基于改进APG的鲁棒光谱回归在人脸识别中的应用

Application of Robust Spectral Regressn Improved by APG in Face Recognition

针对人脸识别中传统的光谱回归方法对于由类标签错误和面部遮挡产生的误差很敏感的问题,提出了基于加速近似梯度的鲁棒光谱回归方法。首先将鲁棒判别子空间的学习问题转化成最大相关熵问题,得出光谱目标和预测之间的最大相关解;然后对相关熵采用总变分正则化,进而学习空间平滑的人脸结构;最后基于半二次化优化的附加形式,将最大相关熵问题投影到混合正则化模型,并通过加速近似梯度算法进行有效优化。在FRGC人脸数据库上的实验结果表明,所提方法对于类标签错误和面部遮挡具有鲁棒性和有效性,与其它几种常用的线性回归、光谱回归方法相比,不仅提高了识别率,而且大大地降低了计算开销。

For the problem that traditional spectral regression approaches are sensitive to the errors incurred by class label error and occlusion in face recognition, robust spectral regression based on accelerated proximal gradient （APG-RSR） is proposed. Firstly, the robust discriminant subspace learning problem is formulated as a maximum conrrentropy problem, which can help to find the most correlation solution between spectral targets and predictions. Then, total variation （TV） regularization is imposed on the conrrentropy objective to learn a spatially smooth face structure. Finally, the maximum conrrentropy problem is casted into a compound regularization model based on the additive form of half-quadratic optimization, which can be efficientIy optimized via an accelerated proximal gradient algorithm. Experiment results on FRGC face databases demonstrate the robustness and effectiveness of our method against inaccurate annotation and occlusion. Also, proposed method has improved recognition rates as well as re- ducing computational cost clearly comparing with several frequently-used linear regression and spectral regression approaches.