Multiple hallucinated deep network for image quality assessment

Image Quality Assessment (IQA) refers to quantitatively evaluating the human perception about the quality of a distorted image. Blind IQA (BIQA) is a type of IQA that does not contain 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 some expert opinion, such as Natural Scene Statistics (NSS), to determine how far the distorted image is from the distribution of pristine samples.By emerging deep networks, several IQA methods have been proposed to use their capability in automatic feature extraction. The main challenge of available deep models is that they need many annotated samples for training to reach a desirable outcome which is costly. In this paper, inspiring the Human Visual System (HVS), we propose a Generative Adversarial Network (GAN) based approach. To this end, we sample multiple images from a submanifold of pristine data manifold by conditioning the network on the corresponding distorted image. Also, NSS features are used to improve the network training and conduct the training process on the right track.