Self‑Supervised Learning for Road Distress Image Recognition

To address the critical issues in existing self‑supervised learning methods for road distress recognition—namely, the lack of domain‑specific priors, singular representational dimensions, insufficient multi‑scale feature modeling, and inadequate cross‑domain generalization capability—an end‑to‑end fully self‑supervised distress image recognition framework tailored to complex road scenes is constructed. Through a multi‑stage coupled functional module system, high‑accuracy distress representation learning and adaptive scene transfer are achieved driven by unlabeled data. This framework abandons the generic training paradigms established for general vision tasks. First, a differentiated masking generation strategy is designed, grounded in the inherent texture and gradient visual features of road distresses, so that focused learning on distress regions is enabled during the pretraining process, and domain‑specific prior knowledge of road scenes is incorporated into the self‑supervised representation learning pipeline. On this basis, a dual‑branch pretraining architecture that couples contrastive learning and masked reconstruction is established, by which global semantic discriminative features and local detailed features are fused, thereby enhancing the representational capacity of the model. To accommodate the wide‑ranging scale distribution characteristics of pavement distresses, an attention‑guided multi‑scale feature adaptive fusion mechanism is introduced into the framework, hierarchically integrating detail‑level, structural, and semantic features from different network stages so that accurate modeling and aggregation of multi‑scale distress features are achieved. Finally, a lightweight visual prompt adaptation mechanism is employed to perform cross‑domain feature alignment, effectively suppressing performance degradation caused by scene domain shift while preserving parameter efficiency. Each core module operates in a synergistic manner, following the progressive logic of feature screening, representation optimization, multi‑scale enhancement, and cross‑domain adaptation, thereby forming a comprehensive self‑supervised learning technical system. The design principles, network architectures, and optimization methodologies of each module are elaborated in detail in the following subsections.

2.1 Distress-aware differentiated masking generation strategy

The pretraining performance of masked autoencoders is highly correlated with the masking sampling strategy. The global random masking mechanism adopted by general-purpose algorithms indiscriminately masks and learns from all image patches without leveraging semantic differences in image content to optimize training emphasis. The feature distribution of road distress images exhibits significant scene-specific characteristics, with distress regions densely concentrating edge mutations and texture anomalies, which constitute the core effective information supporting distress recognition tasks. A uniform masking sampling approach attenuates the training weight of distress regions, rendering the model inefficient in capturing distress-specific features and substantially degrading the learning quality of self-supervised pretraining. To address this issue, a distress-aware differentiated masking generation strategy is constructed, grounded in the visual distribution patterns of road distresses, by which adaptive masking sampling is achieved through quantitative assessment of regional distress saliency, guiding the model to focus on high-value distress regions during the pretraining phase and accomplishing domain-specific representation learning tailored to road scenes.

image003.png

Figure 1. Flowchart of the distress-aware differentiated masking generation strategy

To precisely characterize the intra-image differences in distress saliency, a refined saliency detection system is built by integrating multi-directional gradient features and adaptive texture features, enabling accurate localization of potential distress regions. The flowchart of the distress-aware differentiated masking generation strategy is presented in Figure 1. Gradient feature extraction is performed using an eight-direction Sobel operator, which extends the conventional four-direction operator by incorporating diagonal gradient computations, thereby effectively enhancing the feature response to irregular distresses such as diagonal cracks. The global gradient response map is computed as follows:

$G=\sum_{d \in D}\left|\nabla_d X\right|, D=\left\{0^{\circ}, 22.5^{\circ}, 45^{\circ}, 67.5^{\circ}, 90^{\circ}, 112.5^{\circ}, 135^{\circ}, 157.5^{\circ}\right\}$   (1)

where, X denotes the input road distress image, and $\nabla_d X$ represents the gradient response of the image at the corresponding angle. Meanwhile, an adaptive window strategy is introduced for local texture variance computation, in which the statistical window size is dynamically adjusted according to the local gradient density. Large windows are employed in texture-uniform pavement regions, whereas small windows are adopted in texture-complex regions suspected of containing distresses, thereby improving the accuracy of texture feature detection while maintaining computational efficiency. The local texture variance is formulated as:

