Title:Image Quality Assessment: Exploring Regional Heterogeneity via Response of Adaptive Multiple Quality Factors in Dictionary Space

Abstract:Given that the factors influencing image quality vary significantly with scene, content, and distortion type, particularly in the context of regional heterogeneity, we propose an adaptive multi-quality factor (AMqF) framework to represent image quality in a dictionary space, enabling the precise capture of quality features in non-uniformly distorted regions. By designing an adapter, the framework can flexibly decompose quality factors (such as brightness, structure, contrast, etc.) that best align with human visual perception and quantify them into discrete visual words. These visual words respond to the constructed dictionary basis vector, and by obtaining the corresponding coordinate vectors, we can measure visual similarity. Our method offers two key contributions. First, an adaptive mechanism that extracts and decomposes quality factors according to human visual perception principles enhances their representation ability through reconstruction constraints. Second, the construction of a comprehensive and discriminative dictionary space and basis vector allows quality factors to respond effectively to the dictionary basis vector and capture non-uniform distortion patterns in images, significantly improving the accuracy of visual similarity measurement. The experimental results demonstrate that the proposed method outperforms existing state-of-the-art approaches in handling various types of distorted images. The source code is available at this https URL.