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 bridge inspection lacks paired normal-/low-light images and evidence on recovering damage-relevant information. The paper combines ISP-aware low-light synthesis with detector-agnostic DaL-MoE restoration, reporting higher YOLOv11m box mAP50 on real nighttime imagery.

  • Problem

    The paper addresses the lack of paired normal-/low-light bridge images and evidence on recovering damage-relevant information for detection and segmentation.

    Source: §2.3 · p. 6

  • Method

    DaL-MoE combines ISP-aware low-light synthesis with degradation-guided restoration and adaptive complementary experts in a detector-agnostic front end.

    Source: §3 · p. 6, §5.6 · p. 23, §7 · p. 28

  • Results

    0.129 absolute gain in YOLOv11m box mAP50 (0.123 to 0.252) on 200 real nighttime UAV images, alongside higher recognition accuracy across architectures on paired synthetic data.

    Source: §7 · p. 28

  • Takeaways & Limitations

    Under the evaluated conditions, restored images provide cross-architecture utility for bridge damage recognition.

    Source: §7 · p. 28

    Canny edge density is an auxiliary visibility descriptor, not a direct measure of structural fidelity.

    Source: §6.1.2 · p. 25

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

UAV bridge inspection supports automated damage recognition, but site-dependent low-light degradation and scarce paired training data limit damage visibility and reliable detection. The paper addresses these challenges with ISP-aware synthetic pairs and DaL-MoE, a detector-agnostic restoration front end evaluated on synthetic and real imagery.

1. Introduction

  • Bridge surface damage can signal declining durability and service performance, making timely identification essential for condition assessment and maintenance.UAV inspection supports a workflow from image acquisition to automated recognition and condition-information extraction.Source: §1 · pp. 1–2
  • Low-light conditions cause underexposure, contrast loss, amplified dark-region noise, and blurred damage boundaries, impairing damage recognition, localization, and delineation.Shadows and brightness variation are especially difficult to control beneath bridges, near occlusions, and across differing weather, inspection times, and viewpoints.Source: §1 · p. 2
  • Strictly paired field images and broadly annotated low-light detector datasets are difficult and costly to obtain, motivating controllable training samples without real paired data.UAV pose, distance, shadows, weather, and camera settings hinder spatially aligned capture; severe underexposure also makes annotation harder.Source: §1 · p. 2
  • Experiments report consistent improvements on the synthetic benchmark and evidence of sim-to-real transfer to practical UAV inspection scenes.Evaluation covers full-reference restoration on paired synthetic data, downstream detection and instance segmentation, and real imagery without paired normal-light references.Source: §1 · p. 3
  • An ISP-aware synthesis pipeline creates spatially aligned low-light samples from normal-light UAV images while retaining shared detection and instance-segmentation annotations.Exposure attenuation and sensor noise are simulated in a RAW-like domain while preserving damage locations and scene geometry.Source: §1 · pp. 2–3
  • DaL-MoE combines degradation-aware guidance, multi-scale restoration, and adaptive experts for noise suppression, color adjustment, and structural-detail recovery as a plug-and-play, detector-agnostic front end.It is trained independently of downstream models and requires neither changes to nor joint optimization with detection or segmentation frameworks.Source: §1 · pp. 2–3

2. Related Work

Prior work advances bridge damage datasets and low-light restoration, but remains limited by missing paired-light bridge data, domain mismatch, and insufficient validation of detector-agnostic, multi-class recognition. This paper positions restoration around recovering damage-relevant visual evidence under compound illumination degradation.

2.1. Datasets for Bridge Damage and Low-Light Vision

  • Bridge datasets evolved from image-level crack classification and multi-label CODEBRIM to multi-class detection and segmentation, but lack aligned paired-light annotations.BridgeDamage provides over 2,800 images across five defect categories, while GYU-DET provides 11,123 images with boxes across six categories; existing datasets lack verified annotations shared across paired lighting states.Source: §2.1 · pp. 3–4
  • Generic low-light benchmarks [15] and synthesis pipelines do not directly provide bridge-specific semantics or preserve defect annotations for controlled multi-class evaluation.Civil-infrastructure low-light datasets remain scarce and crack-centric, leaving strict spatial pairing, physics-grounded degradation, and multi-class detection and segmentation annotations unresolved.Source: §2.1 · pp. 4–5

