EUVIMEDEUVIMED entdecken
Crossref

PeC-AGIQA: Perceptual Feature-Guided Contrastive Learning for AI-Generated Image Quality Assessment

Yang Yang, Yunhao Zhang, Xi Yang, Weiming Zhang

IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 17259-17273

DOI: 10.1109/tcsvt.2026.3718652 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Yang Yang und 3 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Innovation Program for Quantum Science and Technology, Science and Technology Major Project of Anhui Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The emergence of Artificial Intelligence Generated Content has revolutionized digital image creation, with AI-generated images (AGIs) now permeating diverse domains from creative arts to scientific visualization. Despite remarkable advancements, AGIs face a fundamental challenge: misalignment between generated images and both textual prompts and human visual perception, creating a semantic gap in quality assessment. Different from conventional Image Quality Assessment that focuses on content-agnostic distortions like noise and blur, AI-generated Image Quality Assessment (AGIQA) needs to address the issues of semantic consistency and perceptual alignment. This paper introduces PeC-AGIQA, a framework that learns feature representations guided by perceptual alignment which systematically shapes the feature embedding space to inherently encode human subjective quality relationships. The framework integrates three core components: (1) a dual-stream CLIP architecture that leverages both frozen and trainable encoders to balance established knowledge with adaptive learning for multimodal representation; (2) a lightweight alignment perception module that enhances text-image correspondence while reducing computational overhead; and (3) a perceptual feature-guided contrastive learning strategy that transforms continuous human opinion scores into structured quality intervals, enabling the model to recognize subtle quality distinctions similar to human evaluators by contrastive learning with the extracted feature. This hierarchical design progressively refines the model’s ability to capture perceptual nuances, moving from fundamental representation to the way that correlates more strongly with human judgments. Experiments on AGIQA benchmarks validate our approach’s superior performance, and effectiveness in achieving improved agreement with human subjective assessment on three public benchmarks.

Förderung

  • National Natural Science Foundation of China
  • Innovation Program for Quantum Science and Technology
  • Science and Technology Major Project of Anhui Province

Zitieren

APA

Yang, Y., Zhang, Y., Yang, X., & Zhang, W. (2026). PeC-AGIQA: Perceptual Feature-Guided Contrastive Learning for AI-Generated Image Quality Assessment. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 17259-17273. https://doi.org/10.1109/tcsvt.2026.3718652

Vancouver

Yang Y, Zhang Y, Yang X, Zhang W. PeC-AGIQA: Perceptual Feature-Guided Contrastive Learning for AI-Generated Image Quality Assessment. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):17259-17273. doi:10.1109/tcsvt.2026.3718652

BibTeX
@article{Yang2026,
  author = {Yang, Yang and Zhang, Yunhao and Yang, Xi and Zhang, Weiming},
  title = {PeC-AGIQA: Perceptual Feature-Guided Contrastive Learning for AI-Generated Image Quality Assessment},
  journal = {IEEE Transactions on Circuits and Systems for Video Technology},
  year = {2026},
  volume = {36},
  number = {11},
  pages = {17259--17273},
  doi = {10.1109/tcsvt.2026.3718652}
}
Kostenlos starten →