ELFLN: An Efficient Lightweight Facial Landmark Network Based on Hybrid Knowledge Distillation

A facial landmark detector based on coordinate regression has minimal parameters and memory consumption, making it suitable for deployment on mobile devices. Typically, it employs a lightweight network as the backbone, but this network is not capable of effectively extracting features. Knowledge distillation is a promising methodology for developing a precise and lightweight network. The existing lightweight student networks do not have corresponding teacher networks, thereby hindering their ability to leverage the knowledge distillation technique. To tackle the inefficiency of the lightweight network in extracting features, we present an efficient lightweight network in this study, named ELFLN. In addition, we propose a novel hybrid knowledge distillation (HKD) framework to address the problem that the inadequacy of the lightweight network in carrying out features knowledge distillation. Finally, we augment our ELFLN by integrating facial landmark detection with head pose estimation, thereby enhancing the network generalization capability. To verify the efficacy of our proposed approach, we perform comprehensive experiments on 300W and WFLW datasets, achieving NME reach of 3.20% on 300W and 4.12% on WFLW with ELFLN+HKD. The number of parameters of ELFLN is observed to reduce by 62%, and FLOPs diminish by 87% compared to state-of-the-art SLPT.