Human-label-free Image Quality Assessment via Bayesian Inference and Adaptive Transition Matrix

Blind image quality assessment (BIQA) models heavily depend on expensive human subjective ratings. To reduce this dependency, full-reference IQA (FR-IQA) metrics are used as surrogates for generating pseudo-labels. However, FR-IQA-based methods introduce substantial label noise that degrades assessment accuracy. In this research, an Adaptive Transition Matrix Learning-based BIQA (ATML-BIQA) method is proposed that learns from synthetically degraded images and multiple annotators without requiring human labels or suffering from noisy label effects. The method operates in four stages: First, distorted images are synthesized from high-quality references and paired randomly. Ten FR-IQA models then assign pseudo-binary labels indicating relative perceptual quality. Second, a CNN-based BIQA model is trained on these binary labels to generate Bayesian optimal labels. Third, an Adaptive Instance Transition Matrix (AITM) is designed to model the relationship between noisy pseudo-labels and optimal labels, capturing the correlation between noisy annotations and latent ground truth for effective label noise correction. Finally, the BIQA model is retrained with corrected labels to produce final quality scores. This research significantly reduces the reliance of BIQA on expensive manual annotations and overcomes the accuracy limitations of traditional proxy labels by adaptively modeling annotation noise. Experimental results demonstrate that while enhancing model generalization, this method provides an efficient and robust new paradigm for quality assessment tasks in unsupervised or weakly supervised environments.