Training Quality-Aware Filters for No-Reference Image Quality Assessment

With the rapid increase of digital imaging and communication technology usage, there's now great demand for fast and practical image quality assessment (IQA) algorithms that can predict an image's quality as consistently as humans. The authors propose a general-purpose, no-reference image quality assessment (NR-IQA) with the goal of developing a model that does not require prior knowledge about nondistorted reference images and the types of distortions. The key is to obtain effective image representations using learning quality-aware filters (QAFs). Unlike other regression models, they also use a random forest to train the mapping from the feature space. Extensive experiments conducted on the LIVE and CSIQ datasets demonstrate that the proposed NR-IQA metric QAF can achieve better prediction performance than the other state-of-the-art approaches in terms of both prediction accuracy and generalization capability.

Previous Image Quality Assessment Approaches

For simplicity, most early NR-IQA algorithms assume that the image under consideration is affected by one or several types of distortion, such as blocking, ringing, blur, and compression. Generally, these approaches extract distortion-specific features that relate to the loss of visual quality. They can perform blind IQA only when the type of distortion is known before-hand. Hence, the applicability of these kinds of methods is limited.