Image Quality Assessment using Semi-Supervised Representation Learning

Abstract

In this paper, we propose a framework for learning feature representations for Image Quality Assessment (IQA) using contrastive learning. To account for the absence of large-scale IQA dataset, we pretrain an image encoder to cluster images based on the image quality using synthetically distorted versions of pristine unlabeled images. Images of similar quality are grouped closer in embedding space, while simultaneously pushing apart images of dissimilar quality. In addition we show that, augmenting the contrastive learning task with downstream aware joint supervision results in feature representations that are more suitable and easily transferable for IQA specific tasks. We study the effectiveness of the learnt representations in downstream task of image quality prediction and show that our model achieves superior performance on both synthetically and authentically distorted IQA datasets when compared to other deep feature-based IQA methods.