An improved network model crack identification segmentation method for complex texture background

Abstract

The application discloses a kind of improved network model crack identification segmentation methods for complex texture background, its method includes: obtaining the image of road surface is input after pre-processing improved DeepLabv3+ network model, MobileNetV2 module obtains low-level feature F low And high-level feature F high , high-level feature F high It is extracted to feature F aspp By multi-scale feature extraction;LBP-ECA module is used for direction information and local texture information and is output high-level feature by gate fusion module, alignment matching, and the fusion feature is obtained by fusing high-level feature and low-level feature F low ; the fusion feature and high-level feature are concatenated and spliced in channel dimension and are decoded processing, classification convolution layer processing, up-sampling processing output feature map after crack identification segmentation.The application significantly improves the detection capability of small crack, broken crack and weak texture crack under complex road surface scene, realizes the significant improvement of crack segmentation precision and stability under the condition of complex road surface image and complex background interference factor.

Description

Technical Field

[0001] This invention relates to the fields of visual processing and intelligent detection of road defects, and in particular to an improved network model crack recognition and segmentation method for complex texture backgrounds. Background Technology

[0002] Road surface cracks are among the most common and easily spreadable defects in transportation infrastructure such as urban roads, industrial parks, and airport runways. Their early and accurate identification is crucial for driving safety and lifecycle maintenance. Traditional image processing methods based on threshold segmentation, edge detection, and morphological operations are simple to implement and have low overhead, but they are highly sensitive to complex backgrounds such as lighting changes, shadows, stains, and gravel, easily leading to false positives and false negatives, and have limited generalization ability. In recent years, deep learning semantic segmentation models (such as U-Net, PSPNet, and DeepLabv3+) and networks specifically built for crack detection, such as DeepCrack, have made technological progress in crack extraction tasks, leveraging multi-scale context and end-to-end learning to improve segmentation accuracy. However, in real-world road scenarios, gravel / sand and cracks are highly similar in grayscale and texture, and crack morphology is often elongated, discontinuous, and has blurred boundaries, causing existing models to still face problems such as loss of detail, insufficient connectivity, and decreased robustness. Furthermore, some high-precision models rely on deep stacking and multi-branch structures, resulting in a large number of parameters and computational costs, making them difficult to meet deployment requirements. Therefore, there is an urgent need for an improved network model crack identification and segmentation method that can simultaneously achieve accuracy, robustness, and lightweight design under complex interference conditions, in order to improve its usability and engineering feasibility in real inspection situations. Summary of the Invention

[0003] The purpose of this invention is to provide an improved network model crack identification and segmentation method for complex texture backgrounds. An improved DeepLabv3+ network model is constructed, which includes a MobileNetV2 module, a void space pyramid pooling module ASPP, an LBP-ECA module, and a multi-scale feature pyramid network module MFPN. This method significantly improves the crack segmentation accuracy and stability under complex road surface images and complex background interference factors, improves the detection rate of small cracks and the integrity of large-scale cracks, and the output crack segmentation results have clear boundaries and good connectivity, which has significant engineering and application value.

[0004] The objective of this invention is achieved through the following technical solution: An improved network model crack recognition and segmentation method for complex texture backgrounds, the method includes: S1. Construct an improved DeepLabv3+ network model that includes a MobileNetV2 module, a void space pyramid pooling module ASPP, an LBP-ECA module, and a multi-scale feature pyramid fusion module MFPN, and train it for crack recognition and segmentation using an image sample label dataset. The image sample label data includes road surface image samples and corresponding crack labels.

[0005] S2. After preprocessing, the acquired road surface image is input into the improved DeepLabv3+ network model. The MobileNetV2 module obtains the low-level features F. low and high-level characteristics F high The hollow space pyramid pooling module ASPP is used for high-level features F high Multi-scale feature extraction and fusion are performed to obtain context-enhanced feature F. aspp ; The S3 and LBP-ECA modules are used to extract feature F. aspp The directional information and local texture information are combined and then the high-level feature F is output through a gated fusion module. eca , to high-level features F eca Alignment matching to feature F low High-level features are obtained at high resolution. The Multi-Scale Feature Pyramid Fusion Module (MFPN) integrates high-level features. With low-level features F low Fusion yields fusion characteristics .

