.png){kind=icon}
.png){kind=icon}
Return
MP-IQA: a multidimensional perceptual fusion framework for AIGC image quality assessment
.png){kind=icon}
PRE
AI
L
J
C
DOI：10.1007/s00371-026-04682-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The rapid development of artificial intelligence-generated content (AIGC) has created an urgent need for reliable image quality assessment. However, objectively and effectively evaluating the visual quality and semantic fidelity of AIGC images remains challenging, limiting model optimization and fair comparison. Existing image quality assessment (IQA) metrics are mainly designed for natural images and often rely on single-dimensional features, making them insufficient for capturing the complex perceptual quality of AIGC images. To address this problem, we first constructed the AIGCIQA-3 M dataset, which contains 200 manually verified prompts, 3600 images generated by DALL $$\cdot $$ E, Stable Diffusion, and Midjourney, and multidimensional human subjective scores. A systematic analysis of 17 existing IQA metrics revealed generally weak correlations with human perception of AIGC content. Based on this analysis, we propose MP-IQA, a lightweight multidimensional perceptual fusion framework that integrates human preference signals and quality-anomaly perception signals to predict the subjective quality of AIGC images more stably. MP-IQA employs a two-step feature selection strategy and ridge regression to learn perceptually relevant feature representations. Experimental results show that MP-IQA outperforms single metrics in both PLCC and SROCC. By incorporating multidimensional perceptual cues, the proposed framework better aligns with human perceptual mechanisms and provides a more robust and comprehensive benchmark for AIGC image quality assessment. Data are available at https://www.kaggle.com/datasets/drmalin/mp-iqa.
Keywords:
AI-generated content
Image quality assessment
Multidimensional fusion
Mean opinion score
Feature fusion
Journal
IF:
2.9
Papers:
4.5K
Citations:
6.5K
Organization
Cited Papers