💻 computer science

A Structural Preservation Framework for Denoiser Selection in YOLO-Based Pedestrian Detection under Sensor Noise

This paper introduces the Structural Preservation Score (SPS) to evaluate denoiser effectiveness for YOLO-based pedestrian detection, revealing that while specific variants like gradient-magnitude correlation can rank denoisers within known noise levels, no single image-level metric can universally predict detection performance across unseen denoisers or noise conditions due to non-linear, architecture-dependent factors.

Original authors: Vo Thanh Kiet, Rene Jaros, Minh Ly Duc, Petr Bilik, Radek Martinek

Published 2026-09-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vo Thanh Kiet, Rene Jaros, Minh Ly Duc, Petr Bilik, Radek Martinek

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of modern technology, cameras are the eyes of machines. From self-driving cars navigating city streets to security systems watching over public squares, these devices rely on artificial intelligence to recognize people and objects instantly. For years, engineers have trained these systems to be incredibly sharp, teaching them to spot details in perfect, clear images. However, the real world is rarely perfect. Cameras often struggle when the light is dim, the weather is bad, or the sensor itself is imperfect, resulting in images that look grainy or speckled with static. This noise can confuse the machine, causing it to miss a pedestrian or mistake a mannequin for a real person. To fix this, a common engineering instinct is to add a cleaning step before the machine looks at the image, hoping that a clearer picture will lead to a smarter decision.

But a new study suggests that this instinct is often wrong. Researchers have discovered that simply making an image look cleaner to the human eye does not necessarily help the machine see better. In fact, some methods that produce the most beautiful, smooth pictures can actually blind the detector to the very details it needs to function. The team set out to find a way to predict which cleaning method would actually help a machine see a person, rather than just making the photo look nice. They tested various cleaning techniques on a dataset of pedestrian images, ranging from clear photos to those heavily distorted by noise, and measured how well different versions of a popular detection system performed afterward.

The researchers found that the relationship between image quality and machine vision is far more complex than previously thought. They developed a new way of measuring an image that focuses not on how smooth it looks, but on how well it keeps the sharp edges and textures that a machine uses to identify shapes. When they applied this new measurement to their tests, they found that some traditional cleaning methods, which had been largely overshadowed by newer, more complex artificial intelligence tools, were actually better at preserving these critical details. One older method, which works by comparing small blocks of the image to find patterns, kept the essential structure of the scene intact. In contrast, several modern deep-learning tools, which are trained to minimize pixel-by-pixel errors, tended to smooth out the image too much. These tools removed the noise, but in doing so, they also erased the fine lines and textures that the machine needed to locate a person, causing its accuracy to drop significantly.

The study revealed a surprising truth about how these systems learn. Modern detection tools are already trained to handle a certain amount of messiness. When the researchers retrained the detectors on the noisy images, the machines learned to cope with the static on their own. In this scenario, adding a cleaning step often provided no benefit, and sometimes made things worse if the cleaner removed too much detail. However, in a more realistic scenario where the machine is already trained and cannot be retrained, the cleaning step became vital. Here, the choice of cleaner mattered immensely. The researchers discovered that the effectiveness of a cleaner depends heavily on the level of noise. At low levels of noise, a specific measure of how well an image keeps its edges could strongly predict which cleaner would work best. But if you mixed all the noise levels together, the prediction failed completely, because the severity of the noise itself was the dominant factor.

Perhaps the most significant finding was that there is no single, universal rule for choosing a cleaner. The researchers tried to build a model that could predict the performance of a cleaning method they had never seen before, based on how it performed on the ones they had tested. This attempt failed. The relationship between the structural quality of an image and the machine's success was specific to the type of cleaner being used. A method that worked well for one type of algorithm did not necessarily predict success for a different type. This means that while there is no magic formula to pick the perfect cleaner for any situation, there is a clear path forward for engineers. They can use this new structural measurement to rank a shortlist of known cleaning tools for a specific camera and noise level, but they cannot rely on it to guess the performance of a completely new tool.

