The maintenance of transportation infrastructure, particularly paved roads, is a critical component of national development strategies. Regular and accurate inspection for surface damage, such as cracks, potholes, and patches, is essential for ensuring safety and planning timely repairs. Traditional manual inspection methods are labor-intensive, inefficient, pose safety risks, and are susceptible to weather conditions. In contrast, Unmanned Aerial Vehicles (UAVs or drones) offer unparalleled advantages for large-scale, efficient infrastructure monitoring. Their high mobility and flexibility allow them to cover extensive areas and navigate complex environments, effectively overcoming the limitations of ground-based surveys. Equipped with high-resolution imaging systems, UAV drones can rapidly capture detailed aerial imagery of road surfaces over long distances, providing the high-quality data necessary for pixel-level segmentation of pavement distress. This capability is foundational for subsequent quantitative analysis of damage parameters, significantly enhancing inspection efficiency and data reliability.
Conventional methods for pavement damage detection often rely on low-level features like color histograms, gradient information, or texture descriptors. While useful in constrained scenarios, these approaches generally lack the robustness and accuracy required for complex, real-world road environments characterized by varying lighting, shadows, road markings, and debris. The advent of deep learning, particularly Convolutional Neural Networks (CNNs), has revolutionized this field. Models such as HRNet, which maintains high-resolution representations throughout the network, and PSPNet, which utilizes pyramid pooling to capture multi-scale context, have demonstrated significant improvements in segmentation tasks. Specific applications to pavement inspection include refined DeepLabV3+ models with post-processing for tunnel cracks, lightweight networks based on MobileNetV2, and integrated detection-segmentation frameworks. However, despite their powerful feature extraction capabilities, CNNs are inherently limited by their local receptive fields, which can hinder the modeling of long-range dependencies crucial for understanding the global context of a road scene, especially for elongated cracks or distributed damage patterns.
The introduction of Vision Transformers, built upon self-attention mechanisms, has addressed this limitation by enabling effective global context modeling. Architectures incorporating hierarchical transformers or models like Swin Transformer have shown promise in improving crack segmentation, particularly for thin and discontinuous fractures. Hybrid models that attempt to marry the local feature extraction prowess of CNNs with the global relational strength of Transformers represent a promising direction. Yet, these hybrid approaches often come at the cost of increased model complexity and computational burden. Furthermore, a predominant focus in existing research has been on crack detection, with relatively less attention paid to other common pavement distress types like potholes and patches. The significant morphological differences between these damage types, combined with challenges like blurred boundaries and complex background interference (e.g., shadows, oil stains, lane markings), continue to pose substantial difficulties for achieving precise and robust segmentation from UAV drone imagery.
To address these challenges, this work proposes an enhanced segmentation model named MDPR-DeepLabV3+, specifically designed for processing UAV-captured aerial pavement images. The proposed framework introduces several key innovations to improve segmentation accuracy, computational efficiency, and robustness in complex environments.
1. The MDPR-DeepLabV3+ Framework for Pavement Damage Segmentation
While the DeepLabV3+ architecture has proven effective for semantic segmentation, including some pavement analysis tasks, it faces limitations when applied to UAV drone imagery. Its capacity to identify fine, intricate cracks and to handle the wide variety of damage scales (from hairline cracks to large potholes) can be insufficient. Moreover, the original model’s parameter count is substantial, leading to high computational complexity that may hinder deployment in resource-constrained scenarios common in UAV-based edge computing. Our proposed MDPR-DeepLabV3+ model builds upon the DeepLabV3+ foundation and incorporates four major components: a lightweight backbone network, an enhanced encoder module, an attention mechanism for low-level features, and a refined decoder fusion module. The overall architecture is illustrated in the structure diagram.
1.1 Lightweight Backbone: MobilenetV2
The original Xception backbone in DeepLabV3+, while powerful, is deep and computationally expensive. To enhance operational efficiency suitable for UAV drone applications, we replace it with MobileNetV2. This network maintains a strong balance between feature representation capability and computational footprint through its innovative design elements: inverted residual blocks, linear bottlenecks, and depthwise separable convolutions. The inverted residual structure first expands the channel dimension using a 1×1 convolution to learn rich features in a higher-dimensional space. It then applies lightweight depthwise separable convolutions for spatial feature extraction before projecting features back to a lower dimension. Residual connections are incorporated where applicable to facilitate gradient flow. The linear bottleneck replaces non-linear activations in the narrow layers to prevent information loss. This design makes MobileNetV2 an ideal, efficient feature extractor for the high-resolution images typically acquired by UAV drones during road inspection missions.
