As shown in Fig. 2, the ARPD algorithm consists of three steps:

UAV-Based Pavement Dataset Construction: As illustrated in Fig. 2, the UAV-PDD2023 dataset is constructed using aerial images captured by unmanned aerial vehicles (UAVs). The dataset is randomly divided into training, validation, and testing subsets to ensure diversity and independence across sets.

As illustrated in Fig. 3, the backbone of the ARPD algorithm integrates four key modules: VssMamba, VCM, SPPF, and C2PSA. The detailed configuration parameters of the backbone are presented in Table 1.

(1)

VssMamba (Vision State Space Mamba) Module: As shown in Fig. 3, the VssMamba module incorporates Depthwise Convolution (DWConv) [32] for enhanced feature extraction at the input stage. This design enables the network to capture deeper and more expressive feature representations. To maintain efficiency and stability during training and inference, Batch Normalization (BN) and Layer Normalization (LN) are employed. The computation is governed by the following equations:

The S³ (Select State Space) module maps a univariate input sequence x(t) ∈ R to an output sequence y(t) via an implicit intermediate hidden state h(t) ∈ R^N, as defined by the following first-order differential equation: (4)h′=Ah(t)+Bx(t) (5)y(t)=Ch(t)where A∈RN×N, B∈RN×1 and C∈RN×1 are the state transition matrix, input projection matrix, and output mapping matrix, respectively. To enhance subsequent feature representation, Zero-Order Hold (ZOH) is employed in the S³M module for discretizing continuous feature signals [28]. For a continuous-time segment [t_a, t_b], the latent state representation l(tb) is given by: (6)l(tb)=eA⋅Δtl(ta)+∫tatbeA(tb−τ)Bx(τ)dτwhere Δt=tb−ta. In this expression, the first term represents the free evolution of the latent state, while the second term reflects the cumulative influence of the input sequence x(t) over the interval. In practice, approximate numerical techniques, such as matrix exponential approximation, are employed to compute eA⋅Δt, enabling the transition to a discrete-time formulation of state evolution. (2)

Vision Clue Merge (VCM) Module: Although CNN and Transformer-based architectures typically utilize convolution operations for downsampling, directional feature scanning may introduce interference during multi-path extraction. To address this issue, VMamba [33] employs 1 × 1 convolutions for dimensionality reduction, while MambaYOLO [34] utilizes 4× compressed pointwise convolutions for downsampling. Inspired by these methods, the proposed VCM module adopts a 3 × 3 convolution with stride 2 for spatial downsampling and complements it with pointwise convolution to preserve informative clues during resolution reduction.

(8)αc=σ(wc⋅GAP(X)+bc) (9)αs=σ(ws⋅GAP(X)+bs) (10)χo=αc⊙(αs⊙X)where αc and αs represent the channel and spatial attention weights, respectively, GAP denotes Global Average Pooling, w and b are learnable parameters, σ is the sigmoid activation function, and ⊙ denotes element-wise multiplication.

Existing neck structures predominantly rely on single-task feature pyramid networks (FPNs) to further process and enhance the features extracted by the backbone. In multi-scale object detection, classical backbone-neck-head architectures typically adopt FPN or Path Aggregation Network (PAN) for feature fusion, which have demonstrated promising results in transportation infrastructure maintenance scenarios. However, such neck designs often restrict inter-layer information transmission to intermediate layers only. To address this limitation, this study employs a lightweight skip connection-based neck structure for Semantics and Detail Infusion (SDI) [27], enabling more effective fusion of multi-scale pavement defect features. As illustrated in Fig. 4, a Transformer encoder is first applied to extract multi-level feature maps and align their output channels. For the i-th feature map, higher-level features (containing richer semantic information) and lower-level features (capturing finer details) are explicitly injected via simple Hadamard product operations, thereby enhancing both semantic and detailed representations of the i-th feature layer. The refined features are subsequently passed into a decoder for resolution reconstruction and segmentation.

