Cross spatial and Cross-Scale Swin Transformer for fine-grained age estimation

Abstract

Facial age estimation is a classic problem in the field of computer vision. Previous studies have shown that learning discriminative features is crucial for accurate age estimation. Although Swin Transformer has been successfully applied on many computer vision tasks, it cannot effectively capture directional features during the aging process for age estimation task. Moreover, it still exhibits bias towards global features and cannot capture more fine-grained age-related features, ultimately leading to ambiguity in distinguishing adjacent ages. To address these issues, we propose Cross Spatial and Cross-Scale Swin Transformer (CSCS-Swin) that can extract fine-grained age-related features. Firstly, the Cross Spatial Feature Block (CSFB) module is constructed in CSCS-Swin, which extracts facial wrinkle features and craniofacial features along the horizontal and vertical directions, and models feature associations between different facial regions. Secondly, considering that the discrimination power of features at different scales differs in facial regions, Cross-Scale Feature Partition (CSFP) is proposed, which can precisely extract corss-scale fine-grained features. Lastly, the Feature Enhancement Module (FEM) is introduced to further enhance the ability of feature representation. These three modules in CSCS-Swin work together to improve the accuracy of age estimation. Extensive experiments on four popular datasets, namely, MORPH II, UTKFace, AFAD, and CACD, demonstrate the superiority of the proposed method.