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

cs.CVAug 24, 2026

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

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

Organizations: School of Computer Science and Engineering, University of Electronic Science and Technology of China, No. 2006, Xiyuan Avenue, West Hi-Tech Zone, Chengdu, 611731, China · Sustainability X-Lab, The University of Hong Kong, Hong Kong, China · School of Architecture, Building, and Civil Engineering, Loughborough University, Loughborough, UK, LE11 3TU · Department of Engineering Science, University of Oxford, Oxford, UK · Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong, China

Abstract

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.

Figures & tables

Figure 1: Conceptual illustration of the motivation for low-light bridge damage detection.

Research level	Research stream	Typical data or methods	Representative capability	Main limitation relative to the present study
Data level	Bridge damage datasets	Crack classification, multi-label damage classification, multi-class detection, and segmentation 8 , 23 , 15 , 21	Provide bridge-specific damage categories, inspection scenes, and spatial annotations at diverse granularities	Lack strictly pixel-aligned normal-light–low-light pairs and shared detection/segmentation annotations across both illumination conditions
	Generic low-light datasets	Paired low-light restoration, nighttime object detection, and visible–infrared multimodal perception 35 , 7 , 22 , 17	Support full-reference restoration or high-level visual perception under low illumination	Focus on natural, traffic, and surveillance domains, lacking civil engineering semantics and bridge defect contexts
	Synthetic low-light data	sRGB-domain degradation, RAW-domain exposure/noise simulation, and day-to-night synthesis 38 , 2 , 26	Provide controllable physical parameters and spatially aligned normal-light–low-light samples	Are not tailored to multiple bridge defect categories and lack mechanisms to preserve and evaluate instance-level annotations
	Low-light infrastructure datasets	Normal- and low-light crack images and multi-darkness crack segmentation benchmarks 39 , 27	Provide specialized data for low-light enhancement and crack segmentation in civil structures	Remain strictly crack-centric and do not support simultaneous object detection and instance segmentation across multiple defect categories
Model level	Generic low-light enhancement and restoration	Retinex decomposition, illumination estimation, deep restoration backbones, and compound degradation recovery 18 , 14 , 12 , 41 , 42 , 36 , 3 , 37 , 11	Recover illumination, suppress noise, and restore color and fine details	Are optimized and evaluated on pixel reconstruction or perceptual quality, without explicitly preserving defect-critical boundaries and textures
	Low-light infrastructure perception	Low-light crack enhancement, reflectance modeling, and enhancement-assisted crack segmentation 39 , 13 , 27	Demonstrate the efficacy of illumination restoration for civil infrastructure crack identification	Target a single crack category with backbone-coupled pipelines, lacking multi-class evaluation and cross-framework adaptability

Table 1: Positioning of the present study relative to representative data- and model-level research streams for low-light bridge damage analysis.

Figure 2: Overview of the proposed DaL-MoE framework. (a) Overall architecture, consisting of a Degradation-Aware Guidance Estimation (DAGE) module and a U-shaped Adaptive Mixture-of-Experts Restoration (AMER) network. (b) Degradation-Aware Guidance Estimation module, which estimates a spatially varying illumination map and generates noise-aware guidance features from the low-light input. (c) Adaptive Mixture-of-Experts Restoration module, where noise-aware, color adjustment, and detail enhancement experts are dynamically integrated by an adaptive gating network for feature restoration.

Item	Content
Field acquisition
Number of bridges	7
Inspection contexts	4, including urban viaducts, mountainous bridges, railway bridges, and long-span highway bridges
Acquisition system	UAV equipped with a DJI H30T imaging system
Original image resolution	$4032\times 3024$ pixels
Dataset composition and partitioning

Table 2: Overview and class-wise instance distribution of the UAV bridge defect dataset and paired low-light benchmark.

Figure 3: Visualization of defect annotations shared between paired normal-light and synthetic low-light images.

Block	Input size	Output size
Light-up image $I_{0}$	$H\times W\times 3$	$H\times W\times 3$
Feature embedding ( $3\times 3$ Conv)	$H\times W\times 3$	$H\times W\times 40$
MoE Block (Encoder-1)	$H\times W\times 40$	$H\times W\times 40$
Downsampling	$H\times W\times 40$	$H/2\times W/2\times 80$
MoE Block (Encoder-2)	$H/2\times W/2\times 80$	$H/2\times W/2\times 80$
Downsampling	$H/2\times W/2\times 80$	$H/4\times W/4\times 160$

Table 3: Architecture of the proposed DaL-MoE and unified training configurations of the compared low-light image enhancement methods and downstream detector architectures.