$V(i, j)=\frac{1}{k^2} \sum_{\substack{p=i-|k / 2|}}^{i+\mid k / 2]} \sum_{q=j-|k / 2|}^{j+|k / 2|}\left(X(p, q)-\mu_{i, j}\right)^2$   (2)

where, k denotes the adaptive window size and $\mu_{i, j}$ represents the mean pixel value within the local window. After normalization, the gradient feature map and the texture variance map are fused through a weighted summation, and the pixel-level distress saliency score map is obtained via Softmax mapping, expressed as:

$S=\operatorname{softmax}(\alpha \cdot \widehat{G}+\beta \cdot \widehat{V})$   (3)

where, $\widehat{G}$ and $\widehat{V}$ denote the normalized gradient response map and texture variance map, respectively. The fusion weights α and β are set to 0.6 and 0.4, respectively. This parameter combination is determined through grid search on the RDD2022 validation set samples, and is shown to optimally accommodate the feature distribution characteristics of road distresses.

Based on the pixel-level saliency scores, patch-level differentiated masking construction is performed, enabling precise alignment between the masking rules and the distress distribution. The input image is uniformly partitioned into standard 16 × 16 patches, and all patches are globally ranked according to their average saliency scores, by which high-saliency distress regions and low-saliency background regions are distinguished. Differentiated masking probabilities are assigned to these two categories so that adaptive focused learning is achieved during the pretraining task. The generation rule for the final masking matrix is defined as:

$M_i=\left\{\begin{array}{c}1, \text { if } i \in \text {high-saliency region and } \operatorname{rand}()<r_{\text {high }} \\ 1, \text { if } i \in \text {low-saliency region and } \operatorname{rand}()<r_{\text {low }} \\ 0, \text { otherwise }\end{array}\right.$   (4)

where, M_i denotes the masking state of the i-th image patch, with a value of 1 indicating that the patch participates in the masked reconstruction task and a value of 0 indicating that the patch features are directly preserved; rand( ) represents a random number uniformly distributed in the interval; and r_high and r_low correspond to the masking probabilities for high- and low-saliency regions, respectively. Through this mechanism, a higher reconstruction learning weight is imposed on the core distress regions, enabling thorough exploitation of distress texture and structural features, while redundant learning from ineffective background regions is attenuated, thereby enhancing the effectiveness and specificity of the pretraining representations.

To balance model training stability and targeted feature learning, an iteration-aware dynamic masking probability scheduling mechanism is designed, enabling a smooth transition of the training strategy. The masking probabilities for high- and low-saliency regions are adaptively updated with respect to the training epoch, with the scheduling functions formulated as:

$r_{\text {high}}(t)=0.5+0.3 \cdot \sigma\left(\frac{t-50}{20}\right), r_{\text {lows}}(t)=0.5-0.1 \cdot \sigma\left(\frac{t-50}{20}\right)$   (5)

where, t denotes the current training iteration epoch, and $\sigma()$ represents the Sigmoid activation function. In the early training stage, the masking probabilities for the two types of regions are kept consistent, both maintained at a level of 0.5, so that the model preferentially learns the fundamental visual features of road scenes and establishes a holistic understanding of pavement images. As training iterations proceed, the masking probabilities are smoothly adjusted within the specified intervals, ultimately converging to fixed values—the high-saliency region masking probability converges to 0.8, and the low-saliency region masking probability converges to 0.4. This progressive scheduling strategy effectively mitigates the representational learning bias induced by differentiated masking at the early training stage, prevents the model from converging to suboptimal local minima, gradually reinforces the model's capacity for learning distress features, and ensures the overall effectiveness of self-supervised pretraining.

2.2 Contrastive-reconstruction dual-branch collaborative self-supervised pretraining

Masked self-supervised learning and contrastive self-supervised learning possess distinct advantages in feature learning; a single learning paradigm is insufficient to simultaneously satisfy both the global semantic discriminability and the local textural granularity required for road distress recognition. Building upon the differentiated masking sampling mechanism established in the preceding section, a contrastive and reconstruction dual-branch collaborative self-supervised pretraining framework is constructed, in which deep integration of the two self-supervised tasks is achieved through a shared feature encoding network. The architecture of the contrastive-reconstruction dual-branch collaborative self-supervised pretraining framework is illustrated in Figure 2. The overall framework consists of a shared visual Transformer encoder, a contrastive learning task head, and a masked reconstruction task head. The encoder adopts the ViT-B/16 base architecture, comprising twelve Transformer encoding layers, with each layer equipped with twelve attention heads, a hidden feature dimension of 768, and a feed-forward network dimension of 3072. The encoder parameters are first initialized on the generic image dataset ImageNet-21K to acquire basic visual representational capacity, and subsequently domain-adaptive pretraining is performed on unlabeled road distress images, providing a stable and generalizable feature extraction foundation for the dual-branch collaborative learning.

