CALP-FD: Cross Attention-Based Facial Landmarks Prompt for Face De-Morphing

Face morphing attacks have been demonstrated to present a significant risk to current Automatic Border Control systems. Despite research on morphing attack detection has made certain progress, the forensic analysis and restoration of accomplices’ facial images remain in the initial stage. Currently, face de-morphing methods based on deep learning technology usually ignore the role of facial landmarks, which are crucial to the structure of facial images. Therefore, the paper proposes CALP-FD, a novel method for face de-morphing using cross-attention-based facial landmarks prompt. By integrating facial landmarks, CALP-FD pays more attention to the features around the facial features and assigns greater weight to pixels around the landmarks. In addition, in order to effectively separate the identities of accomplices, CALP-FD applies cross-attention mechanism for representation interaction in the separation network to improve the restoration accuracy and image quality. In the experimental section, the paper conducts a comprehensive evaluation on HNU_FM, and the experimental results unequivocally demonstrate the superior restoration accuracy and robustness of our method compared to existing approaches.