Method	Venue	PSNR (dB) $\uparrow$	SSIM $\uparrow$
MIRNet 41	ECCV’20	20.2218	0.7903
Restormer 42	CVPR’22	20.6620	0.7624
LLFormer 33	AAAI’23	20.7207	0.7717
FourLLIE 31	ACM MM’23	21.1467	0.7490
Retinexformer 3	ICCV’23	21.9448	0.7565
UHDformer 32	AAAI’24	21.3763	0.7183

Table 4: Comparison of representative low-light image enhancement methods.

Detector	Task	Metric	Reference Inputs		Enhancement Methods
			Base	Normal	Retinexformer 3	HVI-CIDNet 37	DarkIR 11	Ours
YOLOv8m 29	Box detection	$\mathrm{mAP}_{50}$ $\mathrm{mAP}_{50{:}95}$	0.2787 0.1837	0.6710 0.4744	0.4336 0.2745	0.4193 0.2656	0.3908 0.2466	0.4806 0.3341
	Mask segmentation	$\mathrm{mAP}_{50}$ $\mathrm{mAP}_{50{:}95}$	0.1985 0.0853	0.3913 0.1644	0.2721 0.1169	0.2704 0.1105	0.2291 0.0953	0.2863 0.1251
YOLOv11m 19	Box detection	$\mathrm{mAP}_{50}$ $\mathrm{mAP}_{50{:}95}$	0.3097 0.2105	0.7723 0.5855	0.4495 0.2899	0.4232 0.2701	0.3140 0.1904	0.4923 0.3578
	Mask segmentation	$\mathrm{mAP}_{50}$ $\mathrm{mAP}_{50{:}95}$	0.2281 0.1051	0.5236 0.2409	0.2985 0.1260	0.2836 0.1209	0.2061 0.0844	0.3529 0.1655
YOLOv8n 29	Box detection	$\mathrm{mAP}_{50}$ $\mathrm{mAP}_{50{:}95}$	0.3267 0.2210	0.5720 0.3759	0.3611 0.2169	0.3606 0.2191	0.3099 0.1815	0.4167 0.2718

Table 5: Downstream bridge damage detection and instance segmentation performance across different detector architectures and image enhancement methods.

Figure 4: Visual comparison of low-light bridge image enhancement results. Input denotes the original low-light image, GT denotes the corresponding normal-light reference, and the remaining columns show the enhanced results produced by different methods.

Figure 5: Visual comparison of YOLOv11m bridge damage detection results under different input conditions. The same detector is applied to the low-light input, the enhanced images, and the corresponding normal-light reference.

Variant	Main Modules		AMER Components				Performance
	DAGE	AMER	$\boldsymbol{E_{1}}$	$\boldsymbol{E_{2}}$	$\boldsymbol{E_{3}}$	Gate	PSNR (dB) $\uparrow$	SSIM $\uparrow$
A. Main architecture modules
w/o AMER	✓	✗	–	–	–	–	21.65	0.7157
w/o DAGE	✗	✓	✓	✓	✓	✓	22.73	0.7716
B. AMER components
w/o Expert 1	✓	✓	✗	✓	✓	✓	22.03	0.7682

Table 6: Ablation study of the proposed DaL-MoE method.

Figure 6: Dataset-averaged grayscale intensity distributions of 200 real-world nighttime bridge images. DaL-MoE increases the mean intensity from 40.2 to 135.0.

Figure 7: Canny edge densities of 200 real-world nighttime bridge images, sorted by the density of the low-light inputs. The mean density increases from $1.55\%$ to $4.07\%$ .

Figure 8: Structural-cue comparison on representative real-world nighttime bridge images. From left to right: low-light input, input Canny edges, DaL-MoE output, and output Canny edges.

Figure 9: Qualitative YOLOv11m detection results on real-world nighttime bridge images. Label , Low-light , and Ours denote manual annotations, predictions on the original low-light images, and predictions after DaL-MoE enhancement, respectively. The detector and inference settings are fixed.