image020.png

Figure 2. Architecture of the contrastive-reconstruction dual-branch collaborative self-supervised pretraining framework

The contrastive learning branch is centered on the instance discrimination task, by which the model is constrained to learn global semantic features through multi-view image transformations, thereby enhancing the discriminative capability across different road scenes and distress types. A multi-composition image augmentation strategy is designed to construct paired training samples, where two semantically consistent but visually distinct views of the same image are generated through joint transformations including scale-adaptive random cropping, multi-level color perturbation, Gaussian blurring, and horizontal flipping, effectively expanding the feature space of road scenes. The augmented views are fed into the shared encoder for feature extraction, and subsequently passed through a three-layer stacked multi-layer perceptron projection network for dimensionality reduction and semantic mapping, with the layer dimensions sequentially set to 768, 2048, 2048, and 128, and batch normalization and activation functions configured at each layer for feature regularization. The branch is optimized using the normalized temperature-scaled cross-entropy loss, by which the representational distribution is optimized through sample feature similarity measurement. The loss function is formulated as:

$L_{\text {cont}}=-\log \frac{\exp \left(\operatorname{sim}\left(z_i, z_j\right) / \tau\right)}{\sum_{k=1}^{2 B} 1_{[k \neq i]} \exp \left(\operatorname{sim}\left(z_i, z_k\right) / \tau\right)}$   (6)

where, $z_i$ and $z_j$ denote the dual-view features generated from the same original image, sim represents the cosine similarity function, $\tau$ denotes the temperature scaling parameter, set to 0.07 in this work for modulating the smoothness of the feature distribution, B denotes the training batch size, and the indicator function is employed to exclude the similarity computation of the sample with itself, thereby ensuring the validity of the instance discrimination task.

The masked reconstruction branch is dedicated to fitting local pixel and texture details, compensating for the deficiency of contrastive learning in capturing fine-grained distress features. A lightweight Transformer decoder is designed for this branch, comprising eight decoding layers, each equipped with eight attention heads and a hidden dimension of 512, by which reconstruction accuracy is ensured while computational overhead is reduced. The differentiated masking matrix is applied to the original input image, with the pixel information of unmasked regions preserved to construct corrupted images. The corrupted images are fed into the shared encoder for hidden feature extraction, and the complete pavement image information is subsequently reconstructed through the decoder. During training, the pixel reconstruction error is computed exclusively for the masked regions, concentrating feature learning on high-value distress areas. The adopted mean squared error reconstruction loss is defined as:

$L_{r e c}=\frac{1}{|M|} \sum_{t: M_i=1}\left|X_i-\widehat{X}_i\right|_2^2$   (7)

where, M denotes the image masking matrix, |M| represents the total number of masked patches, and $X_i$ and $\widehat{X}_i$ correspond to the pixel features of the original image and the reconstructed image, respectively. By constraining the pixel reconstruction fidelity in the masked regions, the model is effectively driven to learn the local texture and structural features of cracks, potholes, and other distress types with high precision.

To balance the learning weights of the dual-branch tasks and achieve synergistic optimization of global semantic and local detail features, an iteration-adaptive dynamic loss weighting mechanism is introduced for joint training. The two task branches share the encoder parameters, while only the independent projection head and decoder parameters are retained to adapt to the characteristics of their respective tasks, thereby avoiding feature redundancy while ensuring differentiated learning capability across the two tasks. The overall loss function is formulated as a dynamically weighted combination of the dual-branch losses, expressed as:

$L_{\text {total}}=L_{\text {cont}}+\lambda(t) \cdot L_{\text {rec}}$   (8)

where, λ(t) denotes the iteration-dependent balancing coefficient, which is dynamically updated according to a linear decay rule. The update formula is given by:

$\lambda(t)=\lambda_{\max }-\left(\lambda_{\max }-\lambda_{\min }\right) \cdot t / T$   (9)

where, the upper weight limit $\lambda_{\max}$ is set to 2.0 , the lower weight limit $\lambda_{\min}$ is set to 0.5, t denotes the current training iteration epoch, and T is the total number of pretraining epochs, fixed at 200. Through this mechanism, the reconstruction task is prioritized in the early training stage, enabling rapid learning of basic pavement textures and distress structural information. As training progresses, the contribution weight of the contrastive learning task is gradually increased, thereby reinforcing the model's global semantic discriminative capability. Ultimately, complementary fusion and synergistic optimization of the two types of self-supervised representations are achieved.

