摘要
针对实际应用于"智慧南宁"项目建设时,多重训练样本容易使重构人脸陷入局部最大化以及协方差矩阵分解耗时严重的问题,提出一种基于改进Gabor-PCA分析重构的人脸遮挡物清除算法。在训练样本集选择阶段,通过构建5维8方向的Gabor直方图信息分类器,从人脸库中选择Gabor直方图信息与待重构原始人脸图像在外形轮廓等粗信息更为接近的图像组成训练样本集。同时,在PCA主元分析时,通过SVD分解重构协方差矩阵来降维,从而达到二次减少耗时的目的。实验证明,随着训练样本集的增加,该算法对各种人脸都有很强的适应性,并且得到的清除遮挡物后的人脸图像清晰、无局部最大化现象,与原始图像匹配度高,具备投入实际应用的能力。
When practically applying to the construction of'Smart Nanning City'project,multiple training samples are easy to leading the face reconstruction to falling into local maximum and to severe time consuming in covariance matrix decomposition. In order to solve these problems,this paper presents a face occluded area elimination algorithm which is based on Gabor and PCA analysis reconstruction. In the phase of training sample set selection,by constructing a 5-dimensional and 8-directional Gabor histogram information classifier we select the images from face database which have closer rough information such as the figure outlines between the Gabor histogram information and original face to be reconstructed for forming the training sample set. Meanwhile,when making PCA analysis,we use SVD( singular value decomposition) to reduce the dimensionality of covariance matrix,so as to reach the goal of decreasing time cost. Experiment proves that along with the increase of training sample sets,the algorithm has high adaptability to various faces,moreover,the derived face image with the occluded area eliminated is clear,does not have the phenomenon of local maximum,highly matches the original image,and possesses the capability of putting into practical use.
出处
《计算机应用与软件》
CSCD
2016年第7期234-238,共5页
Computer Applications and Software
基金
南宁市城市应急联动指挥系统升级项目(NNZC2010-1441A)
南宁市城市公共安全与社会综合服务系统人才小高地项目(2011020)