[0006] S4. Improved DeepLabv3+ network model to fuse features Characteristics of high-level personnel The channel dimension is concatenated and decoded to obtain the decoded features. The decoded features are then processed by a classification convolutional layer and upsampling to output the feature map after crack recognition and segmentation.

[0007] To better implement this invention, in method S2, the road surface image undergoes preprocessing including normalization and standardization; the MobileNetV2 module outputs low-level features F through depthwise separable convolution processing, inverted residual processing, and linear bottleneck processing. low and high-level characteristics F high .

[0008] Preferably, in method S2, the Aperture Spatial Pyramid Pooling (ASPP) module employs several multi-scale feature convolutional and pooling modules to perform multi-scale feature extraction and pooling processing, and then fuses them to obtain the context-enhanced feature F. aspp .

[0009] Preferably, the LBP-ECA module includes a directional information attention branch and a local texture information attention branch, and the directional information attention branch processing method is as follows: S301, Regarding feature F aspp Feature F is obtained by performing pooling in both the width and height directions. h With feature F w ; S302, Feature F h With feature F w The dimensionality is concatenated by the Concat+Conv2d feature fusion module, and then compressed features F are obtained by 1×1 convolution dimensionality reduction, batch normalization and non-linear activation function; S303. Decompose the compressed feature F into two directional features according to the spatial dimension. and Then, after processing with a 1×1 convolution and a sigmoid activation function, the horizontal attention weights A are generated. h Vertical attention weight A w ;Utilizing horizontal attention weight A h Vertical attention weight A w We obtain the attention-enhanced feature Y by weighting and fusing directional information.

[0010] Preferably, the local texture information attention branch processing method is as follows: A 1×1 convolutional attention-enhanced feature Y is used for channel mapping to obtain a center response feature map C, and a 3×3 convolutional kernel is used to extract neighborhood sampling point features N. i The binary encoding is obtained by comparing it with the center value of the center response feature map C. The binary results are weighted and combined to obtain normalized texture features. Then, through 1×1 convolution, batch normalization and ReLU activation, the texture enhancement features Y' are obtained through fusion enhancement.

[0011] Preferably, the control fusion module concatenates the attention enhancement feature Y and the texture enhancement feature Y' in the channel dimension, performs 1×1 convolution and sigmoid activation to generate pixel-wise and position-wise dynamic fusion weights, and uses the dynamic fusion weights to adaptively weightedly fuse the attention enhancement feature Y and the texture enhancement feature Y' to obtain the high-level feature F. eca .

[0012] Preferably, in method S3, the multi-scale feature pyramid fusion module MFPN uses high-level features Taking the low-level feature flow as input, the input features undergo 1×1 convolution and batch normalization to unify the number of channels in the two input feature streams to the same dimension. Then, multi-scale feature extraction and interactive fusion are performed, followed by channel-dimensional concatenation and 3×3 convolution to output a fused feature map. .

[0013] Preferably, in method S4, the decoded feature F dec After processing by a classification convolutional layer, pixel-level crack classification categories are obtained, and the crack classification category corresponds to whether the crack is a crack in the crack label. The original image size is restored by upsampling, and the feature map after crack identification and segmentation is obtained.

[0014] Preferably, in method S1, the road surface image samples in the image sample label dataset include crack-free road surface image samples and cracked road surface image samples. The cracks in the cracked road surface image samples include crack samples including weathering, construction joints, fracture boundaries, slender cracks, and network cracks. The crack labels include label information indicating whether it is a crack. The road surface image samples are enhanced by processes including random rotation, scaling, cropping, blurring, and color perturbation. The road surface image samples include complex background interference factors such as oil stains, dirt, and shadows. The road surface image samples are preprocessed before being input into the improved DeepLabv3+ network model.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention creatively constructs an improved DeepLabv3+ network model that includes a MobileNetV2 module, a void space pyramid pooling module ASPP, an LBP-ECA module and a multi-scale feature pyramid network module MFPN. This model significantly improves the accuracy and stability of crack segmentation in the face of complex road images and complex background interference factors, improves the detection rate of small cracks and the integrity of large-scale cracks, and the output crack segmentation results have clear boundaries and good connectivity, which has significant engineering and application value.

[0016] (2) The present invention uses six sets of typical crack images. The crack segmentation result of the improved network model crack recognition and segmentation method of the present invention is close to the crack sample label, realizing high-precision crack segmentation result processing, and significantly improving the detection capability of fine cracks, fracture cracks and weak texture cracks in complex road surface scenarios. In addition to crack detection and segmentation in road surface scenarios, the present invention can also be used for crack detection and segmentation processing in other scenarios (such as ground cracks in coal mining areas). Attached Figure Description

[0017] Figure 1 This is a flowchart of the improved network model crack identification and segmentation method of the present invention; Figure 2 This is a schematic diagram illustrating the structural principle of the improved DeepLabv3+ network model in the embodiment; Figure 3 for Figure 2A schematic diagram of the structural principle of the LBP-ECA module; Figure 4 This is a schematic diagram illustrating the structural principle of the Multi-Scale Feature Pyramid Fusion Module (MFPN) in the embodiment. Figure 5 This is a schematic diagram showing the comparison between crack sample labels and crack segmentation prediction results when using a validation set to verify the crack segmentation prediction results of the improved network model crack identification and segmentation method in the embodiment. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to embodiments: Example like Figure 1 As shown, an improved network model crack recognition and segmentation method for complex texture backgrounds is proposed, the method comprising: S1, such as Figure 2 As shown, an improved DeepLabv3+ network model was constructed, which includes a MobileNetV2 module, a void space pyramid pooling module ASPP, an LBP-ECA module, and a multi-scale feature pyramid fusion module MFPN. The model was trained for crack recognition and segmentation using an image sample label dataset, which includes road surface image samples and their corresponding crack labels. The road surface image samples in this invention's image sample label dataset include both crack-free and cracked road surface image samples. Road surface images or pictures are acquired using conventional molding equipment under natural light conditions. During acquisition, multiple directions, distances, and different elevation angles are ensured to create diverse samples. Road surface image samples can be uniformly scaled to a resolution of 512×512. Cracks in the cracked road surface image samples include weathering, construction joints, fracture boundaries, slender cracks, and network cracks. Crack labels include information indicating whether a crack exists, and labels are applied pixel-by-pixel (e.g., cracks are assigned an attribute of 1, while non-crack backgrounds are assigned an attribute of 0). LabelMe software is used as an example annotation tool, with annotation granularity based on the crack backbone and identifiable boundaries. To ensure consistency, a double-checking and sampling review process is implemented. The road surface image samples undergo enhancement processing including random rotation, scaling, cropping, blurring, and color perturbation (these enhancement strategies enhance the model's robustness to interference from gravel, shadows, oil stains, etc.). The road surface image samples comprise complex background interference factors, including oil stains, dirt, shadows, and (for example, gravel and sand), facilitating more accurate identification of non-crack and crack features. The road surface image samples are preprocessed before being input into the improved DeepLabv3+ network model.

[0019] In some embodiments, the improved DeepLabv3+ network model is built based on the PyTorch 2.5.1 deep learning framework. The operating system used for the improved DeepLabv3+ network model is Windows 10, and the graphics card is an NVIDIA GeForce RTX 4070 Super. The network framework of the DeepLabv3+ network model adopts DeepLabv3+, with MobileNetV2 as the backbone, and training and inference are completed in the same environment. Some configurations for training and data processing are listed in Table 1, covering input resolution, batch size and number of training epochs, optimizer and learning rate scheduling, loss function, normalization and data augmentation strategies, etc. To ensure stable results, a fixed random seed is used, and the best weights are saved based on the performance on the validation set during training for final testing.

