PaperlayerView exact revision on arXiv

Source-linked AI summary

Bridge Damage Detection from Low-Light UAV Imagery via Degradation-Aware Mixture-of-Experts Enhancement

Hu Wang, Hongxu Pu, Zhiqi Hu, Fangzhou Lin, Wang Wang

arXiv:2608.23136v1cs.CV

TL;DR

Low-light UAV imagery can obscure bridge damage, while existing datasets lack aligned normal-light–low-light pairs for controlled restoration and recognition evaluation. The paper introduces detector-agnostic DaL-MoE restoration, which improved detection on synthetic and real low-light imagery, raising real-image box mAP50 from 0.123 to 0.252.

  • Problem

    Low-light bridge inspection lacks aligned image pairs and restoration methods tailored to damage recognition across detection and segmentation frameworks.

  • Method

    DaL-MoE combines ISP-aware low-light synthesis with degradation-aware guidance and complementary experts for detector-agnostic restoration.

  • Results

    0.129 absolute gain raised real-image YOLOv11m box mAP50 from 0.123 to 0.252, while synthetic evaluation also improved detection and segmentation metrics.

  • Takeaways & Limitations

    Restored images provided cross-architecture utility under evaluated conditions and improved recognition on real nighttime UAV imagery without enhancement-specific fine-tuning.

  • Takeaways & Limitations

    Evaluation lacks real paired normal-light–low-light UAV imagery, so controlled restoration assessment relies on synthetic pairs.

Abstract

from arXiv · show

Poor illumination obscures small, low-contrast defects in UAV bridge imagery, reducing the reliability and operational flexibility of automated inspection. This paper investigates whether degradation-aware image restoration can improve bridge damage detection under low-light conditions and transfer from synthetic degradations to real inspection scenes. We propose DaL- MoE, a detector-agnostic restoration front end trained with an ISP-aware low-light synthesis pipeline and equipped with degradation-aware guidance estimation and complementary experts for noise suppression, color adjustment, and structural-detail recovery. On paired synthetic data, DaL-MoE achieves 23.12 dB PSNR and 0.8482 SSIM, increasing YOLOv11m box mAP50 from 0.3097 to 0.4923 and mask mAP50 from 0.2281 to 0.3529. On real low-light UAV imagery without paired normal-light references, sim-to-real evaluation shows improved defect visibility and more complete detections than direct inference on raw low-light inputs. Future work will develop low-light-aware bridge damage detectors with stronger cross-scene generalization across bridge sites, imaging conditions, and illumination levels.

1. Introduction

Bridge surface damage threatens durability, service performance, and maintenance decisions, while UAV inspection is challenged by uncontrolled, highly variable illumination and limited paired low-light training data. The paper addresses these challenges with task-oriented restoration and an ISP-aware dataset that preserves aligned damage annotations.

  • Motivation: Surface damage can signal declining structural durability, service performance, and accumulating maintenance risks, making timely and reliable identification essential.Examples include cracking, spalling, corrosion, seepage, and localized material loss.
  • Motivation: UAV-assisted visual inspection supports bridge data acquisition and recognition, including damage quantification, crack recognition, and multiple surface-damage detection.Prior work has also examined UAV inspection path planning and field data acquisition.
  • Challenges: Bridge inspection imagery often suffers shadows, localized underexposure, and substantial brightness variations in difficult-to-control locations and conditions.Vulnerable areas include bridge undersides, inter-girder regions, bearings, expansion joints, box-girder interiors, and occluded surfaces; weather and inspection time also affect exposure.
  • Challenges: Low-light restoration and damage recognition are constrained by scarce normal-light–low-light pairs showing identical damage locations and viewpoints with strict spatial correspondence.This limitation is especially relevant to UAV-based bridge inspection.
  • Contributions: The paper introduces task-oriented restoration and an ISP-aware dataset that generates spatially aligned synthetic low-light images while retaining shared detection and segmentation annotations.The dataset simulates exposure attenuation and sensor noise in a RAW-like domain while preserving damage locations and scene geometry.

2. Related Work

Related work has advanced bridge-damage datasets and low-light restoration, but remains misaligned with UAV inspection because paired multi-class data and recognition-oriented restoration are lacking. This paper addresses the resulting need to recover damage-relevant visual evidence under adverse illumination.