1.2 The DFSP Module in the Encoder
The standard Atrous Spatial Pyramid Pooling (ASPP) module in DeepLabV3+ processes features through parallel convolutions with different dilation rates. However, the branches operate independently, and their outputs are merely concatenated before a final 1×1 convolution, which may not achieve optimal multi-scale feature integration. While a global average pooling branch is included, its ability to model long-range dependencies in complex road scenes remains limited. To overcome these issues, we propose the Dense Feature Synthesis Pyramid (DFSP) module for the encoder, comprising a Dense ASPP component and a Feature Fusion Module (FFM).
The Dense ASPP introduces dense connections between the ASPP branches, allowing features from different receptive fields to interact and complement each other progressively. This design reduces information loss and captures richer multi-scale contextual details, which is vital for detecting damage of varying sizes from UAV drone perspectives. Concurrently, a separate branch passes the deep features from the MobileNetV2 backbone directly forward, providing supplemental structural and textural information for hard-to-distinguish targets.
The core innovation of DFSP is the FFM, which enhances global semantic understanding. The FFM takes two inputs: the deep feature map from the backbone ($FI \in \mathbb{R}^{H \times W \times C}$) and the output prediction map from the Dense ASPP ($PM \in \mathbb{R}^{H \times W \times C}$). It leverages a cross-attention mechanism inspired by Transformer architectures to enable global information exchange between these two feature sequences. To manage computational complexity, a sequence reduction strategy is employed, downsampling the key and value vectors by a factor $R$, reducing the complexity from $O(N^2)$ to $O(N^2/R)$.
Multi-head cross-attention is used to extend the model’s expressive power. For each attention head $i$, queries, keys, and values are derived for both $FI$ and $PM$:
$$Q_{FI}^{i}, K_{FI}^{i}, V_{FI}^{i} = \text{Linear}_{i}(FI), \quad Q_{PM}^{i}, K_{PM}^{i}, V_{PM}^{i} = \text{Linear}_{i}(PM)$$
The attention matrices are computed as:
$$
\begin{aligned}
S_{FI}^{i} &= \text{Softmax}\left(\frac{Q_{FI}^{i} (K_{PM}^{i})^{T}}{\sqrt{d_{head}}}\right), \\
S_{PM}^{i} &= \text{Softmax}\left(\frac{Q_{PM}^{i} (K_{FI}^{i})^{T}}{\sqrt{d_{head}}}\right)
\end{aligned}
$$
The output for each sequence is obtained by applying the attention matrix to the other sequence’s values and concatenating across heads:
$$
Z_{FI} = \text{Concat}_{i}(S_{FI}^{i} V_{PM}^{i}), \quad Z_{PM} = \text{Concat}_{i}(S_{PM}^{i} V_{FI}^{i})
$$
The resulting $Z_{FI}$ and $Z_{PM}$ are reshaped and added to their original input features, followed by a residual convolution block to fuse them into a final enhanced feature map $Fuse \in \mathbb{R}^{H \times W \times C}$. The DFSP module thus effectively integrates multi-scale local details with globally informed context, significantly boosting the model’s perception of pavement damage in complex UAV-captured scenes.
1.3 PSA_M Attention for Edge Recovery
UAV drone imagery of pavements often suffers from complex backgrounds, blurred damage boundaries, and strong interference from lighting and shadows. To better process the shallow feature information before it is passed to the decoder, we incorporate an improved attention mechanism called PSA_M. This module builds upon the Polarized Self-Attention (PSA) but replaces its spatial branch’s global average pooling with a Mixed Pooling Module (MPM). The MPM combines adaptive average pooling for multi-scale context with Strip Pooling to enhance spatial perception along elongated structures—a common characteristic of linear cracks.
For an input feature map $x \in \mathbb{R}^{H \times W}$ of a single channel, strip pooling computes horizontal and vertical pooled vectors:
$$y^{h}_{i} = \frac{1}{W} \sum_{0 \leq j < W} x_{i,j}, \quad y^{v}_{j} = \frac{1}{H} \sum_{0 \leq i < H} x_{i,j}$$
These are then expanded and fused element-wise: $y_{i,j} = y^{h}_{i} + y^{v}_{j}$.
