IEEE Trans Image Process. 2015 Dec;24(12):5868-78. doi: 10.1109/TIP.2015.2492819. Epub 2015 Oct 26.
Representation-based classifiers (RCs) have attracted considerable attention in face recognition in recent years. However, most existing RCs use the mean square error (MSE) criterion as the cost function, which relies on the Gaussianity assumption of the error distribution and is sensitive to non-Gaussian noise. This may severely degrade the performance of MSE-based RCs in recognizing facial images with random occlusion and corruption. In this paper, we present a minimum error entropy-based atomic representation (MEEAR) framework for face recognition. Unlike existing MSE-based RCs, our framework is based on the minimum error entropy criterion, which is not dependent on the error distribution and shown to be more robust to noise. In particular, MEEAR can produce discriminative representation vector by minimizing the atomic norm regularized Renyi's entropy of the reconstruction error. The optimality conditions are provided for general atomic representation model. As a general framework, MEEAR can also be used as a platform to develop new classifiers. Two effective MEE-based RCs are proposed by defining appropriate atomic sets. The experimental results on popular face databases show that MEEAR can improve both the recognition accuracy and the reconstructed results compared with the state-of-the-art MSE-based RCs.
基于表示的分类器 (RCs) 近年来在人脸识别中引起了相当多的关注。然而，大多数现有的 RCs 使用均方误差 (MSE) 准则作为代价函数，该准则依赖于误差分布的高斯性假设，并且对非高斯噪声敏感。这可能会严重降低基于 MSE 的 RCs 在识别具有随机遮挡和损坏的面部图像时的性能。在本文中，我们提出了一种基于最小误差熵的原子表示 (MEEAR) 框架用于人脸识别。与现有的基于 MSE 的 RCs 不同，我们的框架基于最小误差熵准则，该准则不依赖于误差分布，并且被证明对噪声更鲁棒。特别是，MEEAR 通过最小化重建误差的原子范数正则化 Renyi 熵来产生判别性的表示向量。为一般原子表示模型提供了最优条件。作为一个通用框架，MEEAR 也可以作为开发新分类器的平台。通过定义适当的原子集，提出了两种有效的基于 MEE 的 RCs。在流行的人脸数据库上的实验结果表明，与最先进的基于 MSE 的 RCs 相比，MEEAR 可以提高识别精度和重建结果。
