NEW LEARNING FRAMEWORKS FOR BLIND IMAGE QUALITY ASSESSMENT MODEL

The focus of this thesis is on image quality assessment, specifically for problems of assessing the quality of an image blindly or without reference information. There are significant efforts over the last decade in developing objective blind models that can assess image quality as perceived by humans. Various models have been introduced, achieving highly competitive performances and high in correlation with subjective perceptual measures. However, there are still limitations on these models before they can be viable replacements to traditional image metrics over a wide range of image processing applications. This thesis addresses several limitations. The thesis first proposes a new framework to learn a blind image quality model with minimal training requirements, operates locally and has ability to identify distortion in the assessed image. To increase the model’s performance, the thesis then modifies the framework by considering an aspect of human vision tendency, which is often ignored by previous models. Finally, the thesis presents another framework that enable a model to simultaneously learn quality prediction for images affected by different distortion types.