The PSA_M mechanism employs polarization filtering, maintaining two parallel self-attention branches: a channel-only branch and a spatial-only branch. It keeps a higher channel dimension (C/2) in the channel branch and the full spatial dimensions [H, W] in the spatial branch to minimize information loss during dimension reduction. The channel branch uses a Softmax function for attention enhancement, while the spatial branch uses a Sigmoid function, allowing the model to adapt to different feature modeling needs and improve fine-grained regression. The outputs of the two branches are fused with the original input feature $X$:
$$
\begin{aligned}
A_{ch}(X) &= \sigma_{q}(C_{q}(X)) \cdot \text{Softmax}\left(\sigma_{k}(C_{k}(X))\right)^{T} \\
Z_{ch} &= A_{ch}(X) \otimes_{ch} X \\
A_{sp}(X) &= \text{Sigmoid}\left(\text{Softmax}\left(F_{MP}(C_{q}(X))\right) \cdot \sigma_{v}(C_{v}(X))^{T}\right) \\
Z_{sp} &= A_{sp}(X) \otimes_{sp} X \\
\text{PSA}(X) &= Z_{ch} + Z_{sp}
\end{aligned}
$$
where $C_{q}, C_{k}, C_{v}$ are 1×1 convolutions, $\sigma$ are reshape operations, $F_{MP}$ is the mixed pooling function, and $\otimes_{ch}, \otimes_{sp}$ denote channel-wise and spatial-wise multiplication, respectively. By enhancing the extraction of spatial boundary and detailed information, the PSA_M module significantly improves the segmentation of various pavement distress types with ambiguous edges.
1.4 The RCD Module in the Decoder
The original DeepLabV3+ decoder uses standard convolutions to fuse deep and shallow features, which may lead to insufficient interaction and loss of fine-grained details from early layers. While the C2f (Concatenate to Fuse) module from YOLOv8 offers improved gradient flow and feature fusion through split and concatenate operations, its progressive convolutions are standard 3×3 kernels and may not optimally handle the multi-scale nature of damage from potholes to thin cracks.
We propose the Residual Channel Decoupling (RCD) module for the decoder. Based on the C2f framework, RCD replaces its Bottleneck units with a DualConv layer and introduces a residual connection pathway. The DualConv layer employs grouped convolutions, processing the same input feature channels with parallel 3×3 and 1×1 kernels within groups. This reduces parameters and computational cost while enhancing information flow efficiency and feature extraction diversity.
The residual branch preserves a baseline of the original fused features. A learnable weight mechanism dynamically balances the contribution between the feature-enhanced output from the C2f_Dual path and the residual features, preventing information loss and promoting complementarity between deep semantic and shallow spatial features. The fusion process in RCD can be expressed as:
$$O = \sigma(W_{res}) \otimes \text{conv}_{residual}(X) + (1 – \sigma(W_{res})) \otimes C2f_{dual}(X)$$
where $X$ is the input, $W_{res}$ is a learnable weight, $\sigma$ is the Sigmoid function, and $\otimes$ denotes element-wise multiplication. This dual-path design effectively enhances the model’s ability to capture fine details like small cracks and irregular pothole edges, which are critical for accurate UAV drone-based inspection.
2. Model Training and Evaluation
2.1 Hybrid Loss Function
Pavement damage segmentation from UAV drone imagery is a classic class-imbalance problem, where the damaged pixels (foreground) are vastly outnumbered by intact pavement pixels (background). Standard cross-entropy loss can cause the model to be biased towards the background. To address this, we employ a hybrid loss function combining Dice Loss and Focal Loss. Dice Loss optimizes from a region-overlap perspective, improving edge and detail recognition, while Focal Loss focuses learning on hard-to-classify pixels, mitigating the background dominance.
For $N$ classes, with $y_i$ as the ground truth label and $p_i$ as the predicted probability for pixel $i$, and $\epsilon$ as a smoothing factor, the losses are defined as:
$$
\begin{aligned}
L_{Dice} &= 1 – \frac{2 \sum_{i=1}^{N} p_i y_i + \epsilon}{\sum_{i=1}^{N} p_i + \sum_{i=1}^{N} y_i + \epsilon} \\
L_{Focal} &= -\alpha_i (1 – p_i)^{\gamma} \log(p_i)
\end{aligned}
$$
The total loss is: $L = L_{Dice} + L_{Focal}$.
