摘要：

针对人脸图像在复杂环境下存在噪声污染、光照变化和遮挡等情况, 提出一种新的人脸识别方法, 即基于非凸混合范数误差编码的人脸识别方法(NMN), 旨在统一基于向量和基于矩阵的回归方法于一个回归模型中, 以更好地应对多样化的识别挑战。在考虑重构图像的低秩性质的同时, 引入核范数约束捕捉图像的低秩特征。为缓解由异常值引起的偏差问题, 引入非凸函数提高模型的鲁棒性。为进一步提升性能, 充分考虑标签信息, 以更有效地区分不同类别之间的特征差异。在分类阶段, 综合考虑非连续误差和连续误差, 利用类重构误差来判别待识别图像。在Extended Yale B、ORL和AR数据集上进行了实验验证, 结果表明, 相较于其他对比方法, NMN在遮挡人脸识别方面展现出更为卓越的性能。在Extended Yale B数据集上, 针对测试样本添加40%的“狒狒”图像遮挡块, 该方法的识别率达到80.40%, 比其他对比方法至少高出11.68百分点。

关键词: 人脸识别, 非凸函数, 核范数约束, 混合范数误差编码, 复杂环境

Abstract:

In response to the recognition challenges encountered by face images in complex environments with noise pollution, lighting variations, and occlusions, a face recognition method based on Non-convex Mixed-Norm error coding (NMN) is proposed. This method unifies the vector- and matrix-based regression approaches into a single-regression model. Recognizing the low-rank feature of the reconstructed images, nuclear norm constraints are applied to capture their intrinsic features. The introduction of non-convex functions mitigates the bias issues caused by outliers. To enhance the performance of the model, NMN considers label information and imposes constraints on interclass representations. Finally, in the classification phase, considering both noncontinuous and continuous errors, we utilized class reconstruction errors to discriminate against the to-be-recognized images. Experimental validation on the Extended Yale B, ORL, and AR datasets demonstrated that the proposed method exhibited superior recognition performance in the presence of occluded faces compared with the comparative methods. On the Extended Yale B dataset, the recognition rate of the NMN method achieved 80.40% when 40% of ″baboon″ image occlusion blocks were added to the test samples, surpassing other comparative methods by at least 11.68 percentage points.

Key words: face recognition, non-convex function, nuclear norm constraint, mixed norm error coding, complex environment