Objective Automated pavement distress detection supports timely road maintenance and traffic safety, but reliable detection remains difficult in complex backgrounds. Pavement distresses vary substantially in scale and morphology. Longitudinal, transverse, and alligator cracks may be thin, elongated, discontinuous, or distributed over large regions, whereas potholes and repaired areas often have irregular boundaries. RT-DETR offers an end-to-end framework with a favorable accuracy-speed trade-off. However, its conventional local feature extraction, direct cross-layer concatenation, and stride-based downsampling are not specifically designed for these pavement characteristics. Local convolution may fail to preserve long-range crack continuity. Shadows, water stains, manhole covers, lane markings, and complex pavement textures can produce strong shallow responses that are propagated together with useful distress details. Spatial compression may also attenuate weak edges and fine structures. These factors make it difficult to improve detection accuracy without increasing model complexity. To address them, this study developed LCD-DETR, an improved RT-DETR-based detector for stronger structural representation, less background interference during feature fusion, and better edge preservation at a relatively low computational cost.

Methods LCD-DETR modified the baseline at three complementary stages. 1) a parallel dilated inverted residual mobile block (PD-iRMB) was introduced into the backbone. It combined a lightweight inverted-residual branch for local high-frequency details with a parallel dilated depthwise-convolution branch for wider-context sampling. Element-wise addition aligned the two branches, and a convolution smoothed the merged channels. A residual path produced the output. This design represented local crack edges and long-range topological cues within one lightweight unit. Dilation rates 1, 2, 3, and 4 were compared under the same network structure and training settings. 2) a lightweight dynamic gated fusion (LDG-Fusion) mechanism was constructed at cross-layer fusion nodes. Deep semantic features and shallow high-resolution features were aligned and concatenated. Channel and spatial attention refined their joint representation. Two initial gating responses were then generated and normalized with Softmax to assign complementary spatial weights to the deep and shallow branches. Addition of the weighted features produced the fused output. This process explicitly adjusted each feature level's contribution instead of using direct concatenation. 3) deep robust feature downsampling (DRFD) replaced conventional single-path downsampling. DRFD aggregated three parallel paths: interleaved spatial slicing followed by channel projection, stride-2 depthwise convolution for local-neighborhood extraction, and max pooling for local-extrema retention. Channel concatenation and projection combined spatial details, local structures, and salient responses. Experiments were conducted on RDD2022_China. This subset contains five classes: D00 longitudinal cracks, D10 transverse cracks, D20 alligator cracks, D40 potholes, and Repair. The data were divided into training, validation, and test sets at a 7:2:1 ratio. Six ablation configurations were used to evaluate individual modules and their combinations. Comparisons covered five attention mechanisms, four dilation rates, representative YOLO detectors, RT-DETR variants, and recent DETR-based pavement detectors. Cross-domain evaluation was performed on UAV-PDD2023. The reported metrics included precision, recall, F1-Score, mAP@0.5, mAP@0.5:0.95, parameter count, GFLOPs, and FPS.

Results and Discussions On RDD2022_China, the complete LCD-DETR achieved 87.2% precision, 80.6% recall, an F1-Score of 83.7%, 85.4% mAP@0.5, and 56.2% mAP@0.5:0.95, with 17.2 M parameters and 51.9 GFLOPs. Relative to the RT-DETR-r18 baseline, mAP@0.5 and mAP@0.5:0.95 increased by 3.1% and 2.7%, respectively. The parameter count decreased by 13.6%, and computational cost decreased by 5.1 GFLOPs. FPS on the reported RTX 4090 platform increased from 181.1 to 208.3. Class-wise AP@0.5 remained nearly unchanged for D00. It increased by 1.6%, 3.4%, 5.8%, and 5.1% for D10, D20, D40, and Repair, respectively. The ablation results revealed complementary effects at different network stages. PD-iRMB reduced the model to 17.1 M parameters and 51.0 GFLOPs while raising mAP@0.5 and mAP@0.5:0.95 to 83.9% and 54.8%. LDG-Fusion alone produced gains of 3.1% and 2.2% in the two mAP metrics. DRFD alone raised mAP@0.5:0.95 from 53.5% to 56.0%, the highest value under this stricter localization metric among the single-module settings. Their combination linked backbone structural modeling, cross-layer noise-aware fusion, and detail-preserving downsampling rather than repeating the same enhancement at multiple locations. In the dilation-rate study, rate 2 yielded 85.4% mAP@0.5 and 56.2% mAP@0.5:0.95. Larger rates slightly improved recall in one setting but reduced overall accuracy and localization performance under stricter IoU thresholds. This trend indicated a trade-off between contextual coverage and local structural preservation. LDG-Fusion also outperformed the compared CBAM, ECA, CoordAtt, SE, and SimAM configurations in mAP@0.5. This result supported explicit deep-shallow gating for this fusion task. Against RT-DETR-r50, LCD-DETR matched mAP@0.5 and improved mAP@0.5:0.95 by 0.4%. Parameters and computation decreased by approximately 60%. LCD-DETR also retained lower complexity than RSF-DETR and HG-DETR, although its mAP@0.5 was 0.4% below that of RSF-DETR. On UAV-PDD2023, LCD-DETR obtained 72.6% mAP@0.5, 45.4% mAP@0.5:0.95, and a 73.3% F1-Score. These two mAP metrics exceeded the baseline by 1.1% and 2.6%. Qualitative detection and heat-map comparisons showed fewer background false positives, fewer missed fine cracks, more continuous crack responses, and better coverage of irregular distress regions. These observations were consistent with the quantitative improvements but did not replace them.

Conclusions The results show that coordinated improvements to backbone representation, cross-layer fusion, and downsampling can enhance complex-background pavement distress detection without relying on a larger backbone or high-complexity module stacking. PD-iRMB strengthens local and long-range structural modeling, LDG-Fusion controls the transmission of background-sensitive shallow features, and DRFD preserves fine boundary information during spatial compression. Their combination improves accuracy, strict localization, model compactness, and the reported single-platform processing speed, while the UAV-PDD2023 results indicate a degree of cross-domain generalization. However, the FPS result is limited to the reported RTX 4090 environment and does not establish real-time performance on vehicle-mounted or UAV edge devices. Robustness under snow, rain, and nighttime low-light conditions also remains to be evaluated. Future work will therefore examine multi-hardware deployment, pruning and quantization for resource-constrained platforms, and frequency-domain enhancement or multimodal fusion for adverse environments.