2.3 Multi-scale feature adaptive fusion module

Road distresses exhibit extremely non-uniform scale distribution characteristics. Fine cracks contain only a small number of pixel-level texture features, whereas potholes and contiguous alligator cracking cover extensive pavement regions, with scale spans exceeding two orders of magnitude. A single deep feature output can only emphasize either global semantic modeling or local detail capture, failing to accommodate the representational requirements of distresses at different scales, which results in the loss of small-scale distress features and insufficient semantic representation of large-scale distresses. To overcome the limitations of single-scale feature modeling, a multi-scale feature adaptive fusion module is designed, by which hierarchical capture of multi-level feature information and dynamic aggregation are achieved through the synergistic mechanism of scale-adaptive embedding and attention-guided fusion, constructing a unified representation system adapted to full-scale road distresses. The structure of the multi-scale feature adaptive fusion module is illustrated in Figure 3.

image041.png

Figure 3. Structure of the multi-scale feature adaptive fusion module

Leveraging the hierarchical feature representation characteristics of the visual Transformer encoder, a scale-adaptive embedding submodule is constructed to perform the selection and regularization of multi-scale distress features. The Vision Transformer encoder comprises twelve Transformer encoding layers, with features from different network layers exhibiting distinct representational emphases. Shallow-layer features preserve fine-grained high-frequency spatial information and are highly sensitive to small-scale distresses such as fine cracks. Mid-layer features integrate spatial structure and basic semantic information, adapting to medium-scale pavement damage. Deep-layer features focus on global contextual semantic modeling, effectively characterizing the overall distribution patterns of large-area distresses. The feature outputs from the third, seventh, and eleventh layers are selected as the sources of multi-scale representation, corresponding to detail, structure, and semantic core features, respectively. Through independent linear transformations, the three types of features are uniformly projected into a 768-dimensional hidden space, eliminating dimensional discrepancies across hierarchical features and providing regularized and unified feature inputs for subsequent cross-scale fusion.

To overcome the limitations of fixed-weight concatenation in fusion, a lightweight attention-guided fusion mechanism is adopted, by which the fusion weights of each scale feature are adaptively assigned according to the distress distribution state of the input image. First, the regularized multi-scale features are transformed through scale-specific projection matrices to achieve feature space alignment, with the computation formulated as:

$\widetilde{F}_s=W_s F_s s \in\{1,2,3\}$   (10)

where, $F_s$ denotes the original multi-level features extracted by the network, $W_s$ represents the learnable projection matrix specific to each scale, and $\widetilde{F}_s$ denotes the standardized features after dimensional and spatial alignment. Subsequently, the semantic contribution of each hierarchical feature is computed through a global learnable query vector and scale-specific key matrices, and the adaptive fusion weights are obtained after normalization:

$a_s=\frac{\exp \left(q^{\top} K_s \tilde{F}_s\right)}{\sum_{s^{\prime}=1}^3 \exp \left(q^{\top} K_{s^{\prime}} \tilde{F}_{s^{\prime}}\right)}$   (11)

where, q denotes the global learnable query parameter, K_s represents the scale-specific key projection matrix, and the weight coefficients satisfy the normalization constraint. Combined with the attention weights and the value-projected features, adaptive fusion of multi-scale information is accomplished. The final fused feature is computed as:

$F_{\text {fused}}=\sum_{s=1}^3 a_s \cdot V_s \widetilde{F}_s$   (12)

where, V_s denotes the scale-specific value projection matrix. This mechanism dynamically adjusts the contribution ratio of features at each level—enhancing the weight of shallow detail features for fine-crack images while emphasizing deep semantic features for large-area distress images—thereby achieving scale-adaptive intelligent fusion.

To ensure semantic synergy among multi-scale features while preserving their differentiated representations, a cross-scale consistency constraint is introduced, by which semantic drift and representational redundancy across multi-level features are suppressed. This constraint guides the alignment of hierarchical features in the semantic space by measuring their spatial distances. The loss function is defined as:

$L_{\text {scale}}=\sum_{s \neq s^{\prime}}\left\|s g\left(F_s\right)-F_{s^{\prime}}\right\|_2^2$   (13)

where, sg() denotes the stop-gradient operation, which fixes the parameter update path of the reference features and effectively prevents the mutual assimilation of features at different scales during iterative optimization, thereby preserving the unique scale-specific advantages of each level. This loss is incorporated into optimization exclusively during the self-supervised pretraining phase, with a weight coefficient set to 0.1, reinforcing feature synergy without compromising the diversity of multi-scale features, and further enhancing the completeness and robustness of the final fused representations.