  • Bridge damage datasets: Bridge-damage datasets evolved from binary crack classification toward multi-class detection and segmentation with greater defect diversity and annotation granularity.SDNET2018 supported binary crack classification, whereas CODEBRIM, BridgeDamage, and GYU-DET expanded damage categories and task coverage.
  • Data limitations: Existing bridge inspection datasets generally lack strictly spatially aligned normal-light–low-light pairs with shared, verified detection and instance-segmentation annotations.This limits controlled study of low illumination while keeping damage geometry constant and makes real paired UAV data difficult to acquire.
  • Research gap: Low-light infrastructure datasets remain scarce and crack-centric, leaving unresolved the combination of strict pairing, physics-grounded degradation, and multi-class detection and instance-segmentation annotations.This gap motivates controllable training and evaluation without real paired UAV imagery and restoration that remains detector-agnostic.
  • Low-light restoration: Low-light vision progressed from Retinex-style heuristics to deep restoration and physics-based synthesis, including paired benchmarks, zero-reference enhancement, and ISP-aware degradation modeling.LOL and SID provide paired restoration data; Zero-DCE removes paired-data dependence, while unprocessing simulates exposure and sensor noise in a RAW-like domain.
  • Recognition-oriented restoration: Pixel-level fidelity and perceptual brightness do not ensure preservation of weak textures, subtle boundaries, and low-contrast cues needed by downstream damage-recognition models.Aggressive denoising can erase thin cracks, brightness amplification can exaggerate clutter, and sharpening can create misleading artifacts.

3. Methodology

DaL-MoE is a detector-agnostic, task-oriented restoration front end for low-light bridge damage analysis. Its two-stage design uses degradation-aware guidance estimation and adaptive complementary experts to restore illumination, suppress noise, and preserve structural detail.

  • Design principle: DaL-MoE remains detector-agnostic during training while targeting downstream bridge damage analysis and organizing restoration around recognition-relevant low-light degradations.The model does not require a detection loss or embedded detector during training.
  • Adaptive restoration: AMER adaptively balances complementary restoration behaviors because noise suppression, brightness correction, and structural-detail preservation vary across bridge scenes and illumination conditions.This adaptive mixture-of-experts mechanism is designed to address spatially coupled low-light degradation factors.
  • Overall architecture: The architecture contains Degradation-Aware Guidance Estimation (DAGE) and Adaptive Mixture-of-Experts Restoration (AMER) modules.DAGE estimates degradation guidance, while AMER performs adaptive restoration.
  • Degradation-aware guidance estimation: DAGE predicts a spatially varying illumination map and noise-aware guidance feature, providing a preliminarily compensated image and conditional degradation guidance for restoration.The module derives a noise-related degradation prior directly from the predicted illumination map without additional noise supervision.
  • Detail enhancement expert: The Detail Enhancement Expert combines learnable difference extraction, structural reconstruction, and spatially adaptive gating to strengthen boundary- and texture-sensitive responses through a residual connection.It targets fine, low-contrast cues such as thin cracks, spalling boundaries, exposed rebar edges, corrosion contours, and local texture discontinuities.

4. Low-light Data Synthesis Pipeline

The paper synthesizes low-light observations from normal-light images using an ISP-aware pipeline that applies degradations in a RAW-like sensor representation rather than by directly attenuating sRGB intensities. The process reverses key ISP stages, models exposure and sensor noise, and renders the corrupted signal back to sRGB for training and evaluation.

  • Pipeline overview: The pipeline unprocesses normal-light sRGB images into a RAW-like linear representation, applies sensor-domain low-light corruption, and re-renders the result to sRGB.This avoids direct intensity attenuation in the sRGB domain and produces synthetic low-light counterparts for model training and evaluation.
  • RGB-to-RAW unprocessing: RGB-to-RAW unprocessing reverses tone mapping, gamma correction, color-space conversion, and white-balance application.The method uses a three-channel RAW-like representation, allowing color transformations to operate directly on RGB triples.
  • RAW-domain low-light corruption: The corruption stage combines global exposure attenuation sampled within [0.01, 0.1] with a smooth spatial illumination field rescaled to [0.5, 1.5].The illumination field models non-uniform underexposure caused by structural occlusion and local shadows.
  • RAW-domain low-light corruption: Shot-noise and read-noise levels are sampled in the logarithmic domain, with conditional sampling preserving their empirical correlation while ensuring λ_s > 0 and λ_r > 0.The read-noise model is conditioned on the shot-noise level.
  • RAW-to-RGB rendering: Forward rendering adds quantization noise at simulated bit depths B ∈ {12, 14, 16} before applying reciprocal ISP operations and clipping to the valid range.The reciprocal operations restore white balance, convert camera RGB to RGB, gamma-encode, and tone-map the corrupted signal.

