YOLO26-RD: An End-to-End Road Damage Detection Network With Learnable Contrast Enhancement and Edge-Guided Downsampling

Technical Summary: YOLO26-RD

Problem Statement
Conventional pavement-distress detectors typically specialize for small objects by adding a stride-4 detection head and replacing strided convolutions with lossless space-to-depth downsampling. This paper challenges the premise that road damage constitutes "small objects" in region-level survey imagery. Through a systematic audit of a dataset containing 7,618 images, the authors demonstrate that 70.37% of annotated instances are large, while only 1.28% are small at the 640² training resolution. Furthermore, a stride-4 detection level would consume 75.3% of the anchor budget despite covering a scale where essentially no ground truth exists. The failure decomposition of a baseline model confirms that complete misses, rather than localization errors, dominate the detection failures. Additionally, the imagery presents challenges of low, uneven contrast and thin structures that risk being discarded during standard downsampling.

Methodology
The proposed detector, YOLO26-RD, is built upon the YOLO26 architecture (an end-to-end, NMS-free single-stage detector) but undergoes significant structural modifications driven by the data audit rather than design intuition.

1. Architectural Reallocation: The stride-4 (P2/4) detection head is removed entirely. However, the P2/4 branch is retained in the neck to provide high-resolution features for the bottom-up path. This reallocates the assignment and loss budget to scales (P3, P4, P5) where the majority of instances actually exist.
2. LearnableContrast: A new stem module is introduced at the input. Unlike global contrast enhancement or fixed rules like CLAHE, this module applies a per-tile gamma and gain correction. It is a small, differentiable network (494 parameters) trained end-to-end from the detection loss, allowing it to adapt to local illumination variations within a single frame and remain active during inference.
3. EdgeSPD: To address the loss of thin structures during downsampling, the standard strided convolutions at the P2→P3 and P3→P4 transitions are replaced with EdgeSPD. This is a lossless space-to-depth downsampler (rearranging 2×2 pixel blocks into 4 channels) gated by a fixed Sobel gradient prior. The gate amplifies edge-bearing locations before rearrangement, ensuring that thin crack lines are preserved. This adds only 2 parameters per instance but increases channel depth significantly.
4. Inherited Components: The architecture retains Area Attention (A2C2f) and BiFPN-style skips from the YOLOv12 lineage, as well as the NMS-free head of YOLO26, which ensures deterministic prediction counts and latency.

Key Contributions

1. Data-First Audit Protocol: The paper introduces a reusable protocol for auditing annotation geometry and failure modes before designing an architecture. This audit revealed that the conventional "small-object" recipe was misaligned with the region-level annotation geometry of the dataset.
2. Counter-Trend Architecture: The study demonstrates that on region-annotated imagery, the P2/4 detection head is unnecessary. Removing it while retaining features improved mAP50 by 0.028 on the selection split and 0.009 on the validation split, while reducing epoch time by 7.4% and eliminating 75% of anchors.
3. Paired Ablation and Validation: The authors provide a rigorous ablation study showing that module-level gains measured on a selection split can vanish or reverse on a held-out validation split. The proposed architecture's superiority is shown to be a property of the design across scales, not just a tuned configuration.
4. Quantified Performance Ceiling: The paper identifies that the performance ceiling for the "crack" class is bounded by annotation geometry (aspect ratios and ambiguous box edges) rather than model capacity.

Results
Fifteen models were trained from scratch under a single recipe: five scales each of the proposed YOLO26-RD, a recipe-matched YOLO26 baseline, and a recipe-matched YOLOv12 family.

• Performance: Averaged over five scales, YOLO26-RD achieved 0.790 mAP50 and 0.482 mAP50-95, outperforming YOLO26 (0.776/0.471) and YOLOv12 (0.755/0.468).
• Scale Comparison: YOLO26-RD outperformed both reference families at every scale from 's' upward in mAP50. At scales 'm', 'l', and 'x', it led both references on both metrics in all twelve pairwise comparisons on the selection split.
• Best Model: YOLO26-RD-l was the top performer, achieving 0.809 mAP50 and 0.497 mAP50-95. It outperformed the YOLO26-l reference by 0.031 mAP50 and 0.030 mAP50-95, leading in all six per-class entries.
• Efficiency: As a TensorRT FP16 engine, the released model sustains 98 frames per second on an entry-level accelerator (RTX 5060), exceeding the 21 fps requirement for a 100 km/h survey.
• Validation Nuance: While the model led on the selection split, the validation split reproduced the ordering against YOLO26 but reversed the lead against YOLOv12 at the largest scale ('x'), highlighting the importance of validation splits in generalization claims.

Significance and Claims
The paper claims that the primary contribution is methodological: auditing data geometry and failure modes before architectural design can reverse standard assumptions (such as the necessity of a P2/4 head for road damage). The authors assert that the performance gains are architectural properties rather than artifacts of a single tuned configuration.

The work emphasizes that for on-board deployment, the NMS-free nature of the base template provides deterministic latency and post-processing time, which is critical for real-time survey systems. The authors modestly qualify their claims, noting that three of the twelve pairwise margins fall within the dataset's resolution limit (±0.015) and that the "crack" class performance is fundamentally limited by the ambiguity of region-level annotations rather than model capacity. The paper concludes that the proposed architecture is the most accurate detector measured in the study, particularly in the capacity range required for practical deployment.