[0020]

[0021] This invention improves the DeepLabv3+ network model by selecting four metrics—precision, recall, F1 score, and mean Intersection over Union (mIoU)—to evaluate the accuracy of the model's crack segmentation results. Precision represents the proportion of pixels correctly identified as "cracks"—a higher value indicates fewer false positives (misclassifying background as cracks) and more accurate extraction results. Recall represents the proportion of correctly detected real crack pixels—a higher value indicates fewer missed detections and more complete crack coverage. The F1 score considers both precision and recall; a higher value indicates better model extraction performance. mIoU calculates the intersection over union ratios of crack and background classes and averages them, comprehensively reflecting pixel-level region overlap; a higher value indicates better consistency between the segmentation results and the ground truth annotations. Both F1 and mIoU are comprehensive evaluation metrics for the network model.

[0022] ; ; ; ; The above indicators are all calculated from the four types of statistics of the confusion matrix: true positive (TP, crack pixels are correctly identified as cracks), true negative (TN, background pixels are correctly identified as background), false positive (FP, background pixels are misidentified as cracks), and false negative (FN, crack pixels are misidentified as background).

[0023] S2. Acquire road surface images (e.g., using conventional molding equipment, such as a camera; outdoor acquisition under natural light conditions using conventional molding equipment). After preprocessing, input the images into the improved DeepLabv3+ network model. Before inputting the road surface images into the improved DeepLabv3+ network model, the images undergo processing including size unification (e.g., unifying to...). Preprocessing includes resolution and size standardization to ensure consistency between training and inference inputs, and consistency with label size; normalization (e.g., scaling image pixel values ​​to the [0,1] range); and standardization (adjusting the data distribution to mean μ=0 and standard deviation σ=1 based on dataset statistics). Assuming the preprocessed road image X∈R... C×H×W (C is the number of channels, H is the height, and W is the depth). Road surface images are input into the MobileNetV2 module (i.e., the second generation of lightweight convolutional neural networks) to obtain low-level features F. low and high-level characteristics F high Preferably, the MobileNetV2 module outputs low-level features F after depthwise separable convolution processing, inverted residual processing, and linear bottleneck processing. low and high-level characteristics F high Inverted residual processing is used for high-dimensional feature extraction and corresponds to high-level feature F. high Linear bottleneck processing is used for low-dimensional feature extraction and corresponds to low-level feature F. low .

[0024] like Figure 2 As shown, the void space pyramid pooling module ASPP is used for high-level features F high Multi-scale feature extraction and fusion are performed to obtain context-enhanced feature F. aspp Specifically, the Spatial Pyramid Pooling Module (ASPP) employs several multi-scale feature convolutional modules (including 1×1 convolutional modules and multiple 3×3 convolutional modules) and pooling modules (in this embodiment, image pooling is used) to perform multi-scale feature extraction and pooling processing, and then fuses them to obtain the context-enhanced feature F. aspp .

[0025] The S3 and LBP-ECA modules are used to extract feature F. aspp The directional information and local texture information are combined and then the high-level feature F is output through a gated fusion module. eca .like Figure 3 As shown, the LBP-ECA module (a combination of Local Binary Pattern (LBP) and Efficient Channel Attention (ECA)) includes a directional information attention branch (corresponding to Efficient Channel Attention ECA) and a local texture information attention branch (corresponding to Local Binary Pattern (LBP), which generates binary codes and captures local texture features by comparing pixels with their neighbors), such as... Figure 3 As shown, the method for handling directional information attention branches is as follows: S301, Regarding feature F aspp Feature F is obtained by performing pooling in both the width and height directions. h With feature F w In some embodiments, feature F h XAvgPool (averaging along the width W) pooling is used to perform global averaging on each channel and each row, resulting in a one-dimensional description that retains only the height position; feature F w YAvgPool (averaging along the height H direction) pooling is used to perform global averaging on each channel and each column, resulting in a one-dimensional description that retains only the width position. Feature F h With feature F w The expression is as follows: ; Where X represents the input feature map; C is the number of feature channels; H and W are the height and width of the feature map, respectively; X(c,i,j) represents the feature value of the position X(c,i,j) at the Cth channel, the i-th row, and the j-th column; F h This represents the feature after global average pooling along the width direction, retaining information about the height position i. Therefore, its spatial size is H×1, and the overall dimension is C×H×1; F w This represents the feature after global average pooling along the height direction, retaining the information of the width position j. Therefore, its spatial size is 1×W, and the overall dimension is C×1×W. , This is a normalization factor to ensure that the average pooling does not change its numerical scale with changes in size.

