Degradation Category Rating Overview

• Degradation Category Rating is a framework that translates raw degradation evidence into operational categories by quantifying type, severity, and confidence.
• It spans diverse fields—from image enhancement to software engineering—adapting methods like forced-choice classification, spectral analysis, and composite scoring.
• The approach integrates layered assessments (categorization, confidence measurement, and downstream routing) to guide targeted restoration, maintenance, or risk mitigation.

Degradation Category Rating denotes a family of rating schemes that convert degradation evidence into operational categories, severity scores, or health/risk labels. Recent usages are not standardized across fields: in image enhancement it denotes a forced-choice degradation classifier and routing signal; in image robustness it denotes a jointly causal and perceptual categorization with measured severity; in robust multimodal reasoning it is tied to normalized degradation intensities and structured reasoning chains; in all-in-one restoration it is expressed through a frequency-domain representation; in software engineering it refers to composite ratings over architectural, code, and process debt; and in quality engineering it denotes model-based health categories derived from degradation paths, failure thresholds, and near-term reliability (Cai et al., 5 Jun 2025, Becker et al., 15 May 2026, Tang et al., 19 Dec 2025, Huang et al., 17 May 2026, Ahmad et al., 19 Jul 2025, Clark et al., 19 Jul 2025).

1. Conceptual scope and recurring structure

A recurring pattern across these usages is the separation of degradation assessment into three layers. First, a system identifies one or more degradation categories. Second, it quantifies the strength, confidence, or risk associated with those categories. Third, it uses the resulting rating to drive a downstream action such as restoration routing, robustness analysis, maintenance prioritization, or architectural remediation. This suggests that “Degradation Category Rating” is best understood as an operational interface between raw degradation evidence and a decision process rather than as a single canonical metric.

The literature also differs on whether the rating scheme is native to the original method or a derived operationalization. The vision-language image-enhancement framework provides an exact four-way classification prompt and routing logic, while its probability-based rating layer is a practical synthesis from token log-probabilities rather than a separately trained calibration head (Cai et al., 5 Jun 2025). The architectural degradation review provides definitions, causes, metrics, measurement techniques, tools, and remediation gaps, but explicitly notes that the composite scoring rubric is proposed rather than prescribed by the review itself (Ahmad et al., 19 Jul 2025). The quality-engineering material likewise motivates a DCR as a category mapping from validated degradation models to actionable health states, but industry-specific cutoffs remain customizable (Clark et al., 19 Jul 2025).

A common misconception is that category labels alone are sufficient. Several of the cited works argue, directly or indirectly, that nominal category identifiers are inadequate when degradations are mixed, when backend severities are incomparable, or when organizational and process factors amplify technical symptoms. This is why recent formulations add confidence margins, normalized intensities, measured severity tuples, composite weighted sums, or first-passage reliability.

2. Vision-language classification and restoration routing

In "Degradation-Aware Image Enhancement via Vision-Language Classification" (Cai et al., 5 Jun 2025), the rating problem is framed as a single-pass, zero-shot, forced-choice decision over four mutually exclusive categories: A. Super-resolution degradation, including noise, blur, JPEG compression, and general low-resolution or low-quality signal; B. Reflection artifacts; C. Motion blur; and D. No visible degradation. The exact prompt is: “Analyze this image and determine the type of image degradation it exhibits. Categorize it into one of the following degradation types: A. Super-resolution degradation (including noise, blur, JPEG compression); B. Reflection artifacts; C. Motion blur; D. No visible degradation (high-quality image). Provide a simple result, i.e. A, B, C, or D.” The paper uses Qwen2.5-VL as-is, in zero-shot mode, without fine-tuning.

A practical Degradation Category Rating can be derived from the per-option token log-probabilities for A, B, C, and D. Let $z \in \mathbb{R}^4$ denote the logits or log-probabilities for the four options. A probability vector is then computed by temperature scaling,

$$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$

with $T=1$ if no calibration is performed. This yields a per-category rating $\mathrm{Rating}(c)=p_c$, a confidence margin

$m = p_c - \max_{j \ne c} p_j,$

and an overall degraded-versus-clean severity proxy

$S_{\mathrm{degraded}} = p_A + p_B + p_C.$

A user-facing scale is then

$$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$

with rounding. The details further recommend routing to $\hat{c}=\arg\max_i p_i$, passing through only if $\hat{c}=D$ and $S_{\mathrm{degraded}}<0.3$, and treating $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$0 as ambiguous.

