Image recognition-based method for determining milling boundary of old road shoulder pavement in highway reconstruction and expansion

Abstract

The application discloses a kind of based on image recognition's highway reconstruction and extension old road shoulder pavement milling boundary determination method, realize the automatic identification and positioning of old road shoulder milling boundary.For the highway reconstruction and extension old road shoulder pavement milling boundary, obtain the multi-view image of road surface using unmanned aerial vehicle oblique photography.Construct improved YOLO11 detection model, utilize training set to carry out model training, optimize main network, neck network and head network, and introduce all-purpose image repair network suitable for a variety of complex scenes in main network, add bidirectional feature pyramid network in neck network, replace detection head with dynamically improved detection head with pertinence, to improve calculation efficiency and accuracy.Improve the YOLO11 detection model to identify shoulder milling boundary features, solve the problems of low efficiency, poor accuracy and insufficient safety in traditional manual survey, realize the intelligent identification of highway reconstruction and extension old road shoulder milling boundary.

Description

Technical Field

[0001] This invention belongs to the technical field of highway reconstruction and expansion engineering, and in particular relates to a method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition. Background Technology

[0002] In the field of highway reconstruction and expansion engineering, the effective connection between the old and new pavement structural layers is crucial to the overall integrity and service performance of the expanded road. Among these, the precise positioning of the milled boundaries of the original shoulder pavement is the core prerequisite and primary technical step for achieving a high-quality connection. Current engineering practices rely on traditional techniques that have significant limitations in efficiency, accuracy, and safety, becoming a prominent bottleneck restricting construction quality and project progress.

[0003] The most basic and widely used on-site manual marking method mainly relies on construction workers visually surveying the area using chalk or paint. This method is inherently limited by the operator's subjective judgment and manual operation, making it difficult to overcome random deviations caused by human visual errors, posture changes, and environmental interference. This results in significant fluctuations in the final milled boundary lines, failing to meet the stringent requirements of high-grade highways for smooth alignment and precise alignment of structural layers. More fundamentally, manual methods relying purely on visual observation lack the ability to detect defects hidden within the structure, such as pavement layer delamination and loose base layers. Improper boundary setting can easily lead to material waste due to over-milling or quality problems such as failure of the bond between old and new layers due to missed milling. At the same time, construction workers face higher safety risks when working on traffic sections for extended periods.

[0004] To reduce reliance on manual labor, the industry has attempted to introduce various automated or semi-automated inspection equipment. For example, contact or non-contact ranging devices such as laser profilers are used to acquire road surface profile data. However, these technologies are typically based on single-point or line scanning principles, and the collected data is essentially a discrete set of elevation points, unable to directly generate the continuous and complete geometric boundary lines required for construction. Furthermore, their optical or laser ranging modules are highly susceptible to interference from residual water stains, oil, or other contaminants on the road surface, leading to inaccurate ranging values ​​and insufficient reliability in complex actual road environments. Another approach is to use drones equipped with mapping cameras for aerial photography, generating digital surface models through post-processing. While this technology can acquire continuous spatial information, its complete process is highly specialized, encompassing flight data acquisition and extensive subsequent offline data processing (such as aerial triangulation, point cloud generation, and modeling). It typically requires cross-disciplinary professional teams, and the time required from data acquisition to the final output of construction-guided drawings is lengthy, often exceeding eight hours, making it difficult to meet the general demand for rapid and efficient construction in renovation and expansion projects in terms of timeliness and economic cost.

[0005] With the advancement of artificial intelligence, especially deep learning and computer vision technologies, automated recognition methods based on image semantic segmentation or object detection have provided new solutions to the aforementioned problems. These methods automatically learn and understand image features through algorithmic models, theoretically possessing recognition accuracy and robustness that surpasses manual rules. However, directly applying general image recognition models to the specific outdoor construction scenario of highway reconstruction and expansion faces severe challenges. The construction site environment is complex and variable, with strong and irregular lighting changes, shadows cast by vehicles and structures, inherent dirt and texture interference from the road surface, and image degradation caused by weather factors (such as rain and fog). Faced with these highly specific interference factors, the performance of general models or traditional algorithms is unstable, and their recognition boundaries are prone to breakage, false detections, or missed detections. To maintain usability, engineers often need to frequently manually adjust algorithm parameters according to the site conditions. This not only increases the complexity of use but also significantly reduces the "automation" level of the method, failing to fully unleash its technological potential.