[0026] S302, Feature F h With feature F w The dimensions are concatenated using the Concat+Conv2d feature fusion module (a combination of Concat and Conv2d convolutional modules), followed by 1×1 convolutional dimensionality reduction, batch normalization, and a non-linear activation function (the 1×1 convolutional dimensionality reduction, batch normalization, and non-linear activation function processes the corresponding...). Figure 3 The compressed feature F is obtained from the BatchNorm+Non-linear module. The expression for the compressed feature F is as follows: ; in, This indicates concatting by channel or specified dimension, used to merge statistical features in the horizontal and vertical directions. 1×1 convolution is used for channel mixing and dimensionality reduction; BN() is batch normalization used for stable training and bottleneck convergence. It is a linear activation function used to enhance expressive power.

[0027] S303, such as Figure 2 As shown, the compressed feature F is split into two directional features according to the spatial dimension. and (The compressed feature F is split into positional dependencies along the height direction and positional dependencies along the width direction), and then processed by a 1×1 convolution and a sigmoid activation function to generate horizontal attention weights A. h Vertical attention weight A w The expression is as follows: , , This is the Sigmoid activation function, used to suppress or enhance the feature response at different locations.

[0028] Using horizontal attention weights A h Vertical attention weight A w We obtain the attention-enhanced feature Y by weighting and fusing directional information (while preserving coordinate and directional information, we enhance the response of crack-related regions).

[0029] In some embodiments, such as Figure 3 As shown, the local texture information attention branch (to further enhance the ability to judge the local texture structure of cracks) processing method is as follows: A 1×1 convolutional attention-enhanced feature Y is used for channel mapping to obtain a center response feature map C. Feature map C has the same spatial size H×W as feature Y and is used as a center reference for subsequent neighborhood comparison. Neighborhood sampling point features N are extracted through 3×3 convolutional kernel sampling. i And compare it with the center value of the center response feature map C to obtain the binary code (corresponding to Figure 3 Binary comparison (in the model) weights and combines the binary results to obtain normalized texture features. (correspond Figure 3 The LBP Code+Normalization module (which is a combination of LBP encoding and normalization) is then used for fusion enhancement through 1×1 convolution, batch normalization, and ReLU activation to obtain the texture enhancement feature Y'. The expression for the texture enhancement feature Y' is as follows: [Y; LBP(Y) indicates splicing in the channel dimension, BN() is batch normalization; ReLU() is nonlinear activation.

[0030] The gated fusion module (which effectively fuses attention-enhanced feature Y and texture-enhanced feature Y') concatenates attention-enhanced feature Y and texture-enhanced feature Y' along the channel dimension, performs 1×1 convolution and sigmoid activation to generate pixel-wise and position-wise dynamic fusion weights, and uses these dynamic fusion weights to perform adaptive weighted fusion of attention-enhanced feature Y and texture-enhanced feature Y' (taking into account both orientation localization information and local texture discrimination information, improving the ability to perceive fine-grained crack structures) to obtain high-level feature F. eca Among them, dynamic fusion weights The expression is as follows: [Y;Y] indicates concatenation along the channel dimension, therefore its dimensions are 2C×H×W; Conv 1×1 () is a 1×1 convolution with 2 output channels, used to generate two gated weight maps; σ() is the Sigmoid activation function, making the gated weights range from (0,1); G∈R 2×H×W , where g1∈R 1×H×W g2∈R 1×H×W The fusion weights for attention-enhancing feature Y and texture-enhancing feature Y' are respectively assigned. High-level feature F eca The fusion expression is as follows: , where · represents element-wise multiplication, , By expanding the channel dimension to match the feature dimensions of C×H×W, adaptive weighted fusion of the two types of features at each spatial location is achieved.