2.2 UAV Pavement Damage Dataset
Given the scarcity of public datasets containing diverse pavement damage types from an aerial UAV drone perspective, we constructed a dedicated dataset. Images were captured using a DJI Matrice 300 RTK UAV drone equipped with a Zenmuse P1 camera. The original high-resolution images (8192×5460) were cropped into 1024×1092 patches to preserve detail. The dataset comprises 1885 images covering asphalt and concrete roads under various conditions (weather, lighting, altitude/speed). It includes three damage categories: cracks, potholes, and patches, with significant background complexities like shadows, road markings, and debris. The dataset was split into training/validation and test sets in an 8:2 ratio.
2.3 Evaluation Metrics
We use three standard semantic segmentation metrics to evaluate model performance:
1. Mean Intersection over Union (mIoU): Measures the average overlap between predicted and ground truth regions across all classes.
2. Mean Pixel Accuracy (mPA): The proportion of correctly classified pixels across all classes.
3. Mean Precision (mPrecision): The average ratio of correctly predicted positive pixels to all pixels predicted as positive for each class.
Let $y_{ij}$ be the number of pixels of class $i$ predicted as class $j$. The metrics are calculated as:
$$
\begin{aligned}
\text{mIoU} &= \frac{1}{N} \sum_{i=1}^{N} \frac{y_{ii}}{\sum_{j=1}^{N} y_{ij} + \sum_{j=1}^{N} y_{ji} – y_{ii}} \\
\text{mPA} &= \frac{1}{N} \sum_{i=1}^{N} \frac{y_{ii}}{\sum_{j=1}^{N} y_{ij}} \\
\text{mPrecision} &= \frac{1}{N} \sum_{i=1}^{N} \frac{y_{ii}}{\sum_{j=1}^{N} y_{ji}}
\end{aligned}
$$
2.4 Implementation and Training Details
The model was implemented using PyTorch and trained on an NVIDIA RTX 3090 GPU. Key training hyperparameters are summarized in Table 1.
| Parameter | Value / Type |
|---|---|
| Training Epochs | 300 |
| Input Size | 1024×1024 |
| Batch Size | 2 |
| Momentum | 0.9 |
| Optimizer | Adam |
| Initial Learning Rate | 5e-4 |
| LR Scheduler | Cosine Annealing |
| Loss Function | Dice + Focal |
3. Experimental Results and Analysis
3.1 Performance on the Custom UAV Dataset
We compared the proposed MDPR-DeepLabV3+ against several state-of-the-art segmentation models on our custom UAV pavement damage dataset. The results are presented in Table 2.
| Model | mIoU (%) | mPA (%) | mPrecision (%) | Params (KB) |
|---|---|---|---|---|
| Xception-DeepLabV3+ | 70.93 | 82.24 | 81.03 | 214745 |
| MobileNetV2-DeepLabV3+ | 69.13 | 84.90 | 75.20 | 22980 |
| MobileNetV3-DeepLabV3+ | 64.56 | 76.28 | 78.59 | 19593 |
| MobileNetV4-DeepLabV3+ | 49.83 | 69.38 | 59.23 | 30478 |
| HRNet | 60.87 | 76.73 | 71.26 | 258248 |
| PSPNet | 61.94 | 83.62 | 68.11 | 182807 |
| U-Net | 71.96 | 86.71 | 78.00 | 97249 |
| U-Net++ | 63.24 | 81.01 | 76.45 | 142336 |
| Segformer | 74.21 | 81.57 | 86.14 | 53505 |
| MDPR-DeepLabV3+ (Ours) | 78.47 | 92.03 | 83.93 | 30235 |
The proposed model achieves the highest mIoU (78.47%) and mPA (92.03%), demonstrating superior overall segmentation accuracy. While Segformer attains a slightly higher mPrecision (86.14%), it does so with nearly double the parameters of our model. Our method effectively balances high performance with model efficiency, making it well-suited for practical UAV drone inspection applications where computational resources may be limited. Visual comparisons further confirm that our model produces more coherent segmentation masks with fewer fractures in crack predictions, better shape adherence for potholes and patches, and reduced misclassification in complex backgrounds.
