Face Detection in DCT Compressed Domain

作　　者：李晓光[1] 李晓华[1] 沈兰荪[1] 

机构地区：[1]北京工业大学信号与信息处理研究室,北京100022

出　　处：《电子学报》2005年第12期2170-2173,共4页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60172045;No.60402036);北京市自然科学基金(No.4042008)

摘　　要：压缩域人脸检测在图像/视频信息处理中具有重要意义.本文提出了一种基于多级梯度能量特征的DCT(Discrete Cosine Transform)压缩域人脸检测算法.依据DCT压缩图像色差信号的直流系数进行肤色分割,减小检测范围.在分割为肤色的区域提取多级梯度能量特征,即利用不同大小的检测窗口提取归一化的特征向量,表示不同大小的人脸.特征向量输入到级联分类器中分类,确定是否表示人脸.级联分类器由若干简单分类器和一个神经网络分类器构成.简单分类器利用一些先验知识排除大部分明显不是人脸的特征向量,通过简单分类器的特征由神经网络最终确定是否表示人脸.多级梯度能量特征与DCT域图像缩放相结合实现了对不同大小人脸的快速检测.对多级梯度能量特征的定义,减少了检测算法中压缩域图像缩放的次数,从而大幅度减少了计算复杂度,提高了检测速度.实验结果表明提出的多级梯度能量特征可有效描述DCT域人脸模式,同时也证明了该算法的快速有效性.Face detection is important in the processing of images and video. Based on multilevel gradient energy （MGE）,an algorithm of face detection in DCT （Discrete Cosine Transform） compressed domain is presented. In preprocessing procedure, skin color segmentation based on the DC of chromatic components is applied to the input image for reducing the detected regions. According to the map of skin segmentation, MGE based feature vector is extracted, viz. normalized feature vectors are extracted from the detecting windows of various sizes to describe faces of different sizes. Then cascade classifier is employed to classify the feature vectors as face or non-face. Cascade classifier is comprised of several simple classifiers and a neural network classifier. Lots of feature vectors that belong to non-face are removed by simple classifiers which embedded preknowledge rules. The left vectors are classified by neural network. We combined MGE features together with image scaling to allow faces of various sizes. The simphcity of feature extraction accelerated detection by reducing the times of image scaling which is more time cost. The experiment results show that the proposed method is efficient and effective.

关 键 词：人脸检测 DCT 压缩域 多级梯度能量 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]