The work underscores a fundamental shift in how we think about image processing for machines. The goal is not to create an image that looks perfect to a human, but to preserve the specific structural cues that a machine relies on. The study shows that the best cleaner is not always the one that produces the highest score on standard quality charts, nor is it always the most advanced artificial intelligence model. Sometimes, a simpler, older method that respects the natural edges of the scene is the most effective choice. By understanding that machines see the world through the lens of structure and texture rather than smoothness, engineers can make better decisions about how to prepare images for safety-critical tasks, ensuring that the eyes of the machine remain sharp even when the world around them is noisy.

Technical Summary: A Structural Preservation Framework for Denoiser Selection in YOLO-Based Pedestrian Detection under Sensor Noise

1. Problem Statement

Modern deep object detectors, particularly the YOLO family, achieve high accuracy on curated benchmarks but suffer significant reliability degradation when input images deviate from clean training conditions due to sensor noise (e.g., CMOS/CCD limitations, low-light acquisition). A common engineering response is to insert a denoising stage upstream of the detector based on the assumption that "better image quality yields better detection." However, this paper demonstrates that this assumption fails systematically for YOLO detectors.

Current YOLO variants incorporate aggressive data augmentation (mosaic, noise injection, photometric jitter), internalizing tolerance to mild degradation. Consequently, the marginal value of a denoising stage depends not on pixel-level noise elimination, but on whether the denoiser preserves or destroys the specific structural cues (edges, gradients, high-frequency texture) the detector relies on. The literature lacks a framework to predict which denoising methods will aid downstream detection or to diagnose why others fail, forcing practitioners to exhaustively retrain and evaluate every candidate denoiser-detector pair—a computationally prohibitive task (180 runs in this study). Furthermore, standard image-quality metrics like PSNR and SSIM correlate poorly with downstream task performance as they are tuned to human perception rather than machine vision cues.

2. Methodology

The authors introduce the Structural Preservation Score (SPS), a reference-based family of three computable variants designed to quantify how faithfully a denoiser preserves detector-relevant structure. The framework operates without a trained detector and is evaluated on the PnPLO/Karthy pedestrian dataset (944 train, 160 val, 235 test images) across a factorial experimental design:

Detectors: Five YOLO medium variants (v8m through v12m).
Noise Levels: Six Gaussian noise levels ( $\sigma \in \{0, 1, 5, 10, 20, 30\}$ ).
Processing Conditions: Five denoising methods (Gaussian, BM3D, DnCNN, Autoencoder, CAE+PSO) plus a noisy baseline.
Total Runs: 180 end-to-end training and evaluation runs.

The Three SPS Variants:

Variant A (Edge Overlap): Measures the overlap between Canny edge maps of the clean and denoised images, capturing the survival of object contours.
Variant B (Gradient-Magnitude Correlation): Computes the Pearson correlation of Sobel gradient magnitudes between clean and denoised images. This captures the preservation of local contrast patterns, even if edges shift sub-pixel.
Variant C (High-Frequency Energy Retention): Measures the retention of spectral energy in the high-frequency region (radial cutoff $r_c = 0.5 r_{max}$ ) via 2D FFT. This quantifies whether fine-scale texture is suppressed (over-smoothing) or spurious high-frequency content is injected.

Evaluation Protocols:

Retraining Protocol: Detectors are retrained for every $(method, noise)$ condition to measure the detection performance shift ( $\Delta F1$ ) relative to the noisy baseline.
Frozen-Detector Protocol: Detectors are trained once on clean data and kept fixed; denoisers are applied only at inference time to simulate deployment scenarios.
Cross-Validation: A leave-one-method-out (LOOCV) protocol tests if the SPS-to- $\Delta F1$ relationship generalizes to unseen denoising methods.

