Block dictionary learning-driven convolutional neural networks for fewshot face recognition

Fewshot face recognition (FFR) in less constrained environment is an important but challenging task due to the lack of sufficient sample information and the impact of occlusion. In this paper, a novel approach called block dictionary learning (BDL) is proposed, which combines sparse representation with convolutional neural networks to address the FFR problem. Based on the key-point locations of face images, the images are divided into four block regions for local feature extraction. Then, highly compact and discriminative features of both holistic and segmented parts are generated by CNN, which further compensates for the shortage of samples. Moreover, the sparse loss is introduced to optimize the performance of CNN by increasing the inter-class variations of features; thus, it develops a global-to-local dictionary learning algorithm to improve the robustness of BDL against complex variations. Finally, extensive experiments on AR and Extended Yale B datasets significantly demonstrate the effectiveness of BDL in comparison with other FFR methods.

Automated summary

The paper introduces a novel approach called block dictionary learning (BDL) which combines sparse representation and convolutional neural networks to address the fewshot face recognition problem. Through local feature extraction and a global-to-local dictionary learning algorithm, BDL demonstrates effectiveness in comparison with other FFR methods on AR and Extended Yale B datasets.