Category	Meaning	Restoration branch
A	Super-resolution-related degradation	InvSR + PaddleOCR + Real-ESRGAN fusion
B	Reflection artifacts	YOLO/YOSO mask + LaMa + NAFNet
C	Motion blur	NAFNet + CodeFormer on faces
D	No visible degradation	Pass-through

The restoration stacks are category-specific. Category A applies InvSR to the whole image, uses PaddleOCR to detect text regions, enhances those patches with Real-ESRGAN, and fuses them back because diffusion SR “excels on textures but struggles with text,” whereas Real-ESRGAN is strong on text and deartifacting. Category B treats strong reflections as occlusions and weak reflections as degradations: YOLO detects strong reflection regions, YOSO refines a pixel-level mask, LaMa inpaints masked areas, and NAFNet reduces residual weak reflections globally. Category C deblurs with NAFNet and, if faces are detected, refines facial details with CodeFormer. Category D is passthrough.

The system is deliberately single-label. Mixed degradations such as motion blur plus compression are forced into one of the four classes, and the paper does not implement multi-label classification or severity estimation within categories. The paper also shows qualitative results only: it does not report PSNR, SSIM, LPIPS, FID numbers, routing accuracy, confusion matrices, or classification dataset splits. Temperature scaling on a small labeled set is recommended for reliable ratings, but this calibration step is not part of the original paper.

3. Causal-source and perceptual-effect ratings for image robustness

"A Causally Grounded Taxonomy for Image Degradation Robustness Evaluation" formalizes Degradation Category Rating as a dual-axis assignment plus a descriptive severity layer (Becker et al., 15 May 2026). Each degradation is labeled by its dominant causal source and dominant perceptual effect. The causal-source axis contains Environment (E), Sensor/Optics (S), ISP/Renderer/Codec (R), and Transfer/System (T). The perceptual-effect axis contains Noise (N), Blur (B), Weather/Medium (WX), Compression/Quantization (CP), Color/White Balance (CL), Illumination/Exposure (IL), Geometry/Spatial (GD), Resolution/Sampling (RZ), Occlusion/Obstruction (OC), Texture/Sharpness/Contrast (TX), and Temporal/Video (TV).

This dual-axis taxonomy is explicitly intended to reconcile communities that traditionally group degradations differently. Synthetic corruption benchmarks group by operator type, IQA literature groups by perceptual distortion, and physical imaging analyses group by pipeline stage. Under the dual-axis scheme, motion blur is typically $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$1, JPEG artifacts are $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$2, haze or fog is $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$3, and rolling shutter is $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$4. The framework makes explicit that similar-looking effects can arise from different causes and therefore interact differently with downstream systems.

Severity is not redefined by replacing native backend parameters. Instead, the framework introduces a measurement layer based on full-reference metrics. For a clean reference $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$5 and degraded image $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$6, it reports:

$$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$7

$$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$8

and

$$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$9

LPIPS is used as a learned perceptual distance. The paper recommends reporting a tri-metric tuple $T=1$0 and often interpreting $T=1$1 as a severity proxy.

The paper demonstrates the framework through COCO-Degradation on COCO val2017. It integrates 19 ImageNet-C style operators, 24 IQA-inspired distortions from KADID-10k/ARNIQA, and 16 real-camera or system-level failures. A central finding is that equal nominal severities are not comparable across backends. Measured strengths make this visible: for ImageNet-C severity 3, motion blur has approximately $T=1$2, $T=1$3 dB, and $T=1$4, whereas JPEG has approximately $T=1$5, $T=1$6 dB, and $T=1$7. This directly challenges the assumption that a shared native severity index has backend-independent semantics.

The paper also notes practical edge cases. Geometric distortions and strong motion can corrupt full-reference measurements through misregistration; alignment by ECC or homography, windowed SSIM, or cautious interpretation is advised. In unpaired scenarios, the full-reference layer does not apply directly, and no-reference IQA becomes necessary.