3.2 Ablation Study
To validate the contribution of each proposed component, we conducted a comprehensive ablation study, with results detailed in Table 3.
| Exp. | MobileNetV2 | DFSP | PSA_M | RCD | mIoU (%) | mPA (%) | mPrecision (%) | Params (KB) |
|---|---|---|---|---|---|---|---|---|
| 1 | 70.93 | 82.24 | 81.03 | 214745 | ||||
| 2 | √ | 69.13 | 84.90 | 75.20 | 22980 | |||
| 3 | √ | √ | 74.73 | 91.91 | 78.69 | 26454 | ||
| 4 | √ | √ | 71.04 | 87.09 | 77.52 | 22766 | ||
| 5 | √ | √ | 74.82 | 85.24 | 84.09 | 26509 | ||
| 6 | √ | √ | √ | 78.17 | 90.49 | 82.32 | 26477 | |
| 7 | √ | √ | √ | 76.55 | 86.80 | 84.81 | 26532 | |
| 8 (Full) | √ | √ | √ | √ | 78.47 | 92.03 | 83.93 | 30235 |
The study reveals several key insights: 1) Simply replacing the backbone with MobileNetV2 (Exp. 2) reduces parameters drastically but also reduces mIoU, indicating a trade-off. 2) The DFSP module (Exp. 3) provides a substantial boost in mIoU and mPA by enhancing multi-scale and global context integration. 3) The PSA_M module (Exp. 4) effectively improves edge-related metrics like mPA. 4) The RCD module (Exp. 5) significantly boosts mPrecision, indicating its strength in correctly identifying positive pixels. 5) The combination of modules progressively improves performance, with the full model (Exp. 8) achieving the best balance of high mIoU, mPA, and competitive mPrecision. This confirms that each component contributes uniquely and synergistically to the model’s success in segmenting UAV drone imagery.
3.3 Generalization on Public Dataset
To further assess the generalization capability of our model, we evaluated it on the public Crack500 dataset and compared it with other methods, including recent improvements to DeepLabV3+. The results are shown in Table 4.
| Model | mIoU (%) | mPA (%) | mPrecision (%) | Params (KB) |
|---|---|---|---|---|
| U-Net | 72.95 | 85.48 | 79.76 | 97248 |
| Xception-DeepLabV3+ | 76.14 | 89.88 | 81.28 | 214733 |
| Segformer | 79.07 | 95.79 | 81.75 | 53503 |
| Ref. [25] (Improved DeepLabV3+) | 57.21 | 75.09 | 70.60 | 14878 |
| Ref. [26] (Improved DeepLabV3+) | 72.50 | 66.00 | 61.00 | 1996 |
| MDPR-DeepLabV3+ (Ours) | 80.93 | 93.41 | 84.83 | 30233 |
Our model achieves the highest mIoU (80.93%) and mPrecision (84.83%) on the Crack500 dataset. While Segformer attains a higher mPA (95.79%), its precision is lower, suggesting it may classify more background pixels as cracks. Our model strikes a better balance, accurately segmenting crack regions with fewer false positives. Notably, compared to other recent DeepLabV3+ variants from the literature which have lower parameter counts, our model delivers significantly superior performance across all key metrics. This experiment confirms that the architectural improvements in MDPR-DeepLabV3+ are not dataset-specific but confer strong generalization ability for pavement crack segmentation, a core task in UAV drone inspections.
4. Conclusion
This paper presented MDPR-DeepLabV3+, an enhanced semantic segmentation model designed to address the challenges of pavement damage detection from UAV drone imagery. The model integrates a lightweight MobileNetV2 backbone for efficiency, a DFSP module in the encoder for robust multi-scale and global context fusion, a PSA_M attention mechanism to recover crucial edge details, and an RCD module in the decoder for effective multi-level feature integration. Extensive experiments on a custom UAV pavement damage dataset and the public Crack500 dataset demonstrate that our model achieves state-of-the-art or highly competitive performance in terms of mIoU, mPA, and mPrecision, while maintaining a reasonable model size. The ablation studies confirm the individual and collective contributions of each proposed component. The framework proves effective at handling the significant morphological variations among crack, pothole, and patch damage types under complex environmental conditions typical of aerial surveys. Future work will focus on further model lightweighting to enable real-time processing on UAV drone platforms, expanding the damage taxonomy, and exploring semi-supervised or self-supervised learning paradigms to reduce annotation dependency for large-scale UAV-based infrastructure inspection.