Given the input feature map I∈RH×W×C (where H, W, and C denote the height, width, and channel number, respectively), the Transformer encoder generates M feature maps {fi0∈{f10,f20,…,fM0},1≤i≤M, each integrated with both channel and spatial attention, enabling local and global information representation across layers. (11)fi1=μic(δis(fi0))where fi1 represents the feature map of the i-th layer after fi0 is fused with attention, μicand δis represent the channel and spatial attention parameters, respectively. Subsequently, fi1 will be adjusted to have c channels through 1 × 1 convolution to obtain fi2∈RH×W×c. Furthermore, the feature map size is adjusted at each j-th layer to be sent to the decoder, calculated as follows: (12)fi,j3={𝒟(fj2,(Hi,Wi)),ifj<iℐ(fj2),ifj=i1≤i,j≤M𝒰(fj2,(Hi,Wi)),ifj>iwhere 𝒟, ℐ, and 𝒰 represent adaptive average pooling, identity mapping, and bilinear interpolation, respectively. Next, a 3 × 3 convolution is applied to each resized feature map fi,j3 to facilitate smoother integration of curved and multi-scale pavement defect features. (13)fi,j4=θi,j(fi,j3)where θi,j denotes the smoothing coefficient. The element-wise Hadamard product is then applied to all resized feature maps to enrich the i-th-level features with additional semantic information and finer details. (14)fi5=H[(fi,14,fi,24,…,fi,M4)]where H(⋅) denotes the Hadamard product. Finally, fi5 is assigned to the i-th decoder for further resolution reconstruction and segmentation, producing output results at three scales: large, medium, and small.

In addition, inspired by the state-of-the-art object detection advancements in YOLOv11, the C3k2 module, which integrates deformable convolutions and bottleneck enhancements, is incorporated into the neck of the ARPD algorithm to better address multi-scale pavement defect detection. As shown in Fig. 5, C3k2 employs CBS blocks with deformable convolutions of various kernel sizes (e.g., 3 × 3, 5 × 5), allowing the model to extract features across multiple scales and better capture complex spatial characteristics.

The APRD algorithm employs a hybrid loss function to regulate gradient updates, consisting of a classification loss Lc and a regression loss Lr, aiming to improve detection accuracy by simultaneously optimizing object classification and bounding box regression [17]. The loss is calculated as follows: (15)ℒ=γ1Lc+γ2Lrwhere γ1 and γ2 denote the weights for classification and regression loss, respectively.

The classification loss typically adopts Binary Cross-Entropy (BCE), which measures the difference between the predicted class probability p_i and the ground truth label y_i for each predicted box [16]. The formula for classification loss is expressed as: (16)Lc=−∑i(yilog⁡(pi)+(1−yi)log⁡(1−pi))

The regression loss is based on Complete Intersection over Union (CIoU), which evaluates the discrepancy between the predicted and ground truth bounding boxes in terms of center point coordinates, width, and height. For each predicted box, the IoU is calculated to derive the loss. The CIoU loss can be formulated as: (17)IoU=b∩bGTb∪bGT (18)Lr=1−IoU+ρ2(b,bGT)c2+ϑ⋅βwhere ρ(b,bGT) is the Euclidean distance between the centers of the predicted and ground truth boxes, c is the diagonal length of the minimum enclosing box covering both predicted and ground truth boxes, and ϑ and β represent the aspect ratio consistency term and the trade-off parameter, respectively.

As described above, the experimental data are sourced from the publicly available UAV-PDD2023 dataset [1], captured by a downward-facing camera mounted on a UAV flying steadily above road surfaces. A total of 2000 images are used for model training, which are randomly split into training, validation, and test sets in a 7:2:1 ratio, ensuring no image overlap across subsets. The dataset is categorized into six defect types based on visual characteristics: longitudinal cracks (LC), transverse cracks (TC), oblique cracks (OC), alligator cracks (AC), potholes (PH), and repairs (RP). The causes and potential hazards of each type are detailed in Table 2. To further assess generalization and real-time capabilities, inference is also performed on the RDD2022 dataset [31] and the test set.

The APRD model is trained, validated, and tested on an Ubuntu 22.04 desktop equipped with an Intel Core i7-12700 CPU and an NVIDIA GeForce RTX 3060 GPU. Key training parameters include a learning rate of 0.01 for balancing convergence speed and stability, and a weight decay of 0.0005 to prevent overfitting. A fixed momentum value of 0.937 is used to enhance gradient descent efficiency. The model is trained for 300 epochs with a batch size of 16 to ensure stability and thorough convergence.

In the comparative experiments, precision (P), recall (R), and mean Average Precision (mAP) are used as evaluation metrics. Precision measures the proportion of correctly identified instances among all predicted positive instances, while recall evaluates the model’s ability to correctly classify relevant instances. The definitions are as follows: (19)P=TPTP+FP×100% (20)P=TPTP+FN×100% (21)AP=∫01P(R)dR(22)mAP=∑incAPincwhere TP, FP, and FN represent true positives (positive samples correctly predicted as positive), false positives (negative samples incorrectly predicted as positive), and false negatives (positive samples incorrectly predicted as negative), respectively. The nc denotes the total number of categories.

