Face Detection in Distorted Images Based on Image Restoration and Meta-learning

Face detection in distorted images is a challenging task, and a natural idea is to conduct image restoration before face detection. Most of current image restoration techniques focus on improving the perceptual quality of the output image for single type of distortion, without taking the subsequent high-level detection task and unknown distortion into account. In this paper, we propose a restoration-based face detector in which the images are restored based on the detection loss instead of the perceptual quality loss, leading to better performance on subsequent detection task. Furthermore, we employ meta-learning to initialize the model with more appropriate parameters, thus our detector can adapt quickly to unseen distortions only using few examples with the corresponding distortion. Experiments on public datasets show that our proposed method could improve the performance of face detection in distorted images, and have a better generalization ability when applied to images with unseen distortions.