基于Haar特性的改进HOG的人脸特征提取算法

发布时间：2018-05-08 16:51

本文选题：特征提取 + 人脸识别　；参考：《计算机科学》2017年01期

【摘要】：现有的大多数特征提取算法在提取人脸特征时,容易受到光照等外界因素的影响,从而导致后期人脸识别率下降。而方向梯度直方图(Histogram of Oriented Gradient,HOG)具有较强的光照鲁棒性,能够很好地减少由光照带来的干扰,但传统HOG在计算梯度幅值和方向时只计算水平和垂直方向上4个像素点对中间像素的影响,当外界环境变化时不能保证稳定性,因此提出一种基于Haar特性的改进HOG的人脸特征提取算法。该算法在计算梯度幅值和方向时考虑水平、垂直以及对角线上8个像素点对中间像素的影响,由于增加计算量导致特征提取时间也随之增加,因此引入Haar,借助Haar型特征运算简单、快捷的特点设计4组Haar型特征编码模式,按照改进的HOG特征计算方式提取人脸特征。在有光照等外界因素影响的FERET人脸数据库和Yale B扩展的人脸测试库中进行实验,实验结果表明,与GFC,LBP和其他文献中的HOG算法相比,该算法对光照具有更好的鲁棒性,能够在光照变化的环境下提高人脸识别率。该算法在FERET探测集fb,fc,dup1和dup2上的识别率分别为95.1%,80.9%,70.1%和63.2%,在Yale B中的识别率为89.1%。
[Abstract]:Most of the existing feature extraction algorithms are easy to be affected by external factors such as illumination when extracting face features, which leads to the decline of face recognition rate in the later stage. The histogram of Oriented gradient histogram has strong illumination robustness and can reduce the interference caused by illumination. However, the traditional HOG can only calculate the influence of 4 pixels in horizontal and vertical directions on the intermediate pixels in the calculation of gradient amplitude and direction, and can not guarantee the stability when the external environment changes. Therefore, an improved HOG based face feature extraction algorithm based on Haar characteristics is proposed. When calculating the magnitude and direction of gradient, the algorithm takes into account the influence of 8 pixels on the vertical and diagonal lines on the intermediate pixels, and the time of feature extraction increases with the increase of computation. Therefore, four groups of Haar type feature coding patterns are designed with the help of the simple and fast Haar type feature calculation, and the face features are extracted according to the improved HOG feature calculation method. Experiments are carried out in FERET face database and Yale B extended face test database with external factors such as illumination. The experimental results show that the algorithm is more robust to illumination than HOG algorithm in other literatures. It can improve the face recognition rate in the environment of changing illumination. The recognition rate of this algorithm on the FERET detection set fbbutu dup1 and dup2 is 95.1% and 63.2%, respectively. The recognition rate in Yale B is 89.1.
【作者单位】：南京邮电大学计算机学院;
【分类号】：TP391.41

【相似文献】