Abstract:

Traditional face recognition algorithms tend to decline when it occurs to low resolution face images. In order to solve this problem, a Light Discriminative Self-normalizing Neural Network（LDSNN） model is proposed, which extracts discriminative features from High-Resolution（HR） images and corresponding Low-Resolution（LR） images, and learns a coupled mapping which transforms the features to a common subspace. While the property of self-normalizing resulting from scaled exponential linear units accelerates the training stage. A loss function is designed to minimize intraclass distances and enlarge interclass distances based on not only the discriminability of HR-LR features, but also the similarity between them. Thus features from the same subject are more compacted together. The recognition rate of the proposed LDSNN model on a standard and two surveillance databases are 95.57%, 94.10%, and 84.56% respectively, better than other algorithms, which demonstrates that the proposed method works well with uncontrolled low-resolution face recognition.

Key words: face recognition, self-normalizing neural network, coupled mapping, sub-space, intraclass distances, interclass distances

摘要：

为了解决传统人脸识别算法对低分辨率人脸图片识别效果不佳的问题，提出了一种轻型判别自归一化神经网络，能够从高分辨率及其对应的低分辨率图像中提取具有判别性的特征，并将特征耦合映射到共同的子空间。该模型引入缩放指数线性单元，具有自归一化属性，能够加速收敛。为了最小化类内距以及扩大类间距，基于高低分辨率图像特征之间的判别性和相似度，对现有的损失函数进行了优化，从而使得相同类别的特征更紧凑。提出的方法在一个标准人脸数据集以及两个监控数据集上的识别率分别达到了95.57%、94.10%和84.56%，优于其他算法，适用于非限制条件下的低分辨率人脸识别。

关键词: 人脸识别, 自归一化神经网络, 耦合映射, 子空间, 类内距, 类间距