2.2. Low-Light Image Enhancement and Task-Oriented Restoration

  • Deep low-light methods model illumination, noise, color, and detail, yet pixel-level fidelity or brighter appearance does not guarantee improved defect localization or boundary delineation.For bridge inspection, aggressive denoising can erase thin cracks, brightness amplification can exaggerate clutter, and sharpening can introduce misleading artifacts.Source: §2.2 · p. 5
  • Infrastructure-oriented methods remain crack-centric or backbone-coupled, leaving benefits for heterogeneous multi-class defects across detection and segmentation frameworks insufficiently verified.Spalling, seepage, corrosion, and efflorescence impose different restoration demands from regular crack patterns.Source: §2.2 · pp. 5–6

2.3. Research Gap and Positioning of This Paper

  • MoE offers specialized, gated restoration pathways, but bridge-specific experts guided by degradation cues and used as detector-agnostic front ends remain largely unexplored.The paper frames the unresolved task as recovering visual evidence for multi-class localization and contour delineation across recognition frameworks, including when real paired UAV imagery is unavailable.Source: §2.2 · p. 6, §2.3 · p. 6

3. Methodology

DaL-MoE is a detector-agnostic restoration front end for low-light bridge imagery that aims to preserve damage-relevant structure while addressing coupled degradations. It estimates degradation-aware guidance and uses adaptive gating to weight and fuse complementary expert representations into a restored feature conditioned on the input degradation.

3. Methodology

  • DaL-MoE combines degradation-aware guidance estimation with a U-shaped restoration backbone that adaptively balances noise suppression, brightness correction, and structural-detail recovery.The enhancement network is trained without a detection loss or embedded detector, while targeting downstream bridge damage analysis.Source: §3 · pp. 6–7

3.1. Degradation-Aware Guidance Estimation Module

  • DAGE predicts a spatially varying illumination map for residual compensation and derives a noise-related prior from that map to generate conditional guidance without extra noise supervision.The compensated image and guidance feature are passed to the restoration module as complementary inputs.Source: §3.1 · p. 8

3.2. Adaptive Mixture-of-Experts Restoration Module

  • AMER processes features with parallel noise-aware, color-adjustment, and detail-enhancement experts, then adaptively integrates their outputs rather than imposing a fixed restoration order.Noise-aware guidance is incorporated into the noise-suppression expert, while the other branches address chromatic distortions and structural detail.Source: §3.2 · p. 9

3.2.1. Noise-Aware Expert

  • The Noise-Aware Expert injects degradation guidance into channel-wise attention aggregation to suppress noise-sensitive responses while preserving structural information.Its transposed attention models global inter-channel dependencies without constructing a spatial token-to-token affinity matrix, and a local positional branch supplies spatial cues.Source: §3.2.1 · p. 10

3.2.2. Color Adjustment Expert

  • The Color Adjustment Expert combines global channel-wise affine modulation with local chromatic refinement to correct illumination-related color shifts and spatially varying distortions.A residual connection preserves the underlying structural representation while the expert adjusts color responses.Source: §3.2.2 · p. 11

3.2.3. Detail Enhancement Expert

  • The Detail Enhancement Expert combines learnable local-difference extraction, structural reconstruction, and spatial gating to strengthen boundary- and texture-sensitive responses while retaining the underlying features.Its target cues include thin cracks, spalling boundaries, exposed rebar edges, corrosion contours, and local texture discontinuities.Source: §3.2.3 · pp. 11–12

3.2.4. Adaptive Gating Network

  • The Adaptive Gating Network estimates sample-adaptive routing weights because the relative contributions of noise suppression, illumination correction, and detail recovery vary across inputs and feature scales.It uses the noise-aware expert representation to construct a routing descriptor, produces three expert-specific logits, and normalizes them with Softmax.Source: §3.2.4 · p. 12
  • Scalar weights for the noise, color, and detail experts are broadcast across feature dimensions and used to fuse their representations into a degradation-conditioned restored feature for subsequent propagation.The experts address noise suppression, illumination and chromatic adjustment, and structural-detail recovery.Source: §3.2.4 · p. 13

4. Low-light Data Synthesis Pipeline

Because paired normal- and low-light images from identical real-world viewpoints are difficult to acquire, the pipeline synthesizes low-light images by degrading normal-light images in a RAW-like sensor domain and rendering them back to sRGB. This ISP-aware process is intended to better approximate physical image formation under low illumination.

