Fine-grained Image Quality Assessment for Perceptual Image Restoration

A research team from multiple institutions has introduced FGResQ, a groundbreaking AI model designed specifically for evaluating the quality of restored images. The model addresses a critical gap in perceptual image restoration (IR), where existing image quality assessment (IQA) metrics struggle to distinguish fine-grained differences between AI-enhanced images. The researchers first created FGRestore, the first comprehensive dataset for this purpose, containing 18,408 restored images across six common IR tasks like denoising and super-resolution, along with 30,886 fine-grained pairwise preference annotations.

FGResQ's architecture features a dual approach combining coarse-grained score regression with fine-grained quality ranking, allowing it to capture subtle quality differences that traditional metrics miss. Extensive benchmarking revealed significant inconsistencies between existing IQA evaluations and actual restoration quality, demonstrating why specialized tools were needed. The model has been accepted by AAAI2026 and its code and weights have been publicly released, enabling developers to better optimize their image restoration algorithms and compare performance more accurately across different approaches.