3. Key Contributions

SPS Framework: Formulation of a reference-based metric family with explicitly separated predictive (Variants A and B) and diagnostic (Variant C) roles, moving beyond standard PSNR/SSIM leaderboards.
Empirical Evidence: Presentation of results from 180 systematic runs. Under retraining, Variant B (Gradient-Magnitude Correlation) is the only metric among PSNR, SSIM, and the three SPS variants to reach unadjusted statistical significance against $\Delta F1$ ( $r = +0.418, p = 0.038$ ). Its effect size is comparable to PSNR and significantly exceeds SSIM. Among structure-aware perceptual metrics, only GMSD matches Variant B's performance.
Diagnostic Mechanism: High-frequency energy analysis (Variant C) reveals that MSE-trained autoencoders retain only 16–21% of clean-image high-frequency energy at $\sigma=20$ , compared to 87% for BM3D. This quantifies the "over-smoothing collapse" of MSE-trained models. Conversely, DnCNN retains 277% of high-frequency energy, indicating spurious injection (ringing) rather than over-smoothing, which the detector tolerates better than structural loss.
Generalization Limits: Leave-one-method-out cross-validation yields negative $R^2$ values (mean $R^2 = -1.443$ ), demonstrating that the linear SPS-to- $\Delta F1$ relationship does not generalize across different denoising method families. Structural preservation is a necessary but not sufficient condition for predicting performance across unseen operators.

4. Key Results

Correlation with Detection Shift: In the active denoising regime ( $\sigma \ge 1$ ), Variant B correlates with $\Delta F1$ ( $r = +0.418$ ). No metric survives multiplicity correction (Holm/Benjamini-Hochberg), but the effect size is robust. The correlation is detector-dependent, reaching significance only for YOLOv9m and YOLOv10m, likely due to their specific gradient-preserving architectural features (PGI pathway, NMS-free heads).
Frozen-Detector Dynamics: When detectors are frozen (deployment scenario), pooling across noise levels masks the relationship entirely ( $r \approx 0$ ). However, within individual noise levels, Variant B ranks denoisers strongly ( $r \ge +0.93$ at $\sigma \le 10$ ). This identifies the noise level as a confounder; metric-based selection must control for the operating noise level.
Method Performance:
- BM3D: Consistently preserves structure (87% high-frequency retention) and provides the most stable, slightly positive or neutral $\Delta F1$ in high-noise regimes.
- MSE Autoencoders: Actively harmful in retraining regimes, causing absolute collapses in $F1$ (e.g., $\Delta F1 \approx -0.075$ at $\sigma=20$ ) due to over-smoothing.
- DnCNN: Produces spurious high-frequency residue but remains close to the noisy baseline, suggesting detectors tolerate residual noise better than structural loss.
ANOVA Validation: At $\sigma=20$ , the choice of denoising method is a statistically significant factor ( $p=0.0033$ ), with pairwise contrasts showing structure-preserving methods (BM3D, DnCNN) outperforming over-smoothing methods (Autoencoder), though individual pairs do not survive family-wise correction.

5. Significance and Claims

The paper claims that structural preservation is a necessary but not sufficient condition for a denoiser to improve object detection performance. The authors explicitly state that SPS should not be viewed as a universal, detector-agnostic predictor for arbitrary denoising operators. Instead, its significance lies in its role as a computationally cheap, statistically principled aid for ranking candidate denoisers within a curated, calibrated shortlist at a known noise level.

The study refutes the assumption that "better image quality" (per PSNR/SSIM) implies better detection. It establishes that:

Gradient-magnitude preservation is a stronger correlate of detection performance than traditional metrics.
MSE-trained direct-prediction denoisers systematically harm detection by suppressing high-frequency structural cues.
The relationship between image quality metrics and detection performance is non-linear and architecture-dependent, preventing simple extrapolation to unseen methods.

The authors conclude that while SPS cannot replace end-to-end evaluation, it can reduce the search space for denoiser selection by an order of magnitude, provided it is used as a "second opinion" within a controlled candidate set rather than as a standalone universal score.