Skip to content

Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution

Conference: ICCV 2025 arXiv: 2502.06476 Code: GitHub Area: LLM Pretraining Keywords: Image Quality Assessment, Intrinsic Scale, Weak Labels, Multi-Scale Perception, Subjective Annotation

TL;DR

This paper introduces Image Intrinsic Scale (IIS)—the maximum scaling factor at which an image exhibits its highest perceptual quality—and proposes the IISA task, constructs a dataset of 785 images with expert annotations, and presents a weak-label training strategy (WIISA) that consistently improves IIS prediction across multiple NR-IQA methods.

Background & Motivation

Image Quality Assessment (IQA) is a core task in computer vision, yet a critically overlooked problem is that the relationship between image quality and spatial resolution (scaling factor) has never been systematically quantified.

In practice, a paradoxical phenomenon is commonly observed: - Downscaling an image can reduce noise grain and make blur less noticeable, potentially improving perceptual quality. - Excessive downscaling causes loss of high-frequency details (e.g., bird feather texture), degrading quality again. - Therefore, an optimal scaling factor exists that strikes the best balance between distortion removal and detail preservation.

This motivates a fundamental question: given an image, what is its optimal display scale?

Limitations of prior work: 1. Traditional camera performance metrics (e.g., P-MP, MTF) measure resolving power only under ideal laboratory conditions and are not applicable to real-world photographs. 2. NR-IQA methods assess quality at a fixed resolution and cannot answer the practical question of "at what size should this image be displayed." 3. The KonX dataset, while providing multi-resolution annotations for the first time, covers only 3 discrete scales, which is far too coarse.

Method

Overall Architecture

This work comprises three mutually reinforcing contributions: 1. Concept Definition: Image Intrinsic Scale (IIS), defined as the maximum scaling factor at which an image exhibits its highest perceptual quality. 2. Dataset Construction: IISA-DB, an expert-annotated dataset of 785 images. 3. Weak-Label Strategy: WIISA, which derives multiple weak-label training samples from a single annotation.

Key Designs

  1. Formal Definition of IIS: Let \(I^s\) denote image \(I\) scaled to factor \(s\), and \(Q(I^s)\) its perceptual quality. IIS is defined as:
\[\Omega(I) = \max\left(\operatorname*{argmax}_{s_{lb} \leq s \leq 1} Q(I^s)\right)\]

where the lower bound is \(s_{lb} = 0.05\), since quality assessment at extremely small sizes is unreliable. The \(\max\) operator selects the most informative scale when multiple equally optimal scales exist.

  1. WIISA Weak-Label Generation Strategy: This is the core methodological contribution. The key insight is that once the IIS \(\Omega(I)\) of the original image is known, the IIS of its downscaled versions can be derived via a piecewise function:
\[\overline{\Omega}(I^s) = \begin{cases} 1 & s_{lb} \leq s \leq \Omega(I) \\ \frac{\Omega(I)}{s} & \Omega(I) < s \end{cases}\]

Intuitive explanation: - When \(s \leq \Omega(I)\), the image is already at its optimal state and needs no further downscaling, so IIS = 1. - When \(s > \Omega(I)\), the image must be downscaled to \(\Omega(I)/s\) to reach the optimum.

Based on this, \(n_{wl}=2\) scales greater than \(\Omega(I)\) are randomly sampled from each annotated pair \((I, \Omega(I))\) to generate weak-label pairs \((I^{s_i}, \overline{\Omega}(I^{s_i}))\). This strategy is applied online during training, adding \(B \cdot n_{wl}\) weak-label samples per batch.

  1. Subjective Annotation Methodology: A web-based annotation tool called ZOVI (Zoom Viewer) was developed, in which annotators use a slider to reduce scale from \(s=1\) to \(s_{lb}=0.05\) and identify the maximum scale beyond which no further quality improvement is perceived. Each annotator labels each image twice (with a gap of several days), and batches with SRCC < 0.5 are re-annotated. Final IIS values are aggregated using the geometric mean (due to the nonlinear nature of scale space), referred to as MOIS.

