Answers are grounded in this paper and cite their sources.
Accurate and precise segmentation of asphalt pavement defects is fundamental to automatic road maintenance and infrastructure analysis. While various methods have been proposed, many still rely on single-modality inputs or early-stage fusion strategies, which often fail to capture detailed surface complexity and fine-grained defects. Existing fusion strategies typically lack mechanisms to dynamically assess modal reliability or to explicitly model the anisotropic nature of defects such as cracks. To address these challenges, CODANet is proposed as a dual-stream segmentation framework centered on a cross-modal disagreement-aware (CODA) bottleneck. The architecture employs dual-stream encoders that process each modality independently to preserve modality-specific features. At the bottleneck, the CODA module pioneers a new fusion paradigm: it first computes a dense, per-pixel disagreement map by measuring the Jensen–Shannon divergence between provisional class predictions from each stream. This map is then used to learn spatially adaptive reliability weights, allowing the model to dynamically prioritize the more trustworthy modality. Furthermore, CODA incorporates oriented separable convolutions to diffuse features along horizontal and vertical axes, enhancing the representation of structurally anisotropic defects. An enhanced decoder leverages an edge-aware refinement module to sharpen structural boundaries. Extensive experiments on a large-scale dual-modal asphalt dataset and the public Crack500 benchmark demonstrate that CODANet establishes a new state-of-the-art, achieving a mean intersection over union (IoU) of 91.97% and an F1-score of 95.48% on the private dataset. The model demonstrates exceptional performance in segmenting challenging defect classes, including cracks, sealed joints, and potholes. Comprehensive ablation studies rigorously validate the efficacy of the proposed architectural components, including the CODA bottleneck and the dual-evolution encoder design.
No takes yet. Share an insight, caveat, or question.
Alhaek et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: