YOL026-RD: An End-to-End Road Damage Detection Network With Learnable Contrast Enhancement and Edge-Guided Downsampling

Research Title : YOL026-RD: An End-to-End Road Damage Detection Network With Learnable Contrast Enhancement and Edge-Guided Downsampling

Co-Researchers : Dr. Pawarotorn Chaipetch Miss Hathairat Samaikul and Mr. Theerayut Yonseng

The study investigates limitations of road damage detection systems applied to road survey imagery. Some damage instances are small relative to the image resolution, while Stride-4 downsampling can result in localization errors.

The researchers developed YOL026-RD by modifying the Detection Head, adding Learnable Contrast Enhancement, a module that learns its parameters from the detection loss, and EdgeSPD, a spatially lossless downsampler guided by a Sobel Prior.

In the experiments, YOL026-RD achieved an average mAP50 of 0.790 and mAP50-95 of 0.482 across five scales. It also achieved 98 frames per second on an entry-level accelerator at survey speeds of 21–100 km/h.

The research uses road-condition survey imagery captured by a camera mounted above the pavement and continuously recording frames. The detected information includes damage class, bounding box, confidence, timestamp, and position of each detected distress.