4. Normalized intensity and structured reasoning in robust multimodal models

"Robust-R1: Degradation-Aware Reasoning for Robust Visual Understanding" uses degradation categories as explicit reasoning objects rather than as implicit nuisance variables (Tang et al., 19 Dec 2025). Its dataset synthesizes degradations across four real-world visual processing stages: Acquisition, Transmission, Environment, and Postprocessing. The exact category list is: Lens Blur, Lens Flare, Motion Blur, Dirty Lens, Saturation; Compression, Block Change, Shifting, Scan Lines; Darkness, Atmospheric Turbulence, Noise, Color Diffusion; and Sharpness Change, Graffiti, Watermark Damage. Each degraded sample is parameterized by type $T=1$8 and normalized intensity $T=1$9, with

$\mathrm{Rating}(c)=p_c$0

and intensities sampled as $\mathrm{Rating}(c)=p_c$1.

The model’s reasoning chain explicitly serializes degradation information: <TYPE> ... <TYPE_END>, <INFLUENCE> ... <INFLUENCE_END>, <REASONING> ... <REASONING_END>, <CONCLUSION> ... <CONCLUSION_END>. This decomposition is accompanied by supervised fine-tuning and GRPO-based alignment. The degradation-alignment reward is

$\mathrm{Rating}(c)=p_c$2

the length reward is

$\mathrm{Rating}(c)=p_c$3

and the composite reward is $\mathrm{Rating}(c)=p_c$4. The paper further reports an empirical relation

$\mathrm{Rating}(c)=p_c$5

which underlies dynamic reasoning depth scaling.

A Degradation Category Rating in this setting is direct when the dataset annotation is available: the category is $\mathrm{Rating}(c)=p_c$6 and the severity is $\mathrm{Rating}(c)=p_c$7. The commonly used low, mid, and high bins in evaluation are not mapped from explicit thresholds in the paper; a low/mid/high partition of $\mathrm{Rating}(c)=p_c$8 is therefore a derived discretization rather than a native claim. For compound degradations, the sum of intensities is a natural derived aggregate because the paper’s reasoning-length mechanism scales with total intensity.

The ablations clarify why this formulation matters. Removing structured reasoning yields Overall 0.4471; removing $\mathrm{Rating}(c)=p_c$9 yields 0.4880; removing $m = p_c - \max_{j \ne c} p_j,$0 yields 0.4907; and the full SFT+RL model reaches 0.5017 on R-Bench. This establishes that category recognition, intensity alignment, and adaptive reasoning depth are separable contributors to robustness.

5. Spectral rating with the Degradation Frequency Curve

"Degradation Frequency Curve: An Explicit Frequency-Quantified Representation for All-in-One Image Restoration" replaces discrete labels with a measurable spectral profile (Huang et al., 17 May 2026). Given degraded image $m = p_c - \max_{j \ne c} p_j,$1 and clean image $m = p_c - \max_{j \ne c} p_j,$2 during training, or an estimated clean image $m = p_c - \max_{j \ne c} p_j,$3 during blind inference, the residual is $m = p_c - \max_{j \ne c} p_j,$4 or $m = p_c - \max_{j \ne c} p_j,$5. The frequency plane is partitioned into $m = p_c - \max_{j \ne c} p_j,$6 soft radial bands by Gaussian masks $m = p_c - \max_{j \ne c} p_j,$7, and band-wise residual-to-degraded energy ratios are computed:

$m = p_c - \max_{j \ne c} p_j,$8

$m = p_c - \max_{j \ne c} p_j,$9

Equivalently,

$S_{\mathrm{degraded}} = p_A + p_B + p_C.$0

followed by normalization across bands. The paper uses $S_{\mathrm{degraded}} = p_A + p_B + p_C.$1.

The DFC is not itself a categorical label, but it supports derived category ratings from band-mass patterns. The paper reports that haze is low-frequency dominant, rain is middle-frequency prominent, Gaussian noise is concentrated in high frequencies, and low-light is also low-frequency dominant, with severity reflected in peak amplitude changes. A derived rating can therefore use low-, mid-, and high-band masses, global slope, contrast, and entropy. To separate blur from noise, the details further propose a degraded-energy suppression index for high frequencies,

$S_{\mathrm{degraded}} = p_A + p_B + p_C.$2

where $S_{\mathrm{degraded}} = p_A + p_B + p_C.$3 is the degraded-image energy distribution. This is explicitly a practical extension rather than a native paper metric.

