Multi-Scale Feature Learning Based on RICA for Single Training Sample Face Recognition

摘要

In many face recognition scenarios, there are three main challenges for face recognition include single training sample, partial occlusion and uneven illumination. While conventional face descriptors performed unsatisfied and deep learning networks are always time-consuming and resource-consuming. To address these challenges, a novel multi-scale feature is proposed to encoding method to extract multi-scale feature representation for face recognition in complex scenarios with single training sample. Firstly, we divide each face image into multi-scale regions and take dense samples with multi-scale patches in each region. Secondly, we introduce to learn multi-scale encoding matrixes for different face regions based on RICA. Finally, the face features will be extracted for comparing. Experimental results prove that our approach perform stably to expression, partial occlusion and uneven illumination conditions compared with the existing unsupervised methods.