To evaluate the effectiveness of the proposed ARPD model’s backbone and neck configurations, precision-recall (PR) curves are used to illustrate the balance between P and R. Figs. 6 and 7 present the ablation study results for the backbone and neck modules, respectively.

Backbone Ablation Study: Using a consistent C3k2-based neck structure across all configurations, different backbones are integrated into ARPD (including MobileNetV4 [26], U-NetV2 [27], DarkNet53 [17], and the proposed S³M) for comparison. As shown by the pink curve in Fig. 6, the S³M backbone achieves the best performance in extracting pavement defects, reaching the highest mAP of 80.1%. Notably, for AC and RP classes, the S³M-based model achieves over 90% precision, demonstrating the strong adaptability of S³M to diverse road surface damage types.

Neck Ablation Study: Building on the confirmed superiority of the S³M backbone, additional experiments are conducted by integrating various neck structures into ARPD while keeping the S³M backbone fixed. The tested neck modules include C3 [16], C2f [17], C3k2 [17], and the proposed SDI. As illustrated by the pink curve in Fig. 7g, the ARPD model equipped with both the S³M backbone and SDI neck structure achieves the best overall performance, with a mAP of 86.1%. This also reflects a significant improvement compared to the best result from the backbone ablation study (80.1%), confirming the effectiveness of SDI in multi-scale feature integration for pavement defect detection.

Based on the preceding ablation studies, the ARPD algorithm has demonstrated superior performance on UAV-based pavement defect datasets. To further validate its effectiveness, ARPD is compared against several state-of-the-art object detection models, including the YOLO series [16–18], MobileNet series [26,35], DETR series [36,37] and Mamba-based models [33,34], under the same dataset conditions.

As presented in Table 3, ARPD ranks first among all models with a mAP of 86.1%. The state-of-the-art detection framework YOLO11 demonstrates superior performance over YOLOv8 (83.1% vs. 78.4%), primarily due to its innovative architectural components such as C2PSA and C3k2, ranking second and third, respectively. The latest YOLO12 introduces A2C2f for enhanced attention and hierarchical training, it still struggles with sparse and multi-scale pavement defects, and is even outperformed by YOLOv8 in our experiments. Although Mamba-based architectures excel at image classification and long-range modeling, they typically require additional modules for fine-grained feature extraction. The complex characteristics of pavement defects significantly limit the effectiveness of state-space models in this context, indicating room for improvement. Moreover, while MobileNet variants are commonly adopted for lightweight deployment, they exhibit poor performance (<65%) in pavement scenarios characterized by minimal semantic information and complex visual backgrounds. The DETR series, while offering real-time end-to-end detection capabilities, continues to face significant challenges in balancing inference speed and detection accuracy.

In summary, the proposed combination of S³M and SDI achieves the best detection accuracy among all evaluated models in UAV-based pavement defect recognition, making it a reliable component of the ARPD algorithm for automatic road damage inspection.

Visualization results based on the test set are shown in Fig. 8, where the first and second rows represent the original UAV images and the corresponding ARPD predictions. The proposed algorithm demonstrates accurate localization of elongated pavement defects, including fragmented and irregularly distributed oblique cracks. Furthermore, even when multiple types of defects with similar visual features appear simultaneously, ARPD maintains precise detection performance. These results confirm that ARPD exhibits strong adaptability and robustness even in scenarios with minimal pixel-level defect presence.

Additionally, a separate visualization experiment is conducted on 200 randomly selected images from the public RDD-2022 dataset. As shown in Fig. 9, images captured from an in-vehicle perspective often led to vertical cracks (LC) being misidentified as oblique cracks (OC), as indicated by the red arrows. Despite this visual ambiguity, ARPD successfully detects all defect instances in each image, further demonstrating its generalization capability and robustness, and highlighting its suitability for real-world road inspection tasks involving diverse defect types and imaging conditions.

To further evaluate the reproducibility and generalization capability of the proposed algorithm, additional comparative experiments are conducted on publicly available UAV-captured pavement defect datasets, specifically Drone-based IRP [38] and HighRPD [39]. As shown in Table 4, all benchmark models are evaluated under identical configurations, and our method consistently achieved the highest accuracy. Notably, this experiment also assessed training FPS. Although RT-DETR outperformed the YOLO series in accuracy under ample computational resources, it lacks mobile-device compatibility and is surpassed by MambaYOLO in processing speed. Despite being a recent state-of-the-art detector, YOLOv11 shows performance limitations in complex pavement conditions. Overall, these results highlight the strong generalization and practical deployment potential of the proposed algorithm in real-world UAV-based pavement defect inspection.