Rethinking Image Processing for the Age of AI: A Problem-First Framework for Scientific Progress
Organizations: University of Nottingham Ningbo China, Ningbo, China · University of Nottingham, Nottingham, United Kingdom
Abstract
Modern AI has greatly expanded the capabilities of image processing. However, the ready availability of powerful models, public datasets, and benchmark leaderboards has also en- couraged a model-first research pattern: researchers increasingly begin with an available architecture and optimize it on a public benchmark, rather than beginning with the underlying real-world imaging problem. This can produce impressive benchmark results without necessarily improving our understanding or solution of the real problem. This paper argues for a problem-first approach that distinguishes the physical imaging problem, solution principle, statistical estimator, and computational implementation, while clarifying what modern AI can achieve and which fundamental problems remain unsolved. Through case studies of super- resolution and low-light enhancement, we show how benchmark datasets may define tasks that differ substantially from the real-world problems they are intended to represent, and why performance improvements must be interpreted within the conditions under which they are obtained. We propose a six-stage workflow that places problem formulation, image acquisition, information-loss analysis, assumptions, ambiguity, and evaluation before model and dataset selection. The paper also proposes clearer standards for evidence, reproducibility, uncertainty, and claims of state-of-the-art performance. More fundamentally, it calls for a change in research culture and education so that future researchers learn to understand imaging problems deeply and use modern AI to achieve genuine scientific and technical advancement.
Figures & tables
| Level | Central question | Typical choices | Typical failure when conflated |
|---|---|---|---|
| Problem | What latent quantity should be inferred from what observation, under which acquisition and use? | Sensor-to-radiance recovery, canonical rendering, perceptual enhancement, recognition-oriented preprocessing | A benchmark construction silently becomes the real-world problem |
| Principle | What knowledge makes the problem solvable? | Bayesian inference, regularization, empirical risk minimization, invariance, self-consistency | A generic learning principle is presented as a problem-specific explanation |
| Estimator | Which decision is optimal for the loss and uncertainty? | Conditional mean, median, MAP estimate, posterior sample, robust decision | Different outputs are ranked as if they estimated the same target |
| Implementation | How is the estimator represented and computed? | Filter, optimization, CNN, transformer, diffusion sampler, hybrid system | Architectural novelty is mistaken for new information or new physics |
| Dimension | Designed/model-based processing | Learning by example | Deep and generative learning |
|---|---|---|---|
| Representation | Usually designed; assumptions visible in filters, transforms, and regularizers | Designed features with learned predictors, dictionaries, exemplars, or priors | Representation and processing learned jointly at high capacity |
| Primary knowledge | Physics and analytic assumptions ( ) | Internal or external examples ( ), often with explicit structure | Large external priors, architecture bias, and increasingly physics/internal adaptation |
| Typical inference | Direct filtering or explicit optimization | Regression, lookup, retrieval, or sparse inference | End-to-end regression, unfolding, adversarial inference, or posterior sampling |
| Main strength | Transparency, data efficiency, predictable constraints | Empirical flexibility with moderate interpretability | Capacity, context, transfer, and complex or multimodal conditionals |
| Characteristic risk | Model bias and restrictive priors | Limited examples or representation capacity | Distribution shift, opaque priors, and convincing unsupported detail |
| What remains invariant | Observation formation, information loss, identifiability, loss-dependent decisions, and the need to justify prior information | ||
| Evidence component | What it establishes | What it does not establish alone |
|---|---|---|
| Matched synthetic benchmark | Controlled accuracy for a known degradation and reference | Robustness to unknown acquisition or real sensor pipelines |
| Captured paired benchmark | Performance for a specific capture and registration protocol | Universal real-world generalization or unique ground truth |
| Cross-dataset/sensor transfer | Sensitivity to dataset and device shift | Coverage of all deployments |
| Forward-model stress test | Behaviour under controlled physical parameter changes | Adequacy of the chosen model family |
| Re-degradation consistency | Compatibility of output with observation under an assumed model | Truth of prior-dominated null-space detail |
| Unpaired in-the-wild testing | Qualitative failure modes and deployment diversity | Full-reference fidelity |
| Community actor | Constructive practice | Cultural signal |
|---|---|---|
| Students and researchers | Begin with target, acquisition, information loss, assumptions, and use; treat architecture as a justified design decision | Understanding the problem is part of technical excellence |
| Supervisors and research leaders | Require a problem statement and falsifiable claim before model development; discuss failures and mismatch, not only average scores | Careful formulation and diagnosis are legitimate research outputs |
| Authors | Match claims to evidence; disclose resources and tuning; report robustness, uncertainty, and limitations | Credibility matters more than rhetorical breadth |
| Reviewers | Evaluate whether the method solves the stated problem and whether comparisons are fair; recognize non-architectural novelty | The highest familiar-benchmark score is not the only form of contribution |
| Editors and programme committees | Broaden scope and reviewer guidance; welcome problem, data, measurement, evaluation, and diagnostic papers | Publication rewards cumulative knowledge, not only leaderboard movement |
| Benchmark organizers | Document construction; refresh tests; include shifts, efficiency, uncertainty, and worst-group results | Benchmarks are instruments for inquiry rather than definitions of reality |
Benchmark Design
Benchmark design creates tasks, datasets, metrics, and protocols that make AI capabilities measurable and comparisons credible.
See all 394 papers