[0031] High-level feature F eca Alignment and matching to feature F are achieved through bilinear interpolation upsampling. low High-level features are obtained at high resolution. The Multi-Scale Feature Pyramid Fusion Module (MFPN) integrates high-level features. With low-level features F low Fusion yields fusion characteristics In some embodiments, see Figure 4 The Multi-Scale Feature Pyramid Fusion Module (MFPN) uses high-level features With low-level features F eca As input, the number of channels in the two input feature paths is unified to the same dimension through 1×1 convolution and batch normalization, respectively. Then, multi-scale feature extraction and interactive fusion are performed, followed by channel-dimensional concatenation and 3×3 convolution to output a fused feature map. .

[0032] S4. Improved DeepLabv3+ network model to fuse features Characteristics of high-level personnel The decoded feature F is obtained by concatenating and decoding in the channel dimension. decThe cascaded concatenation expression is F cat =Concat( , `Concat()` represents concatenating features along the channel dimension, requiring both features to have the same spatial size H×W, and the number of channels is added after concatenation. For the decoded feature F... dec After classification convolutional layer processing and upsampling, the output is a feature map for crack identification and segmentation. Specifically, the decoded feature F... dec After processing with a classification convolutional layer, pixel-level crack classification categories are obtained, and each crack classification category corresponds to a crack label indicating whether it is a crack. Upsampling is then used to restore the original image size, resulting in a feature map of the identified and segmented cracks (i.e., the crack segmentation result corresponding to the road surface image). This invention effectively improves the detection capability of fine cracks, fracture cracks, and weakly textured cracks in complex road surface scenarios.

[0033] This invention improves the DeepLabv3+ network model, significantly enhancing crack segmentation accuracy and stability in complex interference scenarios (gravel, stains, shadows, occlusion / fractures, etc.): On a self-built road surface dataset, it achieves F1=94.08% and mIoU=89.21%, representing improvements of approximately 0.64 and 1.06 percentage points respectively compared to the control scheme based on DeepLabv3+. Through a feature enhancement mechanism of "coordinate attention direction encoding + learnable LBP texture modeling + gated fusion," it effectively suppresses false detections / false negatives caused by cracks having "same grayscale and similar texture," resulting in clearer boundaries and better connectivity. Multi-level feature pyramid fusion (MFPN) achieves seamless integration of shallow edges / textures with cracks. The method achieves deep semantic alignment and fusion, improving both the detection rate of small cracks and the integrity of large-scale cracks. Using MobileNetV2 as the backbone and configuring OS=16 and dilated convolutions, it reduces the number of parameters and computational load while maintaining accuracy. It boasts high-efficiency inference advantages with low latency and low memory usage, facilitating real-time / near real-time deployment on vehicle-mounted, handheld, and edge devices. The method is end-to-end trainable, easy to integrate, and can be smoothly migrated to similar crack scenarios such as walls, bridges, runways, and mining areas. It can also interface with upper-layer applications such as crack width estimation and severity classification, comprehensively reducing the cost of false alarms / missed alarms and manual verification during inspections, improving unit computing power output and operational efficiency, and demonstrating significant engineering and application value.

