Mapping License Plate Recoverability Under Extreme Viewing Angles for Oppor-tunistic Urban Sensing

Abstract: Urban environments contain many imaging sensors built for specific purposes, including ATM, body-worn, CCTV, and dashboard cameras. Under the opportunistic sensing paradigm, these sensors can be repurposed for secondary inference tasks such as license plate recognition. Yet objects of interest in such imagery are often noisy, low-resolution, and captured from extreme viewpoints. Recent advances in AI-based restoration can recover use-ful information even from severely degraded images. A central challenge is determining which distortion parame-ters allow reliable recovery and which lead to inference failure. This paper introduces recoverability maps, a task-agnostic method for quantifying this boundary. The method combines a dense synthetic sweep of degrada-tion parameters with two summary measures: boundary area-under-curve, which estimates the recoverable frac-tion of the parameter space, and a reliability score, which captures the frequency and severity of failures within that region. We demonstrate the method on license plate recognition from highly angled views under realistic camera artifacts. Several restoration architectures are trained and evaluated, including U-Net, Restormer, Pix2Pix, and SR3 diffusion. The best model recovers about 93% of the parameter space. Similar results across models sug-gest that sensing geometry, rather than architecture, sets the limit of recovery.

• The paper presents a formal methodology using recoverability maps to delineate feasible and unreachable regions for license plate recognition under oblique angles.
• It demonstrates that discriminative models (e.g., Restormer, U-Net) outperform generative counterparts, achieving up to 93.4% recoverability across the angular grid.
• The analysis reveals significant sensitivity to lateral rotation and validates PSNR as a strong proxy for OCR performance in urban sensing deployments.

Quantitative Analysis of License Plate Recoverability under Extreme Angles for Opportunistic Urban Sensing

Introduction

This work establishes a formal methodology for quantifying the operational limits of AI-based license plate recognition (LPR) under extreme, non-purpose-captured viewing conditions typical in opportunistic urban sensing. The problem arises when ubiquitous urban cameras—such as ATMs, CCTVs, and dashcams—incidentally capture vehicle license plates at challenging oblique angles and with sensor artifacts. The core question addressed is: under what geometric and acquisition constraints is reliable automatic LPR feasible, and where is it fundamentally unattainable, regardless of restoration architecture?

Methodological Contributions

The central methodological innovation is the introduction of recoverability maps, a task-agnostic framework for systematic mapping of the inference boundary in high-dimensional degradation parameter spaces. The framework relies on two summary statistics:

• Boundary area-under-curve (AUC): Quantifies the fraction of parameter space (here, the grid of azimuthal and elevational angles) where the target inference—accurate plate recognition—remains feasible.
• Reliability score $F$: Measures the frequency and spatial depth of failures inside the recoverable region, thus capturing not just the nominal extent but the consistency of performance.

These metrics are computed via dense synthetic sampling using a low-discrepancy Sobol sequence over all integer viewing-angle pairs in $[0°, 89°]^2$, propagating each synthetic plate through a realistic degradation pipeline (geometric warp, blur, jitter, and compression), followed by de-warping and restoration via one of several deep learning models. The evaluation is not tied to LPR per se; any secondary inference task can be studied with the same machinery given an appropriate degradation-parameter space and binary success metric.

Experimental Design

The authors implement and train five restoration models annexing both discriminative and generative paradigms:

• U-Net (baseline): Encoder-decoder with skip connections, optimized for $L_1$ + SSIM loss.
• U-Net Conditional: Incorporates angle-conditional FiLM layers for explicit geometric context injection.
• Restormer: Transformer-based, global context modeling for challenging restoration.
• GAN-Pix2Pix: Patch-based adversarial loss to emphasize perceptual plausibility.
• Diffusion-SR3: Denoising diffusion probabilistic model for generative, iterative refinement.

Model selection is based on validation-set PSNR. Models are trained on two synthetic datasets: a Standard set with mild emphasis on extreme angles and an Extreme set with heavy sampling in the most oblique regions. The test-time evaluation grid exhaustively sweeps the $[0°, 89°]^2$ space.

