针对局部二值模式(LBP)及局部图结构(LGS)方法因非均衡的提取方式导致特征表达能力不强,以及单纯提取局部特征存在对人脸信息描述不全面的局限性,提出基于均衡局部模式的DOG多尺度融合金字塔人脸识别算法。首先针对LBP、LGS的不足,利用...针对局部二值模式(LBP)及局部图结构(LGS)方法因非均衡的提取方式导致特征表达能力不强,以及单纯提取局部特征存在对人脸信息描述不全面的局限性,提出基于均衡局部模式的DOG多尺度融合金字塔人脸识别算法。首先针对LBP、LGS的不足,利用双圆交叉采样和以局部宏观信息为参考的自适应阈值、中心对称的采样图结构实现均衡优化,提出可变参的ECLBP和FLGS方法,合称为均衡局部模式,以增强对关键局部信息的利用;然后将利用高斯核及图像差分生成的DOG金字塔与均衡局部模式方法融合,以多尺度特征图在丰富样本信息的同时实现大尺度全局轮廓和小尺度局部细节的融合,进一步捕捉类间及类内差异特征;最后加权级联所有特征图各子块的统计直方图,得到全面准确的特征向量,采用最近邻分类器匹配,通过特征图、直方图的主客观对比,从理论上验证了均衡局部模式获得了信息更丰富、判别性更强的局部特征。在ORL、AR和LFW数据集上,针对ORL、AR和LFW人脸库和相关典型方法的对比结果发现:在时间耗费相当时,均衡局部模式的识别率提升最高达15.52%;进一步融合DOG金字塔补充多尺度特征后,识别率再次提升,最高可达9.24%。实验结果表明,与现有典型提取方法相比,均衡局部模式特征明显拥有更强的表征能力及鲁棒性,多尺度信息进一步增强了特征性能,尤其在少样本环境,当样本信息有限时,整体算法的优势更加明显。In order to solve the problem that local binary pattern and local graph structure methods lack sufficient feature expression ability because of the unbalanced extraction method, and the limitation that only using local features can’t fully describe face information, this paper proposes a method called Face recognition based on DOG multi-scale fusion of Balanced Local Pattern. Firstly, in view of the shortcomings of LBP and LGS, on the basis of balanced optimization by using double-circle cross-sampling, adaptive threshold based on local macro information, and a center-symmetric sampling graph structure, this paper proposes Extended Cross Local Binary Pattern and Four-angle star Local Graph Structure methods with variable parameters, which are collectively called Balanced Local Pattern, which can enhance the extraction of key feature information. Then, the DOG pyramid generated by the Gaussian kernel and image difference is fused with the balanced local pattern method. The supplemented multi-scale feature map enriches the sample information while achieving the fusion of large-scale global contours and small-scale local details, which can further capture inter-class and intra-class difference characteristics. Finally, the comprehensive and accurate feature vector is obtained by weighted cascading the sub-block histograms of all feature map, and the nearest neighbor classifier is used to complete the recognition. The subjective and objective comparison of feature map and histogram theoretically verifies that the balanced local pattern can obtain local features with richer information and stronger discrimination. On the ORL, AR and LFW datasets, the proposed method is compared with the relevant typical methods, and the results on the ORL, AR and LFW databases show that the recognition rate of Balanced Local patterns is improved by up to 15.52% when the time consumption is the same as that of typical methods;after further integration of DOG pyramid, the recognition rate is increased by up to 9.24% again. The experimental results show that compared with the existing typical extraction methods, the balanced local pattern features have stronger representation ability and robustness, and the multi-scale information further enhances the feature performance, especially in the small sample environment, when the sample information is limited, the advantage of the whole algorithm is more obvious.展开更多
文摘针对局部二值模式(LBP)及局部图结构(LGS)方法因非均衡的提取方式导致特征表达能力不强,以及单纯提取局部特征存在对人脸信息描述不全面的局限性,提出基于均衡局部模式的DOG多尺度融合金字塔人脸识别算法。首先针对LBP、LGS的不足,利用双圆交叉采样和以局部宏观信息为参考的自适应阈值、中心对称的采样图结构实现均衡优化,提出可变参的ECLBP和FLGS方法,合称为均衡局部模式,以增强对关键局部信息的利用;然后将利用高斯核及图像差分生成的DOG金字塔与均衡局部模式方法融合,以多尺度特征图在丰富样本信息的同时实现大尺度全局轮廓和小尺度局部细节的融合,进一步捕捉类间及类内差异特征;最后加权级联所有特征图各子块的统计直方图,得到全面准确的特征向量,采用最近邻分类器匹配,通过特征图、直方图的主客观对比,从理论上验证了均衡局部模式获得了信息更丰富、判别性更强的局部特征。在ORL、AR和LFW数据集上,针对ORL、AR和LFW人脸库和相关典型方法的对比结果发现:在时间耗费相当时,均衡局部模式的识别率提升最高达15.52%;进一步融合DOG金字塔补充多尺度特征后,识别率再次提升,最高可达9.24%。实验结果表明,与现有典型提取方法相比,均衡局部模式特征明显拥有更强的表征能力及鲁棒性,多尺度信息进一步增强了特征性能,尤其在少样本环境,当样本信息有限时,整体算法的优势更加明显。In order to solve the problem that local binary pattern and local graph structure methods lack sufficient feature expression ability because of the unbalanced extraction method, and the limitation that only using local features can’t fully describe face information, this paper proposes a method called Face recognition based on DOG multi-scale fusion of Balanced Local Pattern. Firstly, in view of the shortcomings of LBP and LGS, on the basis of balanced optimization by using double-circle cross-sampling, adaptive threshold based on local macro information, and a center-symmetric sampling graph structure, this paper proposes Extended Cross Local Binary Pattern and Four-angle star Local Graph Structure methods with variable parameters, which are collectively called Balanced Local Pattern, which can enhance the extraction of key feature information. Then, the DOG pyramid generated by the Gaussian kernel and image difference is fused with the balanced local pattern method. The supplemented multi-scale feature map enriches the sample information while achieving the fusion of large-scale global contours and small-scale local details, which can further capture inter-class and intra-class difference characteristics. Finally, the comprehensive and accurate feature vector is obtained by weighted cascading the sub-block histograms of all feature map, and the nearest neighbor classifier is used to complete the recognition. The subjective and objective comparison of feature map and histogram theoretically verifies that the balanced local pattern can obtain local features with richer information and stronger discrimination. On the ORL, AR and LFW datasets, the proposed method is compared with the relevant typical methods, and the results on the ORL, AR and LFW databases show that the recognition rate of Balanced Local patterns is improved by up to 15.52% when the time consumption is the same as that of typical methods;after further integration of DOG pyramid, the recognition rate is increased by up to 9.24% again. The experimental results show that compared with the existing typical extraction methods, the balanced local pattern features have stronger representation ability and robustness, and the multi-scale information further enhances the feature performance, especially in the small sample environment, when the sample information is limited, the advantage of the whole algorithm is more obvious.