Low-light Data Synthesis Pipeline

  • The pipeline unprocesses normal-light sRGB images, applies sensor-domain low-light degradations, and renders the corrupted signal back to sRGB for training and evaluation.The resulting image serves as the synthetic low-light counterpart of the normal-light input.Source: §4 · pp. 13–14

RGB-to-RAW unprocessing

  • Normal-light sRGB is mapped to a three-channel RAW-like linear representation by approximately reversing four ISP stages, following.The camera-RGB conversion uses a randomly selected camera color correction matrix, and white balance is removed to form the RAW-like signal.Source: §4 · p. 13

RAW-domain low-light corruption

  • RAW corruption combines exposure attenuation, spatially varying illumination, and signal-dependent shot noise with signal-independent read noise.The global exposure ratio is sampled from a truncated normal distribution within [0.01, 0.1], while shot- and read-noise levels are conditionally sampled to preserve their empirical correlation and remain positive.Source: §4 · p. 14

RAW-to-RGB rendering

  • The corrupted RAW-like signal receives quantization noise and is rendered through reciprocal ISP operations—white-balance restoration, color conversion, gamma encoding, and tone mapping—before clipping to the valid range.The simulated bit depth is 12, 14, or 16.Source: §4 · p. 14

5. Experiments and Geometric Evaluation

The section evaluates DaL-MoE’s image-enhancement quality and its performance on downstream bridge-damage detection and instance segmentation.

5. Experiments and Geometric Evaluation

  • The experiments cover the platform, datasets, implementation, and evaluation protocols, followed by quantitative and qualitative comparisons, ablations, and real-world evaluation.The evaluated outcomes include image-enhancement quality and downstream bridge-damage detection and instance-segmentation performance.Source: §5 · p. 14

5.1. Experimental platform

The experimental platform combines a UAV bridge-defect dataset and paired low-light benchmark with a high-specification workstation. Synthetic counterparts and spatially separated splits define the evaluation data setup.

5.1. Experimental platform

  • 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.Source: §5.1 · p. 14
  • The dataset table covers UAV bridge-defect class distributions and the paired low-light benchmark.It provides an overview; the supplied caption does not report instance counts or class-wise values.Source: Table 2 · p. 15
  • Each normal-light source image had one spatially aligned synthetic low-light counterpart, with splits separated by bridge spans or mileage sections and no adjacent or overlapping UAV views shared across subsets.The split-size counts refer to normal-light source images; Panel B percentages are column-wise class proportions.Source: §5.1 · p. 15

5.2. Datasets

The dataset combines spatially partitioned real UAV bridge imagery with aligned synthetic low-light counterparts and shared defect annotations. It supports image-enhancement training and controlled detection and segmentation evaluation under low-light conditions.

5.2. Datasets

  • 2,895 UAV images from seven bridges were split by spatially disjoint bridge spans or mileage sections into 2,413 training and 482 test images.Images were acquired with a DJI H30T at 4032×3024 pixels; adjacent or overlapping views from the same section remained in the same subset.Source: §5.2 · p. 15
  • 19,582 annotated defect instances cover eight categories, with 15,731 instances in training and 3,851 in test.The eight categories are spalling, crack, hole, honeycombing, seepage, rebar corrosion, rust, and efflorescence; manual polygon contours generated aligned bounding-box and instance-mask annotations.Source: §5.2 · p. 15
  • An ISP-aware pipeline generated spatially aligned low-light counterparts after dataset partitioning, preserving each source image’s subset and shared damage annotations.Normal-light/low-light pairs were used to train and full-reference evaluate enhancement; synthetic low-light test images supported controlled detection and instance-segmentation evaluation.Source: §5.2 · pp. 15–16

5.3. Implementation details

DaL-MoE and the compared enhancement methods were trained under a unified protocol, with paired augmentations preserving spatial correspondence. Downstream detectors used 640 × 640 inputs.

5.3. Implementation details

  • DaL-MoE was implemented in PyTorch and trained with Adam for 3 × 10^5 iterations, cosine-decayed learning rates from 2 × 10^-4 to 1 × 10^-6, and mean absolute error loss.Training used 256 × 256 crops and a batch size of 8.Source: §5.3 · p. 16
  • Synchronous random rotations and horizontal and vertical flips expanded 2,413 aligned training pairs to 12,065 augmented pairs while preserving correspondence.The augmented pairs were randomly cropped into 256 × 256 patches for training.Source: §5.3 · p. 16
  • Downstream detectors were trained and evaluated at 640 × 640 input resolution.The paper states that Table 3 summarizes DaL-MoE configuration, compared enhancement-method settings, and detector architectures.Source: §5.3 · p. 16

5.4. Evaluation Protocol and Metrics

The evaluation uses three complementary protocols to assess restoration fidelity on a held-out synthetic test set, downstream recognition under controlled input conditions, and transfer to real nighttime UAV imagery. Real-scene visibility measures are treated as descriptive evidence, not full-reference fidelity scores or primary ranking criteria.

