Causal Disentanglement for Full-Reference Image Quality Assessment
A recent study published on arXiv (2604.21654) presents a novel full-reference image quality assessment (FR-IQA) framework that utilizes causal inference and decoupled representation learning, moving away from conventional deep feature comparison techniques. This method conceptualizes degradation estimation as a causal disentanglement process, driven by interventions on latent representations. Initially, it separates degradation from content representations by leveraging content invariance between the reference and distorted images. Additionally, drawing inspiration from human visual masking, a masking module captures the causal links between image content and degradation features, enabling the extraction of content-influenced degradation characteristics from the distorted images. The research team behind this work is credited in the paper.
Key facts
- Paper arXiv:2604.21654 proposes a new FR-IQA paradigm
- Approach uses causal inference and decoupled representation learning
- Degradation estimation is framed as a causal disentanglement process
- Content and degradation representations are decoupled using content invariance
- A masking module models causal relationship between content and degradation
- Method is inspired by human visual masking effect
- Published on arXiv
Entities
Institutions
- arXiv