5. Experiments and Geometric Evaluation · 5.1. Experimental platform

The experiments evaluate DaL-MoE for image enhancement and downstream bridge-damage detection and instance segmentation. The platform uses paired synthetic low-light counterparts, spatially separated data splits, and a high-end Linux workstation.

  • 5. Experiments and Geometric Evaluation: The experiments assess DaL-MoE through enhancement quality, bridge-damage detection, and instance-segmentation performance.The evaluation includes quantitative and qualitative comparisons, ablations, and real-world assessments.
  • 5. Experiments and Geometric Evaluation: The experimental platform covers datasets, implementation details, evaluation protocols, quantitative and qualitative comparisons, ablations, and real-world evaluations.
  • 5.1. Experimental platform: All experiments ran on Linux with an AMD Ryzen 9 9950X CPU, 96 GB RAM, and an NVIDIA GeForce RTX 5090 GPU with 32 GB VRAM.
  • 5.1. Experimental platform: The UAV bridge defect dataset and paired low-light benchmark provide dataset-level and class-wise instance-distribution information.The supplied passage identifies the benchmark overview but does not report its distribution values.
  • 5.1. Experimental platform: One spatially aligned synthetic low-light counterpart was generated for every normal-light source image.Image counts and split sizes refer to the normal-light source images.
  • 5.1. Experimental platform: Training and test images were assigned by mutually exclusive bridge spans or mileage sections, excluding adjacent or spatially overlapping UAV views across subsets.
  • 5.1. Experimental platform: Panel B percentages represent column-wise class proportions.

5.2. Datasets

The dataset comprises spatially partitioned real-world UAV bridge inspection imagery with manually delineated defect annotations across eight categories. An ISP-aware synthesis pipeline generates spatially aligned low-light counterparts that retain the original damage geometry and annotations.

  • Real-world UAV dataset: 2,895 images were collected across seven bridges and spatially distinct spans or mileage sections using a DJI H30T-equipped UAV at 4032×3024 pixels.Partitioning by spatially disjoint bridge spans or mileage sections reduces leakage from adjacent viewpoints or repeated local scenes.
  • Defect annotations: Eight defect categories were annotated with closed polygon contours, from which axis-aligned bounding boxes and instance masks were automatically derived.The dataset contains 19,582 annotated defect instances.
  • Low-light synthesis: The ISP-aware synthesis pipeline generated spatially aligned low-light counterparts from normal-light images because paired real-world references are difficult to acquire.Dataset partitioning was completed before synthesis, keeping each source image and its derived counterpart in the same subset.
  • Paired annotations: Synthetic low-light images inherit the corresponding normal-light images’ bounding boxes and instance masks because synthesis preserves damage locations and scene geometry.The paired images retain identical defect locations, instance masks, and axis-aligned bounding boxes.

5.3. Implementation details

DaL-MoE is implemented in PyTorch and trained under a unified enhancement protocol using Adam, cosine learning-rate decay, and offline augmentation of aligned image pairs.

  • DaL-MoE is implemented using PyTorch under a unified training protocol for controlled comparison across image enhancement methods.
  • 3 × 105 iterations are used to optimize each model with Adam.
  • The learning rate decays from 2 × 10−4 to 1 × 10−6 using cosine annealing.
  • Offline augmentation is applied before training to spatially aligned low-light and normal-light image pairs.Augmentations include random rotations and horizontal and vertical flips.

5.4. Evaluation Protocol and Metrics

The evaluation uses complementary protocols to measure restoration fidelity, downstream bridge-damage recognition, and transfer to real inspection scenes. Synthetic quantitative results use 482 held-out paired images, while real nighttime imagery is assessed with reference-free visibility descriptors and fixed downstream inference.

