一种稀疏性增强的人脸识别改进算法

An improved face recognition algorithm based on sparsity enhancing

针对在现有稀疏表示分类(SRC)算法中,用l1范数取代l0范数并不能得到有效的稀疏解,提出了一种将lp(0≤p<1)范数和新判别规则完美结合的人脸识别方法。首先通过迭代算法求解lp范数最小化问题,以此代替传统SRC中的l1范数来求解编码系数,得到更稀疏和有效的解。为了从稀疏编码系数中捕捉到更多的差分信息,并兼顾残差反映每一类样本的贡献,用系数和与残差之比这一新判别规则来分类测试样本。在AR人脸数据库的实验结果表明,本算法可得到更稀疏有效的解,且可在一定程度上提高识别率,尤其是在伪装情况下,有较为明显的提高。

For existing sparse representation classification (SRC ) algorithm, /' minimization to replace the /0 minimization does not obtain sufficiently sparse coefficients,a face recognition algorithm combing perfectly the lp ( 0 ^ p < 1 ) minimization and new decision rule is proposed. Firstly, solving the lpminimization solution by an iterative algorithm, which replaces the traditional /' minimization to obtain the coding coefficients is sparser and more effective. In order to capture more differential information fromsparse coding coefficient,and taking into account residual reflects the contribution of each class,the ratioof the sum of coefficients and the residual acts as a new decision rule. The experimental results on A Rface database show that the algorithm obtain a sparser solution, and improving the recognition rate to some extent, especially in the guise case, there is a more significant improvement.