[0006] Therefore, existing technologies, whether relying on subjective experience-based manual methods, instrumental methods based on discrete point measurements or heavy post-processing, or even general image recognition algorithms applied directly without specific optimization, have failed to solve the problem of determining the milling boundary of old road shoulders in highway reconstruction and expansion projects in an efficient, accurate, robust, and fully automated manner. The industry urgently needs an innovative technical solution that can overcome the aforementioned shortcomings and achieve intelligent, high-precision, and high-efficiency automatic determination of construction boundaries. Summary of the Invention

[0007] The purpose of this invention is to provide a method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition, so as to solve the problems of low efficiency, poor accuracy and insufficient safety of traditional manual surveying of old road shoulder pavement milling boundaries.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition, comprising:

[0009] Capture road surface images, process and label them, and construct a dataset;

[0010] An improved YOLO11 object detection network is constructed, and the improved YOLO11 object detection network is trained using the dataset to obtain an object recognition model;

[0011] The image of the milled boundary of the old shoulder pavement of the highway reconstruction and expansion is input into the target recognition model for recognition, and the bounding box and confidence score of the milled boundary of the old shoulder pavement of the highway reconstruction and expansion are output.

[0012] The improved YOLO11 object detection network introduces AirNet for image denoising, dehazing, and enhancement into its backbone network, introduces Bi-directional feature pyramid network BiFPN into its neck network, and replaces its detection head with an improved dynamic head.

[0013] Furthermore, the captured road surface image includes:

[0014] A drone equipped with a multispectral sensor was used to collect road surface images. The flight parameters were set to a forward overlap rate of no less than 80%, a lateral overlap rate of no less than 60%, and a camera tilt angle of 30°.

[0015] Further image processing includes:

[0016] It enhances the brightness, contrast, and HSV color space of the image, and performs geometric transformations including translation, flipping, and rotation.

[0017] Furthermore, the AirNet consists of a contrastive degradation encoder (CBDE) and a degradation-guided repair network (DGRN), and its training is based on an overall loss function. satisfy:

[0018]

[0019] in, The contrast loss of the CBDE is... It is the real label image Y t With the restored clean image Y p The reconstruction loss between them, the reconstruction loss Defined as:

[0020]

[0021] In the formula, For the input image The total number of pixels, For pixel index, Clean image recovered by the network At pixel position The predicted value, For the real label image Y t At pixel position The pixel value.

[0022] Furthermore, the BiFPN performs deep fusion of feature maps of different scales from the backbone network through a bidirectional path with learnable weights, both top-down and bottom-up; wherein, the 6th layer of the BiFPN uses top-down intermediate features... and bottom-up output features Defined by the following formulas respectively:

[0023]

[0024]

[0025] In the formula, These are the original input features from layer 6. It is a set of learnable weights. It is another set of independent learnable parameters. For numerical stability, For spatial resolution adjustment operations, For depthwise separable convolution, The original input features are for layer 7. This is the bottom-up output feature for the 5th layer.

[0026] Furthermore, the improved dynamic head utilizes an attention mechanism to unify target detection from three dimensions: scale awareness, spatial awareness, and task awareness. It combines the attention from these three dimensions in a concatenated manner, applying it to the input feature tensor. The processing satisfies generalized attention :

[0027]

[0028] in, , and These represent scale-aware attention, spatial-aware attention, and task-aware attention, respectively. The total number of layers in the feature pyramid. For spatial dimension size, For the number of channels, Representation of characteristic tensor The real number space to which it belongs and its shape dimensions.

[0029] Furthermore, the scale-aware attention Defined as:

[0030]

[0031] in, (⋅) is a linear function approximated by a 1×1 convolution. For hard sigmoid functions, The edge saliency mask is pre-computed by the Sobel operator, which makes the pooling process focus more on the interface between pavement structural layers and the edge region of cracks.

[0032] Furthermore, the spatial awareness attention Defined as:

[0033]

[0034] in, This represents the number of locations for sparse sampling. As a measure of the importance of self-learning, Learnable offsets for deformable convolutions. This is the hierarchical index of the feature pyramid. These are learnable weight coefficients. Indicates sparse sampling points. This represents the number of channels in the feature map. It is the first Initial reference position coordinates of sparse sampling points;

[0035] Furthermore, the sparse sampling points are arranged in a cross-shaped pattern along the horizontal, vertical, and diagonal directions, totaling 12 sampling points; the learnable offset Decomposed into offsets along the boundary tangent direction and offset along the boundary normal direction And satisfy:

[0036]

[0037] Among them, for learnable offsets Decomposed into offsets along the boundary tangent direction and offset along the boundary normal direction Apply differentiated learning constraints.