Loss & Training

  • The original loss functions of each NR-IQA method remain unchanged (e.g., TOPIQ uses MSE + ranking loss).
  • WIISA modifies only the data sampling: downscaled image versions and their weak IIS labels are generated automatically each batch.
  • Lanczos interpolation is used for rescaling, consistent with the annotation phase.
  • 10-fold cross-validation is applied, with median test performance reported.

Key Experimental Results

Main Results

Method Training SRCC ↑ PLCC ↑ RMSE ↓ MAE ↓
DBCNN Base 0.755 0.761 0.093 0.074
DBCNN +WIISA 0.776 0.780 0.090 0.069
TOPIQ Base 0.764 0.762 0.098 0.078
TOPIQ +WIISA 0.808 0.805 0.088 0.069
CONTRIQUE Base 0.618 0.635 0.114 0.090
CONTRIQUE +WIISA 0.631 0.651 0.106 0.083
ARNIQA Base 0.651 0.650 0.105 0.082
ARNIQA +WIISA 0.687 0.672 0.103 0.079

WIISA consistently improves performance across all 6 methods, with a maximum relative gain of 5%.

Ablation Study

Configuration SRCC PLCC RMSE MAE Note
Base (no weak labels) 0.764 0.762 0.098 0.078 Baseline
\(n_{wl}=1\) 0.803 0.801 0.090 0.072 1 weak label
\(n_{wl}=2\) (WIISA) 0.808 0.805 0.088 0.069 Optimal
\(n_{wl}=3\) 0.788 0.785 0.096 0.077 Redundancy from too many weak labels
\(\delta=0.50\) 0.795 0.780 0.097 0.076 Threshold too low
\(\delta=0.80\) 0.802 0.800 0.089 0.069 Threshold too high
Bilinear interpolation 0.799 0.796 0.089 0.070 Minor effect of interpolation method

Key Findings

  • Zero-shot transfer fails: Pretrained NR-IQA models applied directly to IIS prediction perform poorly (TOPIQ on SPAQ achieves only SRCC 0.475), indicating that IIS and conventional quality scores represent distinct tasks.
  • IISA-DB annotation reliability: The average confidence interval is 0.057, comparable to the high-reliability NR-IQA dataset KonX (0.046).
  • Concavity assumption validated: 90% (378/420) of image triplets in KonX satisfy the concave quality-scale function assumption.
  • WIISA method-agnostic: All methods ranging from supervised learning to self-supervised and VLM-based approaches benefit from WIISA.

Highlights & Insights

  • New task definition: IIS addresses an eminently practical yet neglected question—"at what display size does this image look best?"
  • Bootstrapped data augmentation: WIISA elegantly exploits the mathematical properties of IIS to derive multiple training samples from a single annotation.
  • Cross-domain application potential: Applicable to image storage optimization, print size selection, super-resolution evaluation, and dataset construction.
  • Annotation methodology contribution: The ZOVI tool, combined with double annotation and geometric mean aggregation, constitutes a reusable subjective experiment paradigm.

Limitations & Future Work

  • The dataset scale is limited (785 images), which may be insufficient for training large models.
  • The concavity assumption does not hold for all images (10% exceptions), and handling of edge cases warrants further investigation.
  • Weak labels can only be generated for scales \(s > \Omega(I)\), leaving the low-scale regime uncovered.
  • Display device variation is not considered (IIS may shift with different PPI values).
  • Only Lanczos interpolation is evaluated; the effect of learned super-resolution methods on IIS remains unexplored.
  • The distinction from effective resolution is noteworthy: effective resolution concerns information preservation, whereas IIS targets perceptual quality maximization.
  • The weak-label approach is generalizable to other vision tasks requiring continuous-valued annotations (e.g., depth estimation, saliency detection).
  • IISA exhibits greater sensitivity than conventional NR-IQA and may be used to evaluate subtle image processing differences.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Introduces a novel and practical task with a clear and concise formulation.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Six methods with comprehensive ablations and reliability analysis, though dataset scale is modest.
  • Writing Quality: ⭐⭐⭐⭐⭐ Well-structured with a strong balance between mathematical rigor and intuitive explanation.
  • Value: ⭐⭐⭐⭐ Opens a new research direction on quality–resolution interaction with broad application prospects.