Dual-Path Adaptive Channel Attention Network Based on Feature Constraints for Face Anti-Spoofing

Interference factors in visible light image data, such as backgrounds and lighting, often lead to poor performance of RGB-based single-modality face anti-spoofing methods. To address these limitations, we propose an innovative face anti-spoofing framework. Within this framework, we design a convolutional neural network (CNN) based on the Dual-path Adaptive Channel Attention (DACA) module, aiming to filter the features of the input facial images to extract key information. In addition, we develop feature constraints method based on Inner Similarity Estimation (ISE), which effectively enhances intra-class consistency by reducing the distance between samples and their class center. This method narrows the intra-class sample distribution and improves class separability, preventing the model from learning excessive irrelevant information and enhancing the robustness and generalization of face anti-spoofing. We test our method on the CASIA SURF dataset, CASIA SURF-CeFA dataset, and CASIA FASD dataset, which shows that our method has significant advantages in distinguishing between live and spoofed faces.