[0038] Furthermore, in the task-aware attention In this dimension, the detection head is decoupled into a boundary classification branch and a boundary regression branch, and independent learnable hyperparameters are configured for each branch. The activation methods of the boundary classification branch and the boundary regression branch respectively satisfy:

[0039]

[0040]

[0041] in, For the first The characteristic values ​​of each channel, , , , , , , , , All are achieved through hyperfunctions The learned hyperparameters, the hyperfunction In L Global pooling is performed along the dimension, and the hyperparameters are obtained through a fully connected layer and a normalization layer. For the input feature tensor, and These are the task-aware attention functions for the boundary classification branch and the boundary regression branch, respectively.

[0042] Furthermore, the feature maps at different scales include layers P2 to P7; wherein, the BiFPN-optimized P2 and P3 layer feature maps are used to capture minute cracks, small-sized defects, and milled boundary edge textures; the P4 and P5 layer feature maps are used to identify medium-sized pavement structure layer areas and defect blocks; and the P6 and P7 layer feature maps are used to locate large-scale milled areas and understand the overall shoulder layout.

[0043] Compared with existing technologies, the beneficial effects of this invention include: The method and system for automatically determining the milled boundary of old road shoulder in highway reconstruction and expansion based on UAV aerial photography and image recognition utilizes UAV oblique photography to acquire multi-view images of the road surface for the milled boundary of old road shoulder in highway reconstruction and expansion. The images are preprocessed and labeled, and divided into training, validation, and test sets. An improved YOLO11 detection model is constructed, introducing AirNet (an all-around image inpainting network) suitable for various complex scenarios into the backbone network, adding BiFPN (a bidirectional feature pyramid network) to the neck network, and replacing the detection head with an Improved Dynamic Head. The improved YOLO11 model is used to identify the features of the milled road shoulder boundary, solving the problems of low efficiency, poor accuracy, and insufficient security of traditional manual surveying, and achieving intelligent identification of the milled boundary of old road shoulder in highway reconstruction and expansion. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of milling boundary recognition in an embodiment of the present invention;

[0046] Figure 2This is an image of the milled boundary of the old shoulder pavement during highway reconstruction and expansion, acquired by an unmanned aerial vehicle in an embodiment of the present invention.

[0047] Figure 3 This is a schematic diagram illustrating the milling boundary marking of the old road shoulder pavement in the highway reconstruction and expansion embodiment of the present invention;

[0048] Figure 4 This is an improved YOLO11 target detection network structure in this embodiment of the invention;

[0049] Figure 5 This is a network architecture diagram of AirNet in an embodiment of the present invention;

[0050] Figure 6 This is a network architecture diagram of BiFPN in an embodiment of the present invention;

[0051] Figure 7 This is a network architecture diagram of the Improved Dynamic Head in this embodiment of the invention;

[0052] Figure 8 These are the training results of the target detection neural network model with attached detection boxes and classification probabilities in this embodiment of the invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] like Figure 1 As shown, this embodiment provides a method for determining the milling boundary of the old shoulder pavement in highway reconstruction and expansion based on image recognition, specifically including the following steps.

[0055] S1, capture road surface images: use drones to acquire high-definition images of the road surface, perform image processing and annotation on the captured images, and divide them into training set, validation set and test set;

[0056] S101, In this embodiment of the invention, a positioning drone equipped with a multispectral sensor is used to acquire road surface images. Specific flight parameters are set as follows: forward overlap rate not less than 80%, and lateral overlap rate not less than 60%. To optimize the capture of milled boundary features of the old road shoulder, the camera tilt angle is specifically set to 30°. This tilt angle parameter is a preferred solution determined after comprehensive consideration for the specific task of old road shoulder boundary identification in highway reconstruction and expansion scenarios. It aims to simultaneously achieve multiple objectives, including geometric distortion control, acquisition of lateral texture information such as road shoulder elevation and joints, suppression of illumination interference such as road surface reflection and shadows, and highlighting potential defects such as cracks. In some exemplary embodiments, the acquired images are as follows: Figure 2 ;

[0057] In this embodiment, 30° is chosen as a moderate tilt angle, which ensures the acquisition of lateral information while controlling the degree of geometric deformation within an acceptable range. A 30° tilt angle also effectively reduces common road surface lighting interference. If the tilt angle is too small (e.g., near-0° vertical photography), it will be difficult to acquire effective texture and geometric information of the road shoulder facade, resulting in insufficient boundary features. If the tilt angle is too large (e.g., exceeding 45°), it will significantly increase image perspective distortion, introducing errors into subsequent measurements and recognition. It may also exacerbate shadow occlusion or produce strong reflections due to excessively tilted light angles, interfering with image quality. A moderate tilt angle of 30° achieves a balance between acquiring sufficient lateral information and controlling image deformation and interference.

