Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment

TL;DR

BIQA models trained on synthetic distortions often fail to generalize to authentic distortions due to domain gaps. The authors present DGQA, a distortion-guided unsupervised domain adaptation framework that selects source domains most similar to the target via a multi-source domain classifier and trains an IQA regressor on the selected data, thereby reducing negative transfer. They formalize UMDA for BIQA and demonstrate that aligning source-target distributions through targeted source-domain selection yields substantial cross-domain gains on synthetic-to-authentic and synthetic-to-algorithmic tasks, while remaining compatible with existing model-based BIQA methods. Extensive experiments across LIVEC, KonIQ-10k, BID, KADID-10k, KADIS-700k, and PIPAL validate improved performance with substantially less labeled data, highlighting practical benefits for real-world BIQA deployment and annotation efficiency.

Abstract

The annotation of blind image quality assessment (BIQA) is labor-intensive and time-consuming, especially for authentic images. Training on synthetic data is expected to be beneficial, but synthetically trained models often suffer from poor generalization in real domains due to domain gaps. In this work, we make a key observation that introducing more distortion types in the synthetic dataset may not improve or even be harmful to generalizing authentic image quality assessment. To solve this challenge, we propose distortion-guided unsupervised domain adaptation for BIQA (DGQA), a novel framework that leverages adaptive multi-domain selection via prior knowledge from distortion to match the data distribution between the source domains and the target domain, thereby reducing negative transfer from the outlier source domains. Extensive experiments on two cross-domain settings (synthetic distortion to authentic distortion and synthetic distortion to algorithmic distortion) have demonstrated the effectiveness of our proposed DGQA. Besides, DGQA is orthogonal to existing model-based BIQA methods, and can be used in combination with such models to improve performance with less training data.

Figures (3)

• Figure 1: The performance on the authentically distorted dataset LIVEC ghadiyaram2015massive when the baseline is trained on LIVE 2006A, KADID-10k kadid10k, and the sub-set of KADID-10k selected by the proposed DGQA. Ref. and Dist. denote the number of the reference images and the number of the distortion in this dataset, respectively.
• Figure 2: (a) are some representative images in LIVEC, (b) are some typical images in the part domains of KADID-10k whose styles are somewhat similar to that of LIVEC, and (c) are some typical images in the part domains of KADID-10k whose styles are quite different from LIVEC.
• Figure 3: The framework of the proposed DGQA. By performing similar domain selection, the source domains with a large gap to the target domain are sieved out. The sieved similar domains $\mathcal{D}_{sim}$ are used to train the model, which reduces the negative transfer from outlier source domains to the target domain and improves the model's performance on the target domain.