MCIQA-2K: A Multi-Dimensional Dataset and No-Reference Quality Assessment Benchmark for Colorized Images
Abstract Overview
This paper introduces MCIQA-2K, a benchmark for no-reference image quality assessment of colorized images, motivated by the mismatch between human perception and conventional full-reference metrics in the inherently ambiguous colorization task. The dataset contains 2,000 colorized images derived from 400 COCO test images using five representative colorization models, with each image evaluated by 16 human raters. The annotations cover three perceptual dimensions: color smearing, semantic color misalignment, and global naturalness, demonstrating solid inter-rater reliability. Building on this benchmark, the authors propose MCIQA, a multi-branch NR-IQA framework that separately models the three dimensions and combines them for perceptual quality prediction.
Novelty
The primary novelty lies in introducing a colorization-specific, multi-dimensional NR-IQA benchmark paired with a dedicated assessment framework tailored to typical colorization failure modes. The architecture uniquely incorporates specialized branches for color smearing, semantic color misalignment, and global naturalness, generating interpretable spatial score maps via vision-language modeling and spatial masking.
Results
On MCIQA-2K, the specialized branches achieve leading performance on their corresponding dimensions, with MCIQA-CS reaching SRCC/PLCC of 0.8274/0.8213, MCIQA-SCM reaching 0.8902/0.8408, and MCIQA-GN reaching 0.8511/0.8571. Furthermore, the global naturalness branch exhibits strong generalization to broader IQA datasets, including achieving the top reported SRCC of 0.6373 on AGIQA-1K among compared methods.
Key Points
- MCIQA-2K provides 2,000 colorized images produced by five representative models and annotated across three perceptual dimensions by 16 human raters.
- The MCIQA framework utilizes dedicated branches for color smearing, semantic color misalignment, and global naturalness, outputting interpretable spatial score maps.
- Experimental results show that task-specific, dimension-aware modeling substantially outperforms generic full-reference and no-reference IQA baselines on colorized image evaluation.