[0058] S102, in this embodiment of the invention, image processing is performed on the high-definition road surface images collected by the UAV. The processing includes an image data enhancement step, specifically enhancing the brightness, contrast, and HSV color space of the images; simultaneously, geometric transformation enhancements, including translation, flipping, and rotation, are performed to expand the dataset and improve the model's generalization ability. Furthermore, it includes steps for calibrating and synchronizing multi-source data to eliminate system errors in the equipment and acquisition process, ensuring the accuracy and consistency of the data in the spatiotemporal dimensions, and providing a reliable foundation for subsequent identification and analysis.

[0059] S103, In this embodiment, the image dataset of milled boundaries of the old shoulder pavement in highway reconstruction and expansion is constructed based on the YOLO dataset format specification. The dataset is optimized and divided into training set, validation set and test set in an 8:1:1 ratio.

[0060] Of the three sets of data, 80% is allocated to the training set to provide sufficient samples for the deep learning model, enabling it to fully learn the features of road surface structure layers, disease texture patterns, and boundary distribution rules, thereby optimizing its internal parameters. The 10% independent validation set is primarily used to implement early stopping strategies and perform hyperparameter tuning during training to optimize the model structure and prevent overfitting to the training data. The remaining 10% independent test set provides the model with a completely new, untrained data environment for final, objective performance evaluation, accurately reflecting the model's generalization ability and robustness in real-world engineering scenarios.

[0061] S104, image annotation involves identifying and selecting target signals in an image, determining the location and type of target signals at the milling boundary of the old road shoulder during highway reconstruction and expansion, reflecting them on the coordinates of a marker rectangle, and defining the category information based on the content of the marker. In this embodiment, LabelImg software is preferably used to manually identify and select target signals in the image, and the annotation method is as follows: Figure 3 .

[0062] S2, construct an improved YOLO11 object detection network. Train the improved YOLO11 object detection network using the dataset obtained in S1 to obtain the object recognition model. The model network structure is as follows: Figure 4 As shown;

[0063] S20101 introduces AirNet (All-Round Image Inpainting Network), which is applicable to various complex scenarios, into the backbone network. As an integrated inpainting network, AirNet's core function is to denoise, dehaze, and enhance the image before feature extraction. It ensures that the model can still clearly identify the boundary between the old shoulder and the main lane in cloudy weather, morning fog, or dusty road conditions, thus improving the system's environmental robustness.

[0064] AirNet consists of two neural modules: a contrast-based degradation encoder (CBDE) and a degradation-guided repair network (DGRN). It can recover damaged images even when the type and extent of damage are unknown. These two parts work together to handle various types of image degradation without prior knowledge of the specific damage. The network structure of AirNet is as follows: Figure 5 As shown, the input image x is extracted into the embedding space z by CBDE, and z and x are input into the DGGblock of the DGRN module for gradual optimization, and finally the prediction result is output.

[0065] CBDE is primarily responsible for learning damage representations from damaged images. It employs a contrastive learning approach, extracting key damage features by comparing different image damage patterns to support subsequent image restoration. It extracts a latent representation (labeled z) from the input damaged image, capturing the essential features of the image damage without prior knowledge (i.e., the specific damage type and extent). CBDE is trained using a contrastive learning framework, where the network learns damage representations by maximizing consistency between similar damage types (positive sample pairs) and minimizing consistency between different damage types (negative sample pairs). This allows AirNet to flexibly handle various unknown or unforeseen image damage scenarios.

[0066] DGRN uses degradation features learned from CBDE to guide the image restoration process. Its goal is to recover sharp images using these degradation features, even when the damage is unknown. DGRN aims to fully utilize the damage representation (z) extracted by CBDE to restore images based on this information. Some key components and functions of DGRN include:

[0067] Degradation Guided Module (DGM): Each DGM contains deformable convolutional (DCN) layers and spatial feature transform (SFT) layers. These modules dynamically adjust their processing methods based on the input degradation features z.

[0068] Degradation Guided Groups (DGGs): DGRNs consist of multiple DGGs, each containing several DGMs. This grouping structure helps to refine the recovery process in stages, with each group further improving image quality based on the previous one.

