Multiple hallucinated deep network for image quality assessment

Image Quality Assessment (IQA) refers to quantitative evaluation of the human's perception of a distorted image quality. Blind IQA (BIQA) is a type of IQA that does not include any reference or information about the distortion. Since the human brain has no information about the distortion type, BIQA is more reliable and compatible with the real world. Traditional methods in this realm used an expert opinion, such as Natural Scene Statistics (NSS), to measure the distance of a distorted image from the distribution of pristine samples. In recent years, many deep learning-based IQA methods have been proposed to use the ability of deep models in automatic feature extraction. However, the main challenge of these models is the need for many annotated training samples. In this paper, through the inspiration of Human Visual System (HVS), a Generative Adversarial Network (GAN)-based approach was proposed to address this problem. To this end, multiple images were sampled from a submanifold of the pristine data manifold by conditioning the network on the corresponding distorted image. In addition, NSS features were employed to improve the network training and conduct the training process on the right track. The testing results of the proposed method on three datasets confirmed its superiority over other the state-of-the-art methods. (c) 2023 Sharif University of Technology. All rights reserved.