2.4 Cross-domain self-supervised visual prompt adaptation

Through the multi-module collaborative self-supervised pretraining described in the preceding sections, the model is capable of acquiring refined, multi-scale distress representations from source-domain road data. However, significant domain distribution discrepancies exist among road images in real-world engineering scenarios. Variations in pavement materials, environmental illumination, and camera viewpoints across different geographic regions induce feature space shifts, leading to representational misalignment and substantial performance degradation when the pretrained model is directly transferred to unknown target domains. Conventional full fine-tuning approaches require updating all network parameters, which is not only computationally expensive but also prone to corrupting the generalizable distress features acquired during source-domain pretraining and triggering overfitting. To address these issues, a lightweight cross-domain visual prompt adaptation mechanism is introduced, in which all parameters of the pretrained backbone network are completely frozen, and only learnable visual prompt vectors are employed to modulate the feature extraction logic of the model. Through this mechanism, target-domain scene adaptation is achieved with minimal parameter overhead, effectively preserving the high-quality source-domain representations while compensating for the performance deficits induced by domain shift. A schematic diagram of the cross-domain self-supervised visual prompt adaptation mechanism is presented in Figure 4. To generate domain-specific prompt information tailored to the distress feature distribution of the target domain, a lightweight two-layer perceptron is constructed as the prompt generation network, which dynamically outputs prompt vectors conditioned on target-domain image features and distress saliency priors. The network input integrates mid-layer features from the encoder and target-domain saliency score features. The activation features from the sixth layer of the Vision Transformer network are selected as the semantic input, as this layer captures both shallow spatial details and deep semantic information, accurately reflecting the underlying visual distribution of target-domain images. After dimensional concatenation of the input features and saliency features, prompt vector decoding is accomplished through hierarchical transformations and nonlinear activations. The overall mapping process is expressed as:

$P=p_\omega\left(f_\theta\left(X_{\text {target}}\right), S_{\text {target}}\right)$   (14)

where, $p_\omega$ denotes the prompt generation network, $X_{\text {target}}$ represents the targetdomain input image, $S_{\text {target}}$ denotes the corresponding distress saliency score map, and P is the final output visual prompt vector. The dimensional transformation structure of the network is sequentially configured as the input dimension, a hidden layer dimension of 256, and the prompt vector dimension, with a Gaussian error linear unit nonlinear activation function incorporated in the intermediate layer to enhance feature mapping capability. The number of output prompt tokens is fixed at ten, ensuring both sufficient prompt informativeness and lightweight characteristics.

image048.png

Figure 4. Schematic diagram of the cross-domain self-supervised visual prompt adaptation mechanism

To fully leverage the domain-regulating capability of visual prompts, a prefix concatenation strategy is adopted to embed the prompt vectors into each encoding layer of the Transformer network, enabling global-level feature rectification. At each layer, independent prompt tokens are prepended to the input sequence before the image patch tokens, and these tokens participate jointly with the patch features in the self-attention computation. Through this mechanism, the target-domain distress prior information carried by the prompt vectors is able to guide the feature modeling process at each network layer, progressively correcting the representational deviations induced by domain shifts. Independent parameters are learned for the prompt tokens across all layers, with the total newly added parameters amounting to only 92,160—less than 0.1% of the total parameter count of the ViT-B/16 backbone network. This embedding approach requires no modification to the backbone network architecture or pretrained weights, thereby maximally preserving the distress representational capacity acquired during self-supervised pretraining, while achieving precise adaptation to target-domain feature patterns through lightweight parameter learning.

