Psychology-Inspired AGIQA for Improved Image Quality Assessment

This paper addresses the problem of semantic drift in existing AGIQA models, where image embeddings show inconsistent similarities to grade descriptions. It proposes a novel approach inspired by psychometrics, specifically the Graded Response Model (GRM), to improve the reliability and performance of image quality assessment. The use of an Arithmetic GRM (AGQG) module offers a plug-and-play advantage and demonstrates strong generalization capabilities across different image types, suggesting its potential for future IQA models.

• •Addresses the problem of semantic drift in AGIQA models.
• •Proposes a novel approach inspired by psychometrics (GRM).
• •Introduces an Arithmetic GRM (AGQG) module.
• •AGQG offers plug-and-play benefits and improves performance.
• •Demonstrates strong generalization across different image types.