Evaluation Protocol and Metrics

  • Quantitative evaluations on the paired synthetic benchmark use a held-out test set of 482 images, with PSNR and SSIM reported as arithmetic means over test image pairs.The benchmark assesses restoration fidelity.Source: §5.4 · pp. 16, 18
  • Downstream models are trained once on normal-light data and held fixed across Normal, unenhanced Base, and Enhanced inputs, isolating the restoration front end’s influence.The same test images, annotations, checkpoints, and inference settings are used; Normal is excluded from enhancement-method ranking.Source: §5.4 · p. 18
  • Detection and segmentation are scored with box and mask IoU, reporting mAP50 and mAP50:95 across eight defect categories and ten IoU thresholds.mAP50:95 averages AP from IoU 0.50 to 0.95 in 0.05 increments.Source: §5.4 · p. 18
  • For 200 real nighttime UAV images without paired normal-light references, dataset-averaged intensity and Canny edge density describe visibility changes but cannot establish restoration fidelity.The measures are excluded from primary ranking because brightness gains may be inaccurate and edge gains may reflect noise or artifacts; real-scene results provide complementary transfer evidence.Source: §5.4 · pp. 18–19

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

DaL-MoE is compared with eight enhancement methods and evaluated through five fixed detector architectures. It leads the paired-benchmark enhancement metrics and improves downstream recognition, with visual comparisons showing more faithful restoration and more complete detections.

5.5.1. Quantitative Comparison

  • 23.1174 dB PSNR and 0.8482 SSIM make DaL-MoE best on the paired benchmark; on YOLOv11m it records box mAP50 0.4923 and mask mAP50 0.3529, highest among compared enhancements.The enhancement comparison reports gains of 1.1726 dB PSNR over Retinexformer and 0.0579 SSIM over MIRNet; the YOLOv11m results also exceed Retinexformer in box and mask metrics.Source: §5.5.1 · p. 20
  • Against Base, DaL-MoE raises box mAP50 by 0.2019, 0.1826, 0.09, 0.1299, and 0.1127 for YOLOv8m, YOLOv11m, YOLOv8n, YOLOv11n, and SCNet, respectively, while narrowing the gap to Normal.The gains extend across box and mask metrics, supporting DaL-MoE as a detector-agnostic restoration front end; Normal remains the upper-bound reference under evaluated conditions.Source: §5.5.1 · pp. 20–21

5.5.2. Visual Comparison

  • Across bridge scenes, DaL-MoE restores illumination with stable color, local contrast, and structural detail, while YOLOv11m produces more complete detections than on raw low-light inputs or Retinexformer outputs.The visual comparisons show fewer missed detections after enhancement and preservation of damage-relevant cues such as crack boundaries, spalling contours, corrosion patterns, and local texture variations.Source: §5.5.2 · pp. 21–22, Figure 4 · p. 21, Figure 5 · p. 22

5.6. Ablation Study

Ablations on paired synthetic low-light bridge data show that DAGE, AMER’s three experts, and degradation-dependent gating each contribute to restoration quality and preservation of damage-relevant structure.

Ablation Study

  • On a paired synthetic benchmark preserving image correspondence and damage annotations, the ablations support complementary experts and gating for restoring damage-relevant information for detection and instance segmentation.The evaluation reflects recovery of illumination, suppression of degradation, and preservation of structural information useful for subsequent multi-class damage analysis.Source: §5.6 · pp. 22–24

Ablation of DAGE and AMER

  • Removing AMER drops PSNR from 23.12 to 21.65 and SSIM from 0.8482 to 0.7157; removing DAGE yields PSNR 22.73 and SSIM 0.7716.DAGE supplies illumination- and noise-related guidance, while AMER adaptively integrates restoration behaviors.Source: §5.6 · pp. 22–23, Table 6 · p. 23

Ablation of the AMER components

  • Removing the Noise-Aware Expert yields PSNR 22.03 and SSIM 0.7682; Color Adjustment yields PSNR 22.89 and SSIM 0.7449; Detail Enhancement yields PSNR 22.72 and SSIM 0.7553.The branches target noise suppression, color and illumination adjustment, and local-boundary and fine-structure recovery, respectively.Source: §5.6 · p. 23
  • Removing adaptive gating gives PSNR 21.14 and SSIM 0.7397, with the lowest PSNR among variants; combining experts without degradation-dependent routing is insufficient.The gate adjusts each expert’s contribution according to the input’s degradation characteristics and feature scale.Source: §5.6 · p. 23

6. Real-World Evaluation

On 200 real nighttime UAV bridge images without paired normal-light references, DaL-MoE is evaluated through visibility analysis and damage detection, with improved visibility and downstream detection reported.

