Facial landmark points detection using knowledge distillation-based neural networks

Facial landmark detection is a vital step for numerous facial image analysis applications. Although some deep learning-based methods have achieved good performances in this task, they are often not suitable running on mobile devices. Such rely networks with many parameters, which makes the training and inference time-consuming. Training lightweight neural such as MobileNets challenging, models might low accuracy. Inspired by knowledge distillation (KD), paper presents novel loss function to train Student network (e.g., MobileNetV2) detection. We use two Teacher networks, Tolerant-Teacher Tough-Teacher conjunction network. The trained using Soft-landmarks created active shape models, while ground truth (aka Hard-landmarks) points. To utilize points predicted we define an Assistive Loss (ALoss) each Moreover, called KD-Loss that utilizes pre-trained (EfficientNet-b3) guide towards predicting Hard-landmarks. Our experimental results three challenging datasets show proposed architecture will result better-trained can extract high • Applying convolutional efficient face alignment. Using Teachers (a tough tolerant network)