Researchers Use AI for Deblurring Faces in Photos

2020-01-23

It is extremely difficult to capture high-speed object photos with a mobile phone, often resulting in a pile of blurry and indistinguishable images. In search of a solution, scientists from the Artificial Intelligence Research Institute of the United Arab Emirates, Beijing Institute of Technology, and Stony Brook University have jointly developed an artificial intelligence system that can remove blurriness from images during post-production. In a research paper, they pointed out that this technology is capable of perceiving faces, meaning it can deblur faces. Furthermore, compared to state-of-the-art dynamic deblurring methods, it exhibits remarkable performance. Due to the relative motion between the camera and the object, both foreground and background often experience different types of degradation. Additionally, due to the distance between the subject and the image plane, they undergo different types of motion. The researchers' model learns and utilizes human and background masks to capture the blurriness by separating foreground and background. To train this model, the researchers compiled a dataset consisting of thousands of blurry images with various foreground motions, complex backgrounds, and corresponding ground truth clear images. Each pair of images was inputted through a human detection model, which provided "roughly accurate" bounding boxes around the subjects, later improved by human annotators. Using a machine equipped with only an Nvidia Titan X graphics card, the researchers trained their deblurring model on a partial HIDE and GoPro Hero supplementary video frame dataset, totaling 10,742 images. The researchers claim that their model achieves state-of-the-art performance in dynamic deblurring and outperforms several baselines, resulting in better restoration effects. By integrating deblurring features from different domains, the model is capable of reconstructing images with clear structures and semantic details. This design leads to a unified, user-friendly, and meticulous deblurring network. By explicitly and independently modeling various human motion patterns and rich semantics, it brings about better deblurring effects for both foreground subjects and background environments. It is worth noting that this model technology is not unique, as Nvidia, MIT, and Alto University have recently proposed a machine learning technique for reducing image blurriness.

RECENT AI NEWS

RECENT AI TOOLS