To constrain the learning direction of the prompt vectors and reduce the feature distribution discrepancy between the source and target domains, a domain-aware prompt alignment loss function is constructed, by which cross-domain feature alignment is achieved in an unsupervised manner based on the maximum mean discrepancy criterion. A multi-kernel Gaussian radial basis function combination is adopted to construct the metric function, integrating multiple kernel bandwidths to enhance the accuracy of distribution discrepancy measurement. The kernel bandwidths are set to 0.1, 1.0, and 10.0, respectively, with equal weights assigned to each kernel function. The cross-domain alignment loss and the maximum mean discrepancy computation are defined as:

$L_{\text {align}}=M M D\left(E_{X_s \sim D_s}\left[f_\theta\left(X_s\right)\right], E_{X_t \sim D_t}\left[f_\theta\left(X_t \mid P\right)\right]\right)$   (15)

$\operatorname{MMD}(P, Q)=\left\|E_{x \sim P}[\phi(x)]-E_{y \sim Q}[\phi(y)]\right\|_H^2$  (16)

where, $D_s$ and $D_t$ denote the source-domain dataset and the target-domain dataset, respectively, $\phi()$ represents the high-dimensional feature mapping function, and $H$ denotes the reproducing kernel Hilbert space. This loss function backpropagates only to update the parameters of the prompt generation network, while the backbone network weights remain fixed throughout. Leveraging unlabeled target-domain data, feature distribution adaptive alignment is accomplished, thereby achieving efficient and low-cost cross-domain scene transfer while preserving the model's generalization capability.

2.5 Overall training procedure and optimizer configuration

A staged progressive training paradigm is adopted for the proposed model, by which optimization is performed sequentially according to the logic of representation learning, cross-domain adaptation, and downstream task fitting. The training objectives, parameter update strategies, and hyperparameter configurations of each stage are independent of one another, ensuring ordered convergence of the model's feature learning. The first stage consists of source-domain self-supervised pretraining, where the contrastive loss, masked reconstruction loss, and multi-scale consistency loss are jointly optimized on large-scale unlabeled road images, with all parameters of the shared encoder, dual-branch task heads, and multi-scale fusion module updated iteratively in synchronization. The AdamW optimizer is employed for parameter updates in this stage, with an initial learning rate of 1 × 10^-4 and a weight decay coefficient of 0.05, which effectively constrains the magnitude of network parameters and suppresses overfitting. Training is conducted in a distributed manner across eight NVIDIA A100 graphics processing units, with a global batch size of 256 and a per-graphics processing unit batch size of 32. A cosine annealing learning rate schedule is adopted, with 10 warm-up iterations during which the learning rate is linearly increased from 1 × 10^-6  to the initial learning rate, followed by a smooth decay as training proceeds. The total number of pretraining iterations is set to 200 epochs, ensuring that all modules adequately learn the generalizable representational features of road distresses.

The second stage consists of unsupervised cross-domain prompt adaptation training, with the core objective of achieving domain-adaptive transfer of the model to unseen target scenes. In this stage, all parameters of the pretrained backbone encoder are completely frozen, fully preserving the distress representational capacity acquired from the source domain, while only the lightweight prompt generation network is optimized. The domain-aware prompt alignment loss is employed to enforce feature distribution consistency between the target and source domains. The AdamW optimizer is retained for this stage, with the learning rate set to 1 × 10^-3 and the weight decay coefficient adjusted to 0.01 to accommodate the lightweight network training characteristics, and the batch size is set to 128. The learning rate decay strategy is disabled, and a fixed learning rate is maintained throughout 50 training iterations. Through adaptive updates of the minimal set of learnable prompt parameters, the feature deviations induced by domain shift are precisely corrected, thereby efficiently enhancing the cross-scene generalization capability of the model without compromising the performance of the original pretrained model.

The third stage is dedicated to fine-tuning training and inference deployment for downstream tasks of road distress classification and segmentation, tailored to the requirements of real-world roadway network inspection. For the distress classification task, a linear classification layer is appended to the multi-scale fused features, and optimization is performed using the cross-entropy loss for category discrimination. For the pixel-level distress segmentation task, the fused features are upsampled to the original image resolution via transposed convolution, combined with one-dimensional convolution to achieve pixel-wise binary classification of distress and background. A layer-wise parameter update strategy is adopted during fine-tuning, where the parameters of the last four encoder layers are unfrozen to accommodate downstream task feature adaptation. The initial learning rate for the backbone network is set to 1 × 10^-5, while the initial learning rate for the classification and segmentation task heads is set to 1 × 10^-3. This differentiated learning rate configuration effectively balances the preservation of generalizable features and the fitting of task-specific features. The fine-tuning batch size is set to 64, with a total of 50 training epochs. After fine-tuning, the model can be directly deployed for road distress recognition and segmentation inference across various scenes.