基于吸引模式的局部二阶梯度轮廓人脸识别算法

Face recognition algorithm based on attractive local second gradient contours

针对现有LDP类算法在特征提取有效性和特征编码稳定性之间难以平衡的问题,提出一种吸引局部二阶梯度轮廓(ALSGC)模式,以提升人脸识别性能。首先,利用Kirsch算子计算人脸的邻域边缘响应图;其次,引入吸引描述子,参考边缘响应图的局部、全局平均灰度值和邻域中心灰度值完成局部吸引模式编码;再次,遍历整幅图像,得到人脸ALSGC特征图并对ALSGC特征图分块分别计算,得到各个分块中不同模式的统计直方图;最后,级联所有分块的统计直方图后生成对应的特征向量,以支持向量机完成分类识别。所提算法克服了LBP、LDP、LDN等算法提取一阶特征有效性的不足,以及DLDP、CSLDP、GCSLDP等算法提取的二阶特征对表情、姿态、饰物遮挡、光照、随机噪声等变化敏感的缺点,较好地实现了特征提取有效性与特征编码稳定性的平衡,兼顾了识别率和稳健性。

In order to solve the problem that traditional LDP algorithms is difficult to balance the effectiveness of feature extraction and the stability of feature encoding,an attractive local second gradient contours(ALSGC)face feature extraction algorithm was proposed.Firstly,the Kirsch operator was used to calculate the neighborhood edge response image of the face image.Secondly,an attraction descriptor was introduced and the local average gray value and global average gray value of the edge response image and the neighborhood center gray value was combined to complete the local attraction pattern encoding.The entire image was traversed to get the ALSGC face feature map,the ALSGC feature map was divided to blocks and the statistical histograms of different patterns for each block was obtained by calculation.Finally,the statistical histograms of all the blocks were cascaded to generate the corresponding feature vector and support vector machine was used to complete classification and recognition.The proposed algorithm not only overcame the insufficient effectiveness of LBP,LDP,LDN and other algorithms who extract first-gradient features,but also reduced the sensitivity to changes in expression,posture,occlusion,lighting and random noise of methods such as DLDP,CSLDP,GCSLDP who extracting second-gradient features.It better achieved the balance between the effectiveness of feature extraction and the stability of feature encoding,and took into account the recognition rate and robustness.