  • Evaluation protocols: Three complementary protocols assess restoration fidelity, downstream bridge damage recognition utility, and transfer to real inspection scenes on a 482-image held-out synthetic test set.The synthetic benchmark provides paired data for quantitative evaluation.
  • Downstream recognition protocol: Downstream models are trained once on normal-light data, held fixed without enhancement-specific fine-tuning, and evaluated under Normal, Base, and enhanced input settings.Base denotes direct inference on low-light images without enhancement; Normal uses corresponding normal-light images and is excluded from ranking.
  • Object detection and instance segmentation metrics: mAP50 averages AP across eight defect categories at IoU 0.50, whereas mAP50:95 averages ten thresholds from 0.50 to 0.95; box and mask variants use bounding-box and mask IoU.Higher values indicate better classification and localization or instance-boundary delineation performance.
  • Real-world evaluation: For real nighttime UAV imagery, missing normal-light references prevent PSNR and SSIM computation, so dataset-averaged mean intensity and Canny edge density describe visibility changes.The same grayscale conversion, Canny settings, and fixed YOLOv11m inference settings are used for original and enhanced images.
  • Real-world evaluation: Mean intensity and edge density are descriptive rather than full-reference fidelity measures because increased brightness or edge density may reflect inaccurate restoration, residual noise, or enhancement artifacts.These descriptors were therefore excluded from fidelity evaluation.

5.5. Comparison with State-of-the-Art Methods

DaL-MoE outperforms the compared low-light enhancement methods on the paired bridge benchmark and consistently improves downstream bridge damage detection and segmentation across detector architectures. Visual comparisons further show more faithful illumination restoration and more complete, detection-friendly predictions under low-light conditions.

  • Quantitative comparison: The comparison evaluates DaL-MoE against eight low-light enhancement methods and tests enhanced images with five bridge damage detector architectures.The detectors are YOLOv8m, YOLOv11m, YOLOv8n, YOLOv11n, and SCNet.
  • Quantitative comparison: 23.1174 dB PSNR and 0.8482 SSIM are achieved by DaL-MoE, the highest image enhancement performance on the paired bridge benchmark.DaL-MoE improves PSNR over Retinexformer by 1.1726 dB; MIRNet is the strongest baseline for SSIM at 0.7903.
  • Downstream recognition: 0.4923 box mAP50 and 0.3578 box mAP50∶95 are obtained by DaL-MoE on YOLOv11m, outperforming Retinexformer by 0.0428 and 0.0679, respectively.DaL-MoE also achieves the highest values across the reported detector–task combinations for bounding-box detection and instance segmentation.
  • Downstream recognition: 0.2019, 0.1826, 0.09, 0.1299, and 0.1127 box mAP50 gains are reported over Base for YOLOv8m, YOLOv11m, YOLOv8n, YOLOv11n, and SCNet, respectively.The same overall trend is observed for box mAP50∶95, mask mAP50, and mask mAP50∶95, while DaL-MoE narrows the gap to the Normal upper-bound reference.
  • Visual comparison: DaL-MoE restores illumination more faithfully toward normal-light references, making bridge surfaces and damage-relevant cues more visible across challenging scenes.With YOLOv11m, enhancement produces more defect regions and more complete predictions than direct inference on low-light images, while Retinexformer can still yield missed detections.

5.6. Ablation Study

The ablation study evaluates DaL-MoE’s DAGE and AMER modules on paired synthetic low-light bridge imagery, including AMER’s expert branches and adaptive gating. Removing these components reduces restoration quality, supporting their role in recovering information useful for damage detection and segmentation.

  • Ablation scope: The study ablates DAGE, AMER, AMER’s three expert branches, and its adaptive gating mechanism on the paired synthetic low-light bridge benchmark.DAGE and AMER denote the Degradation-Aware Guidance Estimation and Adaptive Mixture-of-Experts Restoration modules; the experts target noise, color, and structural detail.
  • DAGE and AMER: 21.65 PSNR and 0.7157 SSIM result when AMER is removed, versus 23.12 PSNR and 0.8482 SSIM for the full model.The degradation demonstrates the importance of adaptive expert collaboration under coupled illumination, noise, contrast, and damage-boundary degradations.
  • DAGE and AMER: 22.73 PSNR and 0.7716 SSIM result when DAGE is removed, as its absence reduces explicit illumination- and noise-related guidance for expert branches.DAGE estimates spatially varying illumination and noise-related guidance from the low-light input.
  • AMER components: 21.14 PSNR and 0.7397 SSIM result when adaptive gating is removed, with the lowest PSNR among evaluated variants.The result indicates that combining experts without degradation-dependent routing is insufficient; gating dynamically adjusts their contributions.
  • AMER components: Removing any individual expert or the gating mechanism lowers performance, confirming complementary expert capabilities and the necessity of adaptive integration.Together with DAGE, these components support restoration that recovers information useful for subsequent multi-class damage detection and instance segmentation.