[0069] Feature fusion and recursive processing: In DGRN, the outputs of different levels can be recursively fed back into the network for further processing.

[0070] To remove damage from the observed image, the following objective function is proposed:

[0071]

[0072] in, It is the real label image Y t With the restored clean image Y p The reconstruction losses between them. The second loss. It is the contrast loss of CBDE.

[0073] For a given degraded image x, the value at the i-th pixel position is... , representing the real label image In the The pixel value at a given pixel location. The goal is to minimize the true label Y using AirNet. t With the restored clean image Y p The L1 distance between them is specifically defined as:

[0074]

[0075] yes The number of pixels, where i is the pixel index, output It is a clean image recovered from the network. At the same pixel position The predicted value.

[0076] S20102 incorporates BiFPN (Bidirectional Feature Pyramid Network) into the neck network. BiFPN is a highly efficient multi-scale feature fusion network that optimizes the traditional Feature Pyramid Network (FPN) by improving the trade-off between accuracy and efficiency through efficient bidirectional cross-scale connections and repetitive block structures. Milled boundaries are often slender and continuous geometric features. BiFPN can balance shallow location information (where is the boundary) and deep semantic information (what is the boundary), effectively solving the breakage problem in long-distance boundary recognition and improving localization accuracy.

[0077] Figure 6 This demonstrates the specific details of the BiFPN architecture. Specifically, the backbone network of the model uses EfficientNet to extract multi-scale primary features from the input image. The feature network employs a weighted bidirectional feature pyramid network (BiFPN), which receives feature maps of different scales from the backbone network (specifically, layers P2 to P7 in this embodiment) and deeply fuses and optimizes these multi-scale features through bidirectional (top-down and bottom-up) paths with learnable weights.

[0078] The fused features at different levels are specialized for identifying defects and structural targets of varying sizes: the BiFPN-optimized P2 and P3 level feature maps have high resolution and focus on capturing minute cracks, small defects, and high-precision milled boundary edge textures; the P4 and P5 level feature maps are used to identify medium-sized pavement structural layer regions and defect blocks; while the P6 and P7 level feature maps have a larger receptive field and are mainly responsible for locating large-scale milled areas, understanding the overall shoulder layout, and global contextual information. This multi-scale deep fusion structure is designed to maximize the complementary and synergistic effects of feature information at different levels with limited computational overhead, thereby comprehensively improving the model's detection accuracy and robustness for various pavement targets from minute to macroscopic.

[0079] BiFPN establishes bidirectional connections between top-down and bottom-up paths. Each bidirectional path (top-down and bottom-up) is treated as a separate feature network layer, and these layers can be repeated multiple times to facilitate higher-level feature fusion. This allows for more efficient flow and fusion of information between features at different scales, enabling effective information exchange across different scales. Cross-scale connectivity is optimized by removing nodes with only one input edge, adding extra edges between input and output nodes at the same level, and treating each bidirectional path as a feature network layer and repeating it multiple times.

[0080] BiFPN integrates bidirectional cross-scale connectivity and fast normalization fusion, as exemplified in this implementation description. Figure 6 The shown is the top-down intermediate feature of the 6th layer of BiFPN. and bottom-up output features Specifically:

[0081]

[0082]

[0083] in These are the original input features from layer 6. It is an intermediate feature of the 6th layer in the top-down path. It is the output feature of the 6th layer in the bottom-up path, and all other layer features are constructed in a similar way. (i=1, 2, 3, ...) is a learnable weight. It is another set of independent learnable parameters, which can be scalars (per feature), vectors (per channel), or multidimensional tensors (per pixel). , It is a very small positive constant used for numerical stability to prevent the denominator from being zero. This is a spatial resolution adjustment operation, typically using bilinear interpolation (upsampling) or max pooling / stretch convolution (downsampling) to unify feature maps of different scales to the same size. The original input features are for layer 7. This is the bottom-up output feature for the 5th layer.

[0084] S20103 replaces the detection head with an Improved Dynamic Head, utilizing an attention mechanism to unify object perception across three dimensions: scale perception, spatial perception, and task perception. This implementation replaces the original scale perception module's global average pooling with boundary-enhanced global pooling, making the pooling process more focused on the boundaries between pavement structural layers and crack edges. To address the linear extension of milled boundaries, a learnable offset using deformable convolution is employed. The system introduces boundary orientation offset constraints in global average pooling; it decouples the detection head into boundary classification and boundary regression branches, and learns hyperparameters independently for the two branches.