[0034] like Figure 5 As shown, this invention uses six sets of typical crack images to demonstrate the prediction performance of an improved DeepLabv3+ network model. Each set of images, from top to bottom, represents the original image, manually labeled image, and the segmentation result from the improved DeepLabv3+ network model. Numbered from left to right as Figures a to g, they correspond to different types of crack samples, covering various scenarios such as thin cracks, blurred edges, cracks with varying thickness, complex structures, and high-density mesh cracks. Figure 5 As shown in Figures a and b, the cracks are coarse in shape with somewhat blurred edges. In this type of crack, other models (such as PSPNet and DeepCrack models) exhibit significant oversegmentation, with excessive expansion of the crack region and numerous pseudo-crack areas. For example, U-Net shows a break in the middle of the crack, making it difficult to maintain the crack's continuity. In contrast, the improved DeepLabv3+ network model of this invention performs relatively stably and can identify the main crack region. Compared to this, the method of this invention (a road crack identification and segmentation method based on the improved DeepLabv3+ network model) can accurately identify the main crack structure, with smooth edges, and the segmentation results are highly consistent with the actual annotations. Figures c and d illustrate typical scenarios of narrow, elongated cracks. These cracks, with their delicate structures, place higher demands on the model's edge perception and connectivity preservation capabilities. The method of this invention (a road surface crack recognition and segmentation method based on an improved DeepLabv3+ network model) effectively maintains the integrity and detailed connectivity of the crack structure in both types of images, demonstrating stronger feature extraction and edge preservation capabilities. Compared to other models, such as PSPNet, which has limited recognition capabilities in these scenarios and suffers from severe missed detections, and other models like U-Net and DeepCrack, while capable of detecting some structures, exhibit issues with breakage, expansion, or noise artifacts, Figure e shows a slightly curved, medium-width crack. The method of this invention (a road surface crack recognition and segmentation method based on an improved DeepLabv3+ network model) performs best in extracting the crack's main trunk and edge details, producing clear segmentation results with natural contours. Other models, such as U-Net and DeepLabv3+, suffer from offset and blurred edges at crack corners, while DeepCrack exhibits a degree of oversegmentation. Figures f and g show densely packed, complex network cracks, posing a challenge to the model's global perception capability and robustness in complex backgrounds. The method of this invention demonstrates good connectivity and boundary segmentation capabilities in the crack network structure, and is superior in detail recognition and noise suppression, closely matching the real labels and exhibiting stronger generalization ability. Other models, such as PSPNet, U-Net, and DeepLabv3+, perform poorly in these scenarios, often exhibiting poor crack connectivity, structural fragmentation, or misidentification. For example, while DeepCrack can extract some main regions, it suffers from "connected patches" in dense areas, resulting in numerous misclassified regions. In summary, the improved network model crack identification and segmentation method of this invention uses a validation set to verify the crack segmentation prediction results. Six sets of typical crack images are used to compare the crack sample labels and crack segmentation prediction results of the validation set. The crack segmentation results of the improved network model crack identification and segmentation method of this invention are very close to the crack sample labels, achieving high-precision crack segmentation result processing.

[0035] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An improved network model crack identification and segmentation method for complex road surfaces, characterized in that: The methods include: S1. Construct an improved DeepLabv3+ network model that includes a MobileNetV2 module, a void space pyramid pooling module ASPP, an LBP-ECA module, and a multi-scale feature pyramid fusion module MFPN, and train it for crack recognition and segmentation using an image sample label dataset. The image sample label data includes road surface image samples and corresponding crack labels. S2. After preprocessing, the acquired road surface image is input into the improved DeepLabv3+ network model. The MobileNetV2 module obtains the low-level features F. low and high-level characteristics F high The hollow space pyramid pooling module ASPP is used for high-level features F high Multi-scale feature extraction and fusion are performed to obtain context-enhanced feature F. aspp ; The S3 and LBP-ECA modules are used to extract feature F. aspp The directional information and local texture information are combined and then the high-level feature F is output through a gated fusion module. eca , to high-level features F eca Alignment matching to feature F low High-level features are obtained at high resolution. The Multi-Scale Feature Pyramid Fusion Module (MFPN) integrates high-level features. With low-level features F low Fusion yields fusion characteristics ; S4. Improved DeepLabv3+ network model to fuse features Characteristics of high-level personnel The channel dimension is concatenated and decoded to obtain the decoded features. The decoded features are then processed by a classification convolutional layer and upsampling to output the feature map after crack recognition and segmentation.

2. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 1, characterized in that: In method S2, the road surface image undergoes preprocessing including normalization and standardization; the MobileNetV2 module outputs low-level features F through depthwise separable convolution, inverted residual processing, and linear bottleneck processing. low and high-level characteristics F high .

3. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 1, characterized in that: In method S2, the Spatial Pyramid Pooling (ASPP) module employs several multi-scale feature convolutional and pooling modules to perform multi-scale feature extraction and pooling, and then fuses them to obtain the context-enhanced feature F. aspp .

4. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 1, characterized in that: The LBP-ECA module includes a directional information attention branch and a local texture information attention branch. The directional information attention branch processing method is as follows: S301, Regarding feature F aspp Feature F is obtained by performing pooling in both the width and height directions. h With feature F w ; S302, Feature F h With feature F w The dimensionality is concatenated by the Concat+Conv2d feature fusion module, and then compressed features F are obtained by 1×1 convolution dimensionality reduction, batch normalization and non-linear activation function; S303. Decompose the compressed feature F into two directional features according to the spatial dimension. and Then, after processing with a 1×1 convolution and a sigmoid activation function, the horizontal attention weights A are generated. h Vertical attention weight A w ;Utilizing horizontal attention weight A h Vertical attention weight A w We obtain the attention-enhanced feature Y by weighting and fusing directional information.

5. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 4, characterized in that: The local texture information attention branch processing method is as follows: A 1×1 convolutional attention-enhanced feature Y is used for channel mapping to obtain a center response feature map C, and a 3×3 convolutional kernel is used to extract neighborhood sampling point features N. i The binary encoding is obtained by comparing it with the center value of the center response feature map C. The binary results are weighted and combined to obtain normalized texture features. Then, through 1×1 convolution, batch normalization and ReLU activation, the texture enhancement features Y' are obtained through fusion enhancement.

6. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 5, characterized in that: The control fusion module concatenates the attention enhancement feature Y and the texture enhancement feature Y' in the channel dimension, performs 1×1 convolution and Sigmoid activation to generate pixel-wise and position-wise dynamic fusion weights, and uses these dynamic fusion weights to adaptively weight and fuse the attention enhancement feature Y and the texture enhancement feature Y' to obtain the high-level feature F. eca .

7. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 1, characterized in that: In method S3, the Multi-Scale Feature Pyramid Fusion Module (MFPN) uses high-level features Taking the low-level feature flow as input, the input features undergo 1×1 convolution and batch normalization to unify the number of channels in the two input feature streams to the same dimension. Then, multi-scale feature extraction and interactive fusion are performed, followed by channel-dimensional concatenation and 3×3 convolution to output a fused feature map. .

8. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 1, characterized in that: In method S4, the decoding feature F dec After processing by a classification convolutional layer, pixel-level crack classification categories are obtained, and the crack classification category corresponds to whether the crack is a crack in the crack label. The original image size is restored by upsampling, and the feature map after crack identification and segmentation is obtained.

9. The improved network model crack recognition and segmentation method for complex texture backgrounds according to claim 1, characterized in that: In method S1, the road surface image samples in the image sample label dataset include crack-free road surface image samples and cracked road surface image samples. The cracks in the cracked road surface image samples include crack samples including weathering, construction joints, fracture boundaries, slender cracks and network cracks. The crack labels include label information of whether it is a crack. The road surface image samples are enhanced by processes including random rotation, scaling, cropping, blurring and color perturbation. The road surface image samples include complex background interference factors such as oil stains, dirt and shadows. Road surface image samples are preprocessed before being input into the improved DeepLabv3+ network model.

Images