A restoration system for dim bridge-inspection images was followed by higher damage-detection scores in both controlled synthetic tests and a real-night evaluation. The real-world result came from 200 nighttime UAV images and one fixed detector, so it does not settle how the approach will perform more broadly.
The work is an arXiv preprint, version 1, dated 24 August 2026. It presents DaL-MoE as an independently trained, plug-and-play, detector-agnostic restoration front end: a separate image-processing stage intended to work before a detector, without modifying the downstream detector or jointly optimizing the two.
A controlled low-light test
The controlled benchmark contained 2,895 spatially aligned normal-light and low-light image pairs, with shared bounding-box and instance-mask annotations for eight defect categories. The low-light counterparts were generated by unprocessing normal-light sRGB images into a RAW-like representation, applying sensor-domain degradation, and rendering them back to sRGB.
All enhancement methods used the same paired training data and augmentation protocol. Detectors used fixed checkpoints trained on the same normal-light split, with no enhancement-specific fine-tuning.
Synthetic scores and detector performance
On the held-out paired synthetic test set of 482 images, DaL-MoE reached 23.1174 dB in PSNR and 0.8482 in SSIM, the study's two measures of reconstruction against aligned references. Those figures come from the synthetic test, not the separate real-night evaluation.
Using the fixed YOLOv11m detector, enhanced-input results were higher on all four reported measures. The paper uses mean average precision, or mAP, as its summary score; box scores concern defect detections, while mask scores concern instance-segmentation outputs. Box mAP50 changed from 0.3097 to 0.4923, box mAP50-95 from 0.2105 to 0.3578, mask mAP50 from 0.2281 to 0.3529, and mask mAP50-95 from 0.1051 to 0.1655.
The pattern extended across the five detector architectures tested in the paired benchmark. Reported box mAP50 differences between DaL-MoE and Base favored DaL-MoE for YOLOv8m (0.2019), YOLOv11m (0.1826), YOLOv8n (0.09), YOLOv11n (0.1299) and SCNet (0.1127).
An ablation, or component-removal test, had lower reported reconstruction scores. Without DAGE, PSNR/SSIM were 22.73/0.7716; without the adaptive gate, they were 21.14/0.7397.
The real-night check
The transfer check used 200 real-world nighttime bridge images captured during UAV inspections. After enhancement, mean grayscale intensity was 135.0, compared with 40.2 before enhancement; mean Canny edge density was 4.07%, compared with 1.55%.
With the same fixed YOLOv11m checkpoint and inference settings, overall box mAP50 changed from 0.123 on low-light inputs to 0.252 on enhanced inputs, an absolute gain of 0.129. Box mAP50-95 changed from 0.061 to 0.126, an absolute gain of 0.065.
But the real-world quantitative evaluation was limited to 200 images, bounding-box annotations and one fixed YOLOv11m detector. It did not establish real-scene instance segmentation or robustness across multiple detectors and independent sites.
What the results do not establish
The evidence is narrower than a general claim about bridge inspection. DaL-MoE was tested as a detector-agnostic restoration front end, and its real-scene quantitative evaluation remained limited. The paired low-light data were generated through the stated RAW-like degradation and sRGB-rendering pipeline.
Because the paired low-light images were generated, the synthetic reconstruction scores describe performance against those constructed counterparts. The real test adds a transfer check, but it does not establish real-scene instance segmentation or robustness across multiple detectors and independent sites.
The document is identified as arXiv:2608.23136v1 and dated 24 August 2026. Its CRediT statement assigns Funding Acquisition to Hu Wang, the authors declare no commercial or associative conflict of interest, and the authors say data will be made available on request.
Paper data and sources
Original title: Bridge Damage Detection from Low-Light UAV Imagery via Degradation-Aware Mixture-of-Experts Enhancement
Authors: Hu Wang, Hongxu Pu, Zhiqi Hu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text