[0085] In highway scenarios, the shape of the road shoulder and the background (grass, guardrails, curbs) are highly variable. Scale perception adapts to changes in drone altitude; spatial perception helps the model focus on slender road shoulder segments rather than cluttered backgrounds; and task perception better distinguishes between "true milled boundaries" and "ordinary road surface cracks." The improved Dynamic Head primarily utilizes these three perception capabilities in a series combination to enhance detector performance. Figure 7 .

[0086] Input feature tensor , represents a three-dimensional tensor, where the first dimension is the number of layers, the second dimension is the spatial location, and the third dimension is the channel. Representation of characteristic tensor The real space to which it belongs and its shape dimensions. Its generalized attention. It can be described as:

[0087]

[0088] The above attention format is transformed into three sequential attentions, each focusing on only one dimension:

[0089]

[0090] in, , , These represent attention along three dimensions: level-wise, spatial-wise, and channel-wise.

[0091] Scale-aware (Level-wise): Detected targets have different scales, corresponding to feature maps of different scales. The scale-awareness of the detector is improved by altering the expressive power of different levels; each level refers to a feature map at a different layer within the neck.

[0092]

[0093] Where g(⋅) is a linear function, approximated by a 1*1 convolution. This is the hardsigmoid function. S represents the spatial dimension of the feature map, which is height × width (H × W), or the total number of pixels contained in the feature map. This indicates the number of channels in the feature map. The edge saliency mask (value range 0~1) is pre-computed by the Sobel operator, which makes the pooling process focus more on the interface of pavement structure layers and crack edge areas.

[0094] Spatial-wise: The target may appear at any location in the image. The spatial position perception capability of the detector can be improved by changing the expressive power of different spatial features corresponding to different locations.

[0095]

[0096] in, The number of sparsely sampled locations is a self-learning importance metric. It is a learnable offset of variable convolution. The hierarchical index of the feature pyramid. Represents learnable weight coefficients, Indicates the total number of layers in the feature pyramid. This represents the output of the spatial awareness attention mechanism. For each layer... and each sparse sampling point Take the corrected position from the feature map. eigenvalues ​​at , and These are the task-aware attention functions for the boundary classification branch and the boundary regression branch, respectively. It is the first The initial reference position coordinates of the sparse sampling points.

[0097] At the same time, sparse sampling points The arrangement was changed from a rectangular grid to a cross-shaped grid (sampling 3 points each along the horizontal, vertical, and diagonal directions, for a total of 12 points), making it more suitable for the boundary orientation of the two main horizontal and vertical directions in road scenes. The offset of the deformable convolution was adjusted. Decomposed into offsets along the boundary tangent direction and offset along the normal direction And impose differential constraints:

[0098]

[0099] Offset along the boundary tangent direction, allowing a relatively large range (constraint range ±3 pixels).

[0100] Offset along the boundary normal direction, limited to a small range (constraint range ±1 pixel).

[0101] The tangent direction allows for larger offsets to track continuous boundaries, while the normal direction restricts smaller offsets to maintain stable boundary widths. Furthermore, different feature layers independently learn their own offset parameters.

[0102] Task-aware (Channel-wise): Detected targets can have different task information representations, corresponding to different detection heads. The task-aware capability of the detector can be improved by changing the expressive power of different channels.

[0103]

[0104] The detection head is decoupled into a boundary classification branch and a boundary regression branch, and hyperparameters are learned independently for the two branches:

[0105]

[0106]

[0107] Classification branch: Use a steeper slope ( Initialized to 1.2. Initialized to 0.8) to enhance the distinction between positive and negative samples.

[0108] Regression branch: Use a gentler slope ( , ), and allows negative intercept ( It can be learned as a negative value to preserve subpixel-level offset information.

[0109] in, It is a superfunction used to learn and control the activation threshold. The slope of the first linear segment determines the input features. (No. When the feature values ​​of each channel pass through this activation function, the proportion that is preserved or amplified is... The intercept corresponding to the first linear segment determines the offset of the activation function, that is, how many thresholds the feature values ​​need to exceed to be effectively propagated. Its usage is similar to the Dynamic ReLU activation function, first in L Global pooling is performed on the dimension of , then two fully connected layers, one normalization layer, and finally . The function is normalized to output. This implementation obtains these parameters through global pooling and fully connected layers, allowing different channels and different input samples to have different activation functions.

[0110] In this embodiment of the invention, the dynamic head of the target detection network has been specifically improved to optimize its ability to perceive and locate the specific target, the milled boundary of the road surface.