The DFC also acts as a reusable restoration prior. FCMS adaptively decomposes the curve into band-wise spectral tokens, and MMD uses those tokens for token-conditioned multi-band decoding. The training objective combines a spatial reconstruction term and a frequency-domain consistency term,

$S_{\mathrm{degraded}} = p_A + p_B + p_C.$4

with $S_{\mathrm{degraded}} = p_A + p_B + p_C.$5 and $S_{\mathrm{degraded}} = p_A + p_B + p_C.$6. The paper reports that $S_{\mathrm{degraded}} = p_A + p_B + p_C.$7 tokens works best, estimated DFC is reliable relative to GT-DFC with a negligible 0.01 dB gap, and DFC-guided tokenization improves unseen degradations and complex mixtures. On CDD11, DFC-IR exceeds MoCE-IR-S by +1.17 dB PSNR on average; on unseen underwater and desnowing degradations it gains +0.59 dB and +0.45 dB PSNR over MoCE-IR.

A limitation of DFC-based rating is that radial bands ignore orientation. This weakens direct sensitivity to directional artifacts such as compression blocking. The details therefore describe comb-mask augmentation for JPEG-like grid artifacts as an extension, not as an evaluated core component of the paper.

6. Architectural degradation ratings in software systems

In software engineering, “Degradation Category Rating” refers to the conversion of architectural degradation evidence into category scores over structural and socio-technical debt dimensions. "Architectural Degradation: Definition, Motivations, Measurement and Remediation Approaches" defines architectural degradation as the progressive divergence between a system’s implemented and intended architecture, arising from repeated violations of architectural decisions, rules, and principles and from cumulative code-level changes that undermine structural consistency, modularity, cohesion, and separation of concerns (Ahmad et al., 19 Jul 2025). The review describes an evolution from low-level conformance gaps and structural symptoms in the 1990s–2000s, through emphasis on maintenance and documentation breakdowns in 2009–2016, to an explicitly socio-technical conception in 2019–2024.

The review organizes causes into Architectural Debt, Code Debt, and Process Debt. Architectural Debt includes documentation issues, architectural smells, structural dependency issues, low modularity, explicit violations of architectural decisions, legacy architectures, and tooling limitations. Code Debt includes increased code complexity, code smells or antipatterns, duplicate code, uncontrolled code changes, hotspots and co-change patterns, system size increase, and technology evolution. Process Debt includes decentralized or distributed development, inadequate process execution, time pressure and constraints, unsupportive organizational culture, developer turnover, knowledge gaps, and insufficient system understandability.

The paper identifies 54 metrics and 31 measurement techniques. Prominent architectural metrics include architectural smells, coupling, cohesion, structural modularity, Decoupling Level, Dependency Cycle (DC), Cluster Factor (CF), Concern Overload (CO), Scattered Functionality (SF), Lack of cohesion of methods (LCM), Normalised Cumulative Component Dependency (NCCD), and conformance indicators such as # architectural inconsistencies and # divergences. Code-level indicators include # change, # file co-changes, Cross-Module Co-Changes (CMC), Inner-Module Co-Changes (IMC), active hotspots, # LOC, cyclomatic complexity, duplicated code, # package cycles, fan-in, fan-out, and defect-linkage counts. Architectural quality is strongly associated with smells, dependency cycles, high coupling, low cohesion, low modularity, and high dependency counts.

The review also catalogs tooling. Arcan detects architectural smells, Arcade collects architectural-smell data and measures architectural changes, Understand and Sonar support architecture recovery and structural analysis, SotoArc, Axivion/Bauhaus, and Structure101 provide broad architectural analysis, and JArchitect, SonarGraph, JITTAC Medic, and related systems perform conformance checking and rule enforcement. A major conclusion is that tool support is strong for detection but weak for ongoing or preventive remediation; socio-technical factors are especially under-instrumented.

