From Distortion Manifold to Perceptual Quality: a Data Efficient Blind Image Quality Assessment Approach

Though current no-reference image quality assessment (NR-IQA) approaches have achieved impressive performance gain thanks to deep learning techniques, it is claimed that the risk of over-fitting exists. To improve model generalization ability, most of the current researches incorporate mass data to train or tune the data-driven models. However, the process of image data collection and quality label annotation is quite time-consuming and labour-intensive. Therefore, in this paper, we explore an alternative solution to promote model generalizability but with relatively small fractions of training data. Compared with previous approaches which make effort to approximate the whole complex image distribution, we propose to explicitly learn an image distortion manifold first, which lies in a much lower dimension space and also representative in capturing general degradation patterns. We then project the images to their perceived quality from the learned manifold to obtain quality predictions. Since the manifold embeds general distortion features despite of varying image contents, it can be learned with relatively small amount of samples. In order to learn the manifold and quality projection, we introduce a two-branched network to learn both low level distortions and high level semantics. We also propose a simple but efficient training framework, composing of a masked labelling strategy and a gradual weighting curriculum to fulfill the task. Thanks to the learned distortion manifold, the proposed model achieves superior generalizability compared with previous models. Extensive experiments demonstrate its effectiveness in terms of training with limited data, testing on large scale images, and with unseen types of distorted images.