[0111] Specific improvements include:

[0112] To address the characteristic that milled boundaries are located at the junctions of pavement structural layers and belong to strong edge regions in images, this invention replaces the global average pooling operation in the original scale-aware module with boundary-enhancing global pooling to strengthen the model's focus on edge pixels. This improvement allows the feature fusion process to focus more on the feature responses at the junctions of pavement structural layers and crack edges, thereby enhancing the model's scale awareness of key boundaries.

[0113] To address the typically linear extension of milled boundaries, the sampling method of the original deformable convolution is optimized. The standard rectangular grid sampling (e.g., 3×3) is replaced with a cross-shaped sampling along the horizontal, vertical, and two diagonal directions, totaling 12 sampling points. This sampling layout better conforms to the direction of the linear boundary, effectively focusing on boundary cues and reducing interference from irrelevant background areas. Simultaneously, the offset of the deformable convolution (…) It can be decomposed into two orthogonal components: the offset along the boundary tangent direction ( ) and offset along the boundary normal direction ( It also applies differentiated learning constraints to the offsets in these two directions, enabling the convolution kernel to adapt more flexibly and accurately to the direction and position of the boundary.

[0114] Considering that boundary recognition tasks require both classification (determining whether a boundary exists) and regression localization (determining precise location), and that these two tasks have different requirements for underlying feature responses, this invention decouples the detection head into independent boundary classification and boundary regression branches. Each branch is configured with independent learnable parameters, including independently learning the slope and intercept hyperparameters of a dynamic activation function for each branch. This decoupling design allows the classification and regression tasks to optimize features according to their respective objectives, thereby improving the overall task adaptability and accuracy of the detection head.

[0115] Step S3, Intelligent Recognition of Milled Boundary of Old Road Shoulder in Highway Reconstruction and Expansion: Input the image of the milled boundary of the old road shoulder in highway reconstruction and expansion to be identified into the target recognition model for recognition, and output the bounding box and confidence score of the milled boundary of the old road shoulder in highway reconstruction and expansion to complete the automatic recognition.

[0116] The milled boundary images of the old shoulder pavement from the highway reconstruction and expansion in step S1 are randomly divided into training, validation, and test sets in an 8:1:1 ratio; the corresponding training parameters are adjusted and configured, including setting the training iterations to 220 rounds, to better suit the detection task. Specifically:

[0117] In this embodiment, the experimental environment was a computer running Windows 11, with an Intel i5-13500H processor and 32GB of memory. The development tool was PyCharm, the development language was Python 3.11.9, and the deep learning framework was PyTorch 2.0.0. Specific experimental training parameters are shown in Table 1, and the training results of the object detection neural network model with attached detection boxes and classification probabilities are shown in [Table 1]. Figure 8 .

[0118] Table 1 Experimental parameters

[0119]

[0120] The dataset was divided into training, validation, and test sets in a ratio of 8:1:1 for experimentation, and compared with five commonly used object detection algorithms. The experimental results are shown in Table 2.

[0121] Table 2 Comparison Results of Comparative Experiments

[0122]

[0123] In terms of accuracy evaluation metrics, the model achieved the highest precision (P) and recall (R), with mAP@0.5 reaching 96.7%. Overall accuracy was the highest among all models, demonstrating high accuracy and minimizing misclassification of background as targets. It was able to accurately identify milled boundaries of old road shoulders during highway reconstruction and expansion, exhibiting the strongest overall detection and generalization capabilities. mF1 was 96.7%, the highest among all models, indicating extremely high recall while maintaining a good balance between accuracy and recall, detecting more real targets.

[0124] In terms of lightweight evaluation metrics, the improved YOLO11 model in this implementation has the lowest computational complexity, the fewest parameters, the highest detection efficiency, and the fastest detection speed. Its FLOPs are only 5.4*10^6. 6 The computation speed reached its maximum value, indicating that the improved YOLO11 algorithm significantly reduced the number of parameters and computational complexity while improving performance, thus significantly accelerating image inference speed.

