Title: HiRQA: Hierarchical Ranking and Quality Alignment for Opinion-Unaware Image Quality Assessment

Abstract:

Although significant strides have been made in no-reference image quality assessment (NR-IQA), the field remains constrained by dataset biases and an over-reliance on subjective annotations, which limit generalization capabilities. To address these challenges, we introduce HiRQA (Hierarchical Ranking and Quality Alignment), a self-supervised, opinion-unaware framework. This approach generates hierarchical, quality-aware embeddings by integrating ranking and contrastive learning mechanisms. In contrast to existing methods that require pristine reference images or additional modalities during inference, HiRQA derives quality scores solely from the input image.

Our methodology features a novel higher-order ranking loss that guides quality predictions via relational ordering among distortion pairs, alongside an embedding distance loss that ensures alignment between feature distances and perceptual disparities. Additionally, a contrastive alignment loss applied during training, driven by structured textual prompts, refines the learned representations. Although HiRQA is trained exclusively on synthetic image distortions, it demonstrates robust generalization to real-world degradations. Comprehensive evaluations confirm its effectiveness across various unseen distortions, including lens flare, haze, motion blur, and low-light scenarios. For applications requiring real-time performance, we present HiRQA-S, a streamlined variant capable of processing an image in just 3.5 milliseconds. Extensive testing on both synthetic and authentic benchmarks highlights HiRQA’s competitive accuracy, strong generalization potential, and scalability. The HiRQA model and inference pipeline are accessible at: https://github.com/uf-robopi/HiRQA.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC