- Abstract: Deep Learning (DL) algorithms based on Generative Adversarial Network (GAN) have demonstrated great potentials in computer vision tasks such as image restoration. Despite the rapid development of image restoration algorithms using DL and GANs, image restoration for specific scenarios, such as medical image enhancement and super-resolved identity recognition, are still facing challenges. How to ensure visually realistic restoration while avoiding hallucination or mode- collapse? How to make sure the visually plausible results do not contain hallucinated features jeopardizing downstream tasks such as pathology identification and subject identification? Here we propose to resolve these challenges by coupling the GAN based image restoration framework with another task-specific network. With medical imaging restoration as an example, the proposed model conducts additional pathology recognition/classification task to ensure the preservation of detailed structures that are important to this task. Validated on multiple medical datasets, we demonstrate the proposed method leads to improved deep learning based image restoration while preserving the detailed structure and diagnostic features. Additionally, the trained task network show potentials to achieve super-human level performance in identifying pathology and diagnosis. Further validation on super-resolved identity recognition tasks also show that the proposed method can be generalized for diverse image restoration tasks.
- Keywords: Task-GAN: Improving Generative Adversarial Network for Image Restoration
- TL;DR: Couple the GAN based image restoration framework with another task-specific network to generate realistic image while preserving task-specific features.