The composite rating rubric described in the details is explicitly proposed rather than paper-native. It defines an Architectural Degradation Score $S_{\mathrm{degraded}} = p_A + p_B + p_C.$8, a Code Degradation Score $S_{\mathrm{degraded}} = p_A + p_B + p_C.$9, and a Process Degradation Score $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$0, each as a weighted sum of normalized metrics,

$$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$1

followed by an overall score

$$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$2

Suggested starting weights are $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$3, $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$4, and $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$5, and proposed bands are Low $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$6, Moderate $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$7, High $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$8, and Severe $$\mathrm{Score}_{1\text{–}5} = 1 + 4 \cdot S_{\mathrm{degraded}}$$9. The worked example yields $\hat{c}=\arg\max_i p_i$0, which falls in the High band. This operationalization reflects the review’s central claim that degradation is both technical and organizational.

7. Model-based health and risk categories in quality engineering

In quality engineering, Degradation Category Rating denotes a categorical health or risk label derived from a validated degradation model rather than from an image or architectural signal. "What Quality Engineers Need to Know about Degradation Models?" motivates such a rating as a mapping from continuous degradation measurements and predictions to actionable categories using current degradation level relative to a failure threshold $\hat{c}=\arg\max_i p_i$1, near-term failure risk, and degradation dynamics (Clark et al., 19 Jul 2025). Let $\hat{c}=\arg\max_i p_i$2 denote the observed degradation signal, $\hat{c}=\arg\max_i p_i$3 its mean path, and

$\hat{c}=\arg\max_i p_i$4

for increasing paths. At evaluation time $\hat{c}=\arg\max_i p_i$5 with planning horizon $\hat{c}=\arg\max_i p_i$6, the core quantities are distance to threshold, near-term reliability $\hat{c}=\arg\max_i p_i$7, failure probability $\hat{c}=\arg\max_i p_i$8, and rate or acceleration terms such as

$\hat{c}=\arg\max_i p_i$9

The details provide an example category mapping: Category A (Healthy) if $\hat{c}=D$0 and $\hat{c}=D$1; Category B (Watch) if $\hat{c}=D$2 or $\hat{c}=D$3; Category C (High Risk) if $\hat{c}=D$4 or $\hat{c}=D$5, or if $\hat{c}=D$6 is large; and Category D (Failing/Exceeded) if $\hat{c}=D$7 or $\hat{c}=D$8. These cutoffs are presented as customizable by industry and regulatory context.

The framework spans repeated-measures degradation testing, accelerated destructive degradation testing, and dynamic-covariate field tracking. It reviews General Path Models and stochastic-process models. For a Wiener process with drift,

$\hat{c}=D$9

the first-passage time to $S_{\mathrm{degraded}}<0.3$0 is inverse Gaussian, and DCR should use first-passage reliability rather than the simpler instantaneous probability $S_{\mathrm{degraded}}<0.3$1. For a Gamma process,

$S_{\mathrm{degraded}}<0.3$2

and reliability at horizon $S_{\mathrm{degraded}}<0.3$3 is $S_{\mathrm{degraded}}<0.3$4. This distinction is important because conflating state probability with threshold-crossing probability is a common error for non-monotone paths such as Wiener.

The worked Gamma-process example illustrates the logic. With $S_{\mathrm{degraded}}<0.3$5, $S_{\mathrm{degraded}}<0.3$6, $S_{\mathrm{degraded}}<0.3$7, and $S_{\mathrm{degraded}}<0.3$8, one obtains $S_{\mathrm{degraded}}<0.3$9, $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$00, and $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$01, which maps to Category B (Watch) because the expected level is between $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$02 and $$p_i = \mathrm{softmax}(z/T)_i = \frac{\exp(z_i/T)}{\sum_j \exp(z_j/T)},$$03 despite high reliability. The paper further reviews software support including nlme, ADDT, SPREDA, RStan, chemdeg, JMP, and ReliaSoft Weibull++, underscoring that DCR in this domain is inseparable from model validation, uncertainty quantification, acceleration links such as the Arrhenius transform, and conservative decision rules.

Across these domains, Degradation Category Rating serves the same general purpose—turning degradation evidence into an interpretable basis for intervention—but the object being rated differs fundamentally. In some settings it is a probability over forced-choice labels, in others a measured severity tuple, a normalized synthesis parameter, a spectral profile, a socio-technical composite score, or a reliability-grounded risk state. The term therefore denotes a class of operational rating interfaces rather than a single universal formalism.