6. Real-World Evaluation

On 200 real-world nighttime UAV bridge images without paired normal-light references, DaL-MoE improves image-level visibility and supports stronger downstream damage detection with a fixed detector. Real-scene evaluation therefore combines auxiliary visibility descriptors, qualitative comparisons, and task-oriented detection results.

  • Evaluation setup: The real-world evaluation uses 200 nighttime UAV bridge images without spatially aligned normal-light references, so PSNR and SSIM cannot be computed.Evaluation instead examines intensity, edge density, qualitative visibility, and downstream detection.
  • Image-level visibility: Mean intensity increases from 40.2 to 135.0, corresponding to a 3.36× increase without obvious global saturation.The enhanced distribution shifts toward a broader brightness range, improving visibility in underexposed bridge regions.
  • Image-level visibility: Mean Canny edge density increases from 1.55% to 4.07%, indicating richer local edge responses after enhancement.Because edge density may also reflect enhanced texture or residual noise, it is interpreted alongside qualitative examples and detection results.
  • Downstream detection: With detector architecture, parameters, preprocessing, and inference fixed, DaL-MoE makes several previously weak or missed damage regions detectable and better aligned with visible structures.Direct inference on low-light inputs shows missed detections, spurious responses, and inaccurate localization under severe underexposure, darkness, and sensor noise.
  • Downstream detection: 0.129 absolute gain raises overall box mAP50 from 0.123 to 0.252, while box mAP50∶95 increases from 0.061 to 0.126.DaL-MoE improves box detection across all eight damage categories without enhancement-specific fine-tuning.

7. Conclusions

The paper presents DaL-MoE as a degradation-aware restoration front end for low-light UAV bridge damage recognition, evaluated on paired synthetic and real nighttime imagery. Results show improved restoration and downstream recognition, while real-scene scope and cross-site generalization remain limited.

  • Contributions: DaL-MoE addresses low-light bridge damage recognition through controllable synthesis and restoration usable by object detection and instance segmentation models.The framework targets training and evaluation without paired real normal-light–low-light images.
  • Synthetic evaluation: 23.1174 dB PSNR and 0.8482 SSIM were achieved on 482 held-out paired synthetic test images, outperforming compared low-light restoration methods.For YOLOv11m, box mAP50 increased from 0.3097 to 0.4923, and mask mAP50 increased from 0.2281 to 0.3529.
  • Real-world evaluation: 40.2 to 135.0 grayscale intensity and 1.55% to 4.07% Canny edge density were observed across 200 real nighttime UAV images after enhancement.The real images lacked paired normal-light references and were manually annotated with bounding boxes for eight damage categories.
  • Limitations: The real-world quantitative evaluation is limited to 200 images, bounding-box annotations, and one fixed YOLOv11m detector.The synthesis pipeline approximates several imaging degradations but cannot fully reproduce complex field conditions.
  • Limitations: Real-scene instance segmentation performance and robustness across multiple detectors and independent inspection sites remain unestablished.DaL-MoE also retains a performance gap between enhanced and normal-light inputs and lacks explicit detection or segmentation supervision.
  • Future work: Future work will expand the real-world dataset into a larger multi-site low-light bridge benchmark spanning bridge types, structural components, camera systems, UAV platforms, and illumination conditions.Bounding-box annotations will be retained and extended, alongside high-quality instance-mask annotations.

CRediT authorship contribution statement

The authors contributed across drafting, visualization, validation, coding, data curation, figure preparation, conceptualization, experimental design, supervision, and funding acquisition.

  • Hu Wang contributed original drafting, visualization, validation, coding, and funding acquisition.
  • Hongxu Pu contributed validation, coding, and data curation.
  • Zhiqi Hu contributed data curation, coding, and visualization, while Fangzhou Lin contributed data curation and figure preparation.
  • Wang Wang contributed conceptualization, original drafting, experimental design, data curation, coding, and supervision.
Loading 2608.23136v1…