Class	Low-light		Enhanced		Absolute gain
	$\mathrm{AP}_{50}$	$\mathrm{AP}_{50{:}95}$	$\mathrm{AP}_{50}$	$\mathrm{AP}_{50{:}95}$	$\Delta\mathrm{AP}_{50}$	$\Delta\mathrm{AP}_{50{:}95}$
Spalling	0.079	0.033	0.240	0.109	+0.161	+0.076
Crack	0.050	0.019	0.125	0.057	+0.075	+0.038
Hole	0.224	0.096	0.418	0.179	+0.194	+0.083
Honeycombing	0.015	0.007	0.116	0.045	+0.101	+0.038
Seepage	0.186	0.099	0.355	0.205	+0.169	+0.106

Table 7: Class-wise box detection performance and absolute gains on 200 manually annotated real-world nighttime bridge images.

Explore similar work

Robust Lightweight Crack Classification for Real-Time UAV Bridge Inspection
Apr 30, 2026Wei Li, Haisheng Li, Weijie Li +3Convolutional Neural Networks Lightweight CNNs
Bay Area Super Bridge Maintenance Technology Center, Guangdong Provincial Highway Construction Co., Ltd., Guangdong, China · Guangdong AIHISUN Technology Co., Ltd., Guangdong, China
Compositional-Degradation UAV Image Restoration: Conditional Decoupled MoE Network and A Benchmark
Apr 10, 2026Jinquan Yan, Zhicheng Zhao, Zhengzheng Tu +3Image Restoration All-in-One Image Restoration
Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, School of Computer Science and Technology, Anhui University, Hefei 230601, China. · Key Laboratory of Intelligent Computing & Signal Processing (Anhui University), Ministry of Education, Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, School of Artificial Intelligence, Anhui University, Hefei 230601, China.
AMIEOD: Adaptive Multi-Experts Image Enhancement for Object Detection in Low-Illumination Scenes
May 7, 2026Xiaochen Huang, Honggang Chen, Weicheng Zhang +4Mixture of Experts Low-Light Image Enhancement
College of Electronics and Information Engineering, Sichuan University, Chengdu 610065, China · Police Integration Computing Key Laboratory of Sichuan Province, China

Apr 30, 2026J/K move · Enter open · S save

Robust Lightweight Crack Classification for Real-Time UAV Bridge Inspection

Wei Li, Haisheng Li, Weijie Li, Jiandong Wang+2

Bay Area Super Bridge Maintenance Technology Center, Guangdong Provincial Highway Construction Co., Ltd., Guangdong, China · Guangdong AIHISUN Technology Co., Ltd., Guangdong, China

Convolutional Neural Networks Lightweight CNNs Structural Health Monitoring

Open paper AskPDF arXiv

With the widespread application of Unmanned Aerial Vehicles (UAVs) in bridge structural health monitoring, deep learning-based automatic crack detection has become a major research focus. However, practical UAV inspections still face four key challenges: weak crack features, degraded imaging conditions, severe class imbalance, and limited computational resources for practical UAV inspection workflows. To address these issues, this paper proposes a unified lightweight convolutional neural network framework composed of four synergistic components: a lightweight backbone network, a Convolutional Block Attention Module (CBAM) for channel and spatial enhancement, a directed robust augmentation strategy based on inspection-scene priors, and Focal Loss for hard-sample learning under class imbalance. Experiments on the SDNET2018 bridge deck dataset show that the proposed method achieves an inference speed of 825 FPS with only 11.21M parameters and 1.82G FLOPs. Compared with the baseline model, the complete framework improves the F1-score by 2.51% and recall by 3.95%. In addition, Grad-CAM visualizations indicate that the introduced attention module shifts the model's focus from scattered regions to precise tracking along crack trajectories. Overall, this study achieves a strong balance among accuracy, speed, and robustness, providing a practical solution for ground-station assisted real-time deployment in UAV bridge inspections. The source code is available at: https://github.com/skylynf/AttXNet .

Abstract

Explore similar work

Robust Lightweight Crack Classification for Real-Time UAV Bridge Inspection

Compositional-Degradation UAV Image Restoration: Conditional Decoupled MoE Network and A Benchmark

AMIEOD: Adaptive Multi-Experts Image Enhancement for Object Detection in Low-Illumination Scenes