Key Empirical Results

Maximal Recoverability Boundary

The best-performing model, Restormer, recovers approximately 93.4% of the angular parameter grid at $\geq 90\%$ plate OCR accuracy. All tested architectures—regardless of discriminative or generative class—share a tightly clustered boundary, indicating that the physical information loss under severe foreshortening is the primary limiting factor, not the restoration network design. An irreducible region (the $\sim$6.6% in the upper-right, where both azimuth and elevation exceed $\sim$80°) is unrecoverable by any evaluated model due to information collapse.

Asymmetry of Angular Sensitivity

There is a pronounced asymmetry in recoverability: lateral (azimuthal) rotations degrade LPR performance much more rapidly than elevational ones. This results directly from the information geometry of standard plates—horizontal separation encodes critical digit boundaries, so lateral foreshortening is intrinsically more destructive.

Architecture-Dependent Reliability

Discriminative models (Restormer, U-Net, U-Net Conditional) uniformly exhibit higher AUC and lower reliability score $F$ than generative counterparts. Diffusion-SR3, though advanced in other restoration tasks, is unstable on extreme angles, hallucinating visually plausible but incorrect digit sequences with $F$ up to 1.12 on the Extreme dataset. In contrast, discriminative models produce ambiguous or blurred outputs, which are more readily rejected by downstream systems. U-Net Conditional achieves a competitive accuracy-latency tradeoff (7.5 ms inference), making it suitable for real-time urban deployments.

Proxy Metric Validity

PSNR and OCR accuracy are tightly coupled ($R^2 \geq 0.979$), validating the use of PSNR as a training and checkpoint-selection proxy. SSIM, in contrast, displays threshold-like rather than linear behavior and should not be used as a primary surrogate for OCR performance in this context.

Implications

Practical Deployment in Smart Cities

The quantified boundary (up to 93.4% recoverability) provides a direct guideline for urban planners: as long as fixed sensors do not simultaneously exceed ~80° along both angular axes relative to incident plates, software-only restoration is feasible for most incidental captures. Lateral mounting angle (azimuth) is more critical than elevation for hardware placement. Furthermore, the ability to overlay sensor coverage distributions on the recoverability map enables precise ROI estimation and effective forensic triage.

Generality and Framework Transferability

Although demonstrated on LPR, the recoverability-map framework immediately extends to other opportunistic inference tasks—including face recognition under pose and occlusion, scene text under glare/motion blur, and gait recognition from varied viewpoints. The dichotomy between input-induced performance boundaries and model-imposed limits surfaces as a recurring theme for all such secondary-use settings.

Limitations and Future Work

The results are benchmark-bound, relying on synthetic plate generation with fixed font, resolution, and noise characteristics. Unmodeled phenomena in real-world data (e.g., dirt, scale variance, rolling shutter, multi-fonts) will shift absolute error rates, but not the location or structure of the recoverable domain. The OCR backend is Tesseract v4; stronger plate-specific recognizers may extend marginal recoverability but will not resolve the irretrievable-information region at extreme double-foreshortening.

Open directions include extending recoverability mapping to richer acquisition pipelines, alternate inference tasks, and developing benchmarks or controlled datasets that sample real-world deployments at their operational extremes.

Conclusion

This study provides a rigorous quantitative methodology for mapping and interpreting the operational boundaries of deep learning-based LPR under the opportunistic urban sensing paradigm (2604.23814). It establishes that the camera's geometric relationship to the plate fundamentally constrains recoverability; discriminative restorers are superior for reliable deployment, and PSNR can act as an effective proxy for OCR. The general recoverability map technique and its metrics offer a systematic deployment analysis tool for a broad class of repurposed urban inference systems, highlighting the need for both appropriate measurement frameworks and realistic parameter-space coverage in the development and evaluation of smart city sensing algorithms.