Bridging the Perceptual Gap: Residual-Enhanced Downscaling and Manifold-Aware Perception Alignment Adaptation for NR-IQA

Leveraging Large Vision-Language Models like CLIP has recently set new benchmarks for No-Reference Image Quality Assessment (NR-IQA). However, the contrastive pretraining of CLIP inherently prioritizes semantic invariance, which often suppresses subtle perceptual signals, a phenomenon we term perceptual submergence. Furthermore, standard preprocessing techniques (e.g., cropping and interpolation) further exacerbate the loss of critical high-frequency quality cues. In this paper, we propose the Cross-modal Perception Alignment Adapter (CMPA), a manifold-aware framework designed to disentangle perceptual distortions from dominant semantics. CMPA introduces a Perception-Sensitive Feature Extractor (PFE) that projects CLIP features into a compact, low-dimensional subspace, explicitly magnifying distortion-induced off-manifold deviations. Subsequently, a Cross-Modal Perception Alignment Injector (PAI) aligns these features with quality-aware text anchors and re-injects them into the backbone. To ensure input fidelity, we also devise a Residual-enhanced Perceptual Downscaling strategy that adaptively compensates for resolution-induced information loss using Just Noticeable Difference (JND) guided frequency re-injection. Extensive evaluations on several benchmark datasets demonstrate that our approach significantly outperforms state-of-the-art methods, effectively recovering the perceptual signals submerged in semantic-dense representations.

Uncovering Perceptual Signals in Large Vision-Language Models for Image Quality Assessment Context and Challenge No-Reference Image Quality Assessment (NR-IQA)—the ability of algorithms to evaluate image degradation without a pristine reference—is a critical component of modern digital media processing. Recently, Large Vision-Language Models (LVLMs) like CLIP have established new benchmarks in this field. However, because these models are pre-trained to prioritize semantic content (e.g., object recognition), they inherently suppress subtle visual distortion cues. We identify this phenomenon as "perceptual submergence." Furthermore, routine preprocessing steps, such as image cropping and interpolation, exacerbate this issue by discarding critical high-frequency details necessary for accurate quality evaluation. Proposed Solution To overcome the limitations of semantic bias and information loss, we introduce the Cross-modal Perception Alignment Adapter (CMPA). This framework is specifically designed to disentangle subtle perceptual distortions from dominant semantic features. The core mechanism involves a Perception-Sensitive Feature Extractor (PFE) that magnifies distortion-induced variations, isolating them into a focused, low-dimensional subspace. These isolated features are then aligned with quality-aware textual anchors to ensure the model accurately interprets the specific type of degradation. Additionally, to address the loss of detail during image resizing, we implement a Residual-enhanced Perceptual Downscaling strategy. By leveraging the psychophysical concept of Just Noticeable Difference (JND), this strategy adaptively preserves and re-injects critical frequency information, ensuring that high-fidelity input cues are maintained throughout the network. Significance and Impact Extensive empirical evaluations across multiple benchmark datasets demonstrate that our framework significantly outperforms existing state-of-the-art methods. By effectively recovering perceptual signals that are otherwise submerged in semantic-heavy representations, this approach bridges the gap between high-level semantic understanding and low-level perceptual evaluation, offering a more robust and human-aligned solution for automated image quality assessment.