Multiscale Dynamic Graph Representation for Biometric Recognition with Occlusions

• DOI: 10.48550/arxiv.2307.14617
• Publication Date: 2023-01-01
• Authors: Min Ren; Yunlong Wang; Yuhao Zhu; Kunbo Zhang; Zhenan Sun
• Keywords: Representation; Margin (machine learning); Matrix representation

Occlusion is a common problem with biometric recognition in the wild. The generalization ability of CNNs greatly decreases due to adverse effects various occlusions. To this end, we propose novel unified framework integrating merits both and graph models overcome occlusion problems recognition, called multiscale dynamic representation (MS-DGR). More specifically, group deep features reflected on certain subregions recrafted into feature (FG). Each node inside FG deemed characterize specific local region input sample, edges imply co-occurrence non-occluded regions. By analyzing similarities representations measuring topological structures stored adjacent matrix, proposed leverages matching judiciously discard nodes corresponding occluded parts. strategy further incorporated attain more diverse representing regions sizes. Furthermore, exhibits illustrative reasonable inference by showing paired nodes. Extensive experiments demonstrate superiority framework, which boosts accuracy natural occlusion-simulated cases large margin compared that baseline methods.