[0125] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0126] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition, characterized in that, include: Capture road surface images, process and label them, and construct a dataset; An improved YOLO11 object detection network is constructed, and the improved YOLO11 object detection network is trained using the dataset to obtain an object recognition model; The image of the milled boundary of the old shoulder pavement of the highway reconstruction and expansion is input into the target recognition model for recognition, and the bounding box and confidence score of the milled boundary of the old shoulder pavement of the highway reconstruction and expansion are output. The improved YOLO11 target detection network introduces AirNet for image denoising, dehazing and enhancement into its backbone network, introduces Bi-directional feature pyramid network BiFPN into its neck network, and replaces its detection head with an improved dynamic head. The AirNet consists of a contrast base degradation encoder CBDE and a degradation guided repair network DGRN. The BiFPN performs deep fusion of feature maps of different scales from the backbone network through a bidirectional path with learnable weights, both top-down and bottom-up. The feature maps at different scales include levels P2 to P7; among them, the P2 and P3 level feature maps optimized by BiFPN are used to capture minute cracks, small-sized defects, and milled boundary edge textures; the P4 and P5 level feature maps are used to identify medium-sized pavement structure layer areas and defect blocks; and the P6 and P7 level feature maps are used to locate large-scale milled areas and understand the overall shoulder layout. The improved dynamic head utilizes an attention mechanism to unify target detection from three dimensions: scale awareness, spatial awareness, and task awareness. It combines the attention of these three dimensions in a concatenated manner, applying it to the input feature tensor. The processing satisfies generalized attention : ； in, , and These represent scale-aware attention, spatial-aware attention, and task-aware attention, respectively. The total number of layers in the feature pyramid. For spatial dimension size, For the number of channels, Representation of characteristic tensor The real number space to which it belongs and its shape dimensions; The scale-aware attention Defined as: ； in, It is a linear function approximated by a 1×1 convolution. For hard sigmoid functions, The edge saliency mask pre-computed by the Sobel operator makes the pooling process focus more on the interface between pavement structural layers and the edge region of cracks; spatial awareness attention Defined as: ； in, This represents the number of locations for sparse sampling. As a measure of the importance of self-learning, Learnable offsets for deformable convolutions. This is the hierarchical index of the feature pyramid. These are learnable weight coefficients. Indicates sparse sampling points. This represents the number of channels in the feature map. It is the first Initial reference position coordinates of sparse sampling points; Furthermore, the sparse sampling points are arranged in a cross-shaped pattern along the horizontal, vertical, and diagonal directions, totaling 12 sampling points; the learnable offset Decomposed into offsets along the boundary tangent direction and offset along the boundary normal direction And satisfy: ； Among them, for learnable offsets Decomposed into offsets along the boundary tangent direction and offset along the boundary normal direction Apply differentiated learning constraints; offset along the boundary tangent direction. The constraint range is ±3 pixels, and the offset is along the boundary normal direction. The constraint range is ±1 pixel; In the task perception attention In this dimension, the detection head is decoupled into a boundary classification branch and a boundary regression branch, and independent learnable hyperparameters are configured for each branch. The activation methods of the boundary classification branch and the boundary regression branch respectively satisfy: ； ； in, For the first The characteristic values ​​of each channel, , , , , , , , , All are achieved through hyperfunctions The learned hyperparameters, the hyperfunction In L Global pooling is performed along the dimension, and the hyperparameters are obtained through a fully connected layer and a normalization layer. For the input feature tensor, and These are the task-aware attention functions for the boundary classification branch and the boundary regression branch, respectively.

2. The method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition as described in claim 1, characterized in that, The captured road surface images include: A drone equipped with a multispectral sensor was used to collect road surface images. The flight parameters were set to a forward overlap rate of no less than 80%, a lateral overlap rate of no less than 60%, and a camera tilt angle of 30°.

3. The method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition as described in claim 1, characterized in that, Image processing includes: It enhances the brightness, contrast, and HSV color space of the image, and performs geometric transformations including translation, flipping, and rotation.

4. The method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition as described in claim 1, characterized in that, The overall loss function on which AirNet training is based satisfy: ； in, The contrast loss of the CBDE is... It is the real label image Y t With the restored clean image Y p The reconstruction loss between them, the reconstruction loss Defined as: ； In the formula, For the input image The total number of pixels, For pixel index, To recover a clean image At pixel position The predicted value, For the real label image Y t At pixel position The pixel value.

5. The method for determining the milling boundary of old road shoulder pavement in highway reconstruction and expansion based on image recognition as described in claim 1, characterized in that, The 6th layer of the BiFPN, top-down intermediate feature and bottom-up output features Defined by the following formulas respectively: ； ； In the formula, These are the original input features from layer 6. It is a set of learnable weights. It is another set of independent learnable parameters. For numerical stability, For spatial resolution adjustment operations, For depthwise separable convolution, The original input features are for layer 7. This is the bottom-up output feature for the 5th layer.

Images