Enhanced face recognition in unconstrained environments: leveraging 3D morphable model and focal modulation network
Randa Nachet, Tarik Boudghene Stambouli · Brazilian Journal of Technology · 2025
Due to the fact that faces can vary significantly in appearance when viewed from different angles or orientations, face recognition systems can easily be impacted by this issue. In this paper, we propose a novel approach for face frontalization leveraging a 3D morphable model (3DMM) and advanced deep learning techniques. Our architecture comprises two main modules: the 3D face fitting module, where we redesign this strategy using a focal modulation network-based encoder, and the texture completion module, which recovers the unseen regions caused by self-occlusion. Experimental results demonstrate the effectiveness of our method in producing high-quality frontalized face images across different poses, improving the performance of face recognition in unconstrained environments.