6.1.1. Brightness Enhancement

  • The mean grayscale intensity increases from 40.2 to 135.0, improving visibility of underexposed bridge surfaces without a dominant high-intensity accumulation.The improvement is particularly beneficial beneath bridge decks, in locally occluded regions, and around structural components.Source: Figure 6 · p. 24, §6.1.1 · p. 24

6.1.2. Structural Cue Recovery

  • Canny edge density rises from 1.55% to 4.07%, exposing more continuous boundaries and local cues, though it is only an auxiliary visibility descriptor and can include noise or amplified texture.The qualitative examples show stronger responses around boundaries, joints, textures, cracks, and spalling.Source: Figure 7 · p. 25, §6.1.2 · p. 25

6.2. Downstream Bridge Damage Detection

  • Enhanced-image detections generally align better with annotations than direct low-light inference, although missed detections and spurious predictions remain.Low-light inputs show missed detections, spurious responses, and inaccurate localization; enhanced images make several previously weak or missed damage regions detectable.Source: §6.2 · p. 26
  • On 200 annotated real nighttime images, DaL-MoE raises box mAP50 from 0.123 to 0.252 and box mAP50∶95 from 0.061 to 0.126.The same fixed YOLOv11m checkpoint is used without enhancement-specific fine-tuning; gains occur across all eight damage categories, including AP50∶95 for every category.Source: §6.2 · pp. 26–28, Table 7 · p. 27

7. Conclusions

DaL-MoE combines ISP-aware synthetic-pair construction with degradation-aware restoration, improving restoration quality and downstream detection on paired synthetic data and showing transfer to real nighttime imagery. These findings support its use as a restoration front end under the evaluated conditions, while broader validation remains necessary.

7. Conclusions

  • DaL-MoE uses ISP-aware synthesis to create 2,895 aligned normal/low-light pairs and combines degradation-aware guidance with experts for noise, color, and structural-detail recovery.It preserves box and instance-mask annotations for eight damage categories and outputs standardized RGB images that existing recognition models can process without architecture changes or joint training.Source: §7 · p. 28
  • On 482 paired synthetic test images, DaL-MoE achieved 23.1174 dB PSNR and 0.8482 SSIM, while YOLOv11m box/mask mAP50 rose from 0.3097/0.2281 to 0.4923/0.3529.YOLOv11m box mAP50∶95 increased from 0.2105 to 0.3578, and mask mAP50∶95 increased from 0.1051 to 0.1655. Similar improvements were observed for YOLOv8m, YOLOv8n, YOLOv11n, and SCNet; ablations supported contributions from degradation-aware guidance, complementary expert branches, and adaptive gating.Source: §7 · p. 28
  • On 200 real nighttime images, fixed YOLOv11m box mAP50 rose from 0.123 to 0.252 and mAP50∶95 from 0.061 to 0.126 without enhancement-specific fine-tuning.Dataset-average grayscale intensity increased from 40.2 to 135.0 and average Canny edge density from 1.55% to 4.07%; these are auxiliary visibility descriptors, while detection results provide task-level evidence. AP50 and AP50∶95 improved for all eight damage categories.Source: §7 · p. 28
  • The evaluation is limited by a single-system synthetic dataset, imperfect simulation of field imaging, a 200-image real set with box labels, and one real-scene detector.Real-scene instance segmentation and robustness across multiple detectors and independent sites are not established; a gap from normal-light performance remains, and restoration training does not explicitly use detection or segmentation supervision.Source: §7 · pp. 28–29
  • Future work will expand real low-light data across sites, bridge types, components, cameras, UAV platforms, and illumination conditions, adding instance masks and multiple architectures for evaluation.Planned work also includes using camera calibration and RAW sensor data to improve degradation modeling, and investigating high-resolution processing, damage quantification, spatial registration, and integration with BIM or digital twins, inspection records, and human verification.Source: §7 · p. 29

Related papers

Loading 2608.23136v1…