A concrete pavement crack detection method based on progressive context interaction

Abstract

The application belongs to the technical field of image processing, and particularly relates to a concrete pavement crack detection method based on progressive context interaction; the method comprises the following steps: obtaining a road surface image to be processed, inputting the road surface image into a trained concrete pavement crack detection model for processing, and obtaining a concrete pavement crack detection result; the concrete pavement crack detection model adopts a progressive encoding-decoding structure, and the model is divided into three stages, namely a feature extraction stage, a feature enhancement stage and a feature fusion stage; the feature extraction stage comprises an ACRB module and a progressive context interaction mechanism; the feature enhancement stage comprises a triple attention module and a vertical attention and horizontal attention hybrid module; and the feature fusion stage comprises a convolution layer and a MADFM module; the application improves the crack detection accuracy, and has excellent robustness and good application prospect.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for detecting cracks in concrete pavement based on progressive contextual interaction. Background Technology

[0002] In recent years, the rapid increase in the number of concrete roads has highlighted the importance of crack detection. Under the influence of foundation settlement and changes in temperature and pressure, concrete road surfaces are extremely prone to cracking, resulting in various types of fissures. The appearance of cracks affects the strength, durability, and usability of the concrete structure, leading to road damage. Therefore, timely detection and repair of cracks can prevent small cracks from gradually developing into larger, more difficult-to-repair cracks, potholes, and other more complex types of road surface damage, thereby effectively extending the service life of roads, reducing maintenance costs, mitigating driving risks, and maintaining road smoothness. However, traditional manual detection methods are time-consuming and easily affected by environmental noise, making it difficult to effectively identify cracks. At the same time, due to the diversity, anisotropy, topological complexity, low contrast, and noisy texture backgrounds of cracks, coupled with current algorithms still facing problems such as imprecise segmentation and discontinuous crack detection when dealing with thin cracks and low-contrast backgrounds, detecting road surface cracks from images remains a challenge.

[0003] In conclusion, it is necessary to deepen the research on crack detection technology in order to overcome the limitations of existing algorithms and improve the accuracy of crack detection. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a concrete pavement crack detection method based on progressive contextual interaction. The method includes: acquiring a pavement image to be processed, inputting it into a trained concrete pavement crack detection model for processing, and obtaining the concrete pavement crack detection result. The concrete pavement crack detection model adopts a progressive encoder-decoder structure, and the model is divided into three stages: feature extraction, feature enhancement, and feature fusion. The feature extraction stage includes an ACRB module and a progressive contextual interaction mechanism; the feature enhancement stage includes a triple attention module and a hybrid vertical and horizontal attention module; the feature fusion stage includes convolutional layers and a MADFM module.

[0005] Preferably, the data processing steps in the feature extraction stage include:

[0006] The road surface image is processed by the ACRB module and downsampling to obtain feature A; feature A is then processed by the ACRB module to obtain feature B; feature B is upsampling and then processed with the first shallow side output through a progressive context interaction mechanism to obtain the second shallow side input; the second shallow side input is processed by the ACRB module to obtain the second shallow side output.

[0007] Feature B and the first shallow side output, after being downsampled, are then processed by a progressive context interaction mechanism to obtain feature C; feature C is then processed by the ACRB module to obtain feature D; feature D is upsampled and then processed by the second shallow side output through a progressive context interaction mechanism to obtain the third shallow side input; the third shallow side input is processed by the ACRB module to obtain the third shallow side output.

[0008] Feature B is downsampled to obtain feature I; feature I is processed by the ACRB module to obtain feature J; the downsampled features of the second shallow side output, feature D, and feature J are upsampled and then processed by a progressive context interaction mechanism to obtain feature E; feature E is processed by the ACRB module to obtain feature F; feature F is upsampled and then processed by the progressive context interaction mechanism with the third shallow side output to obtain the fourth shallow side input; the fourth shallow side input is processed by the ACRB module to obtain the fourth shallow side output.

[0009] Feature D is downsampled and then processed with feature J through a progressive context interaction mechanism to obtain feature K; feature K is processed by the ACRB module to obtain feature L; the downsampled features of the third shallow side output, feature F, and feature L are upsampled and then processed by a progressive context interaction mechanism to obtain feature G; feature G is processed by the ACRB module to obtain feature H; feature H is upsampled and then processed with the fourth shallow side output through a progressive context interaction mechanism to obtain the fifth shallow side input; the fifth shallow side input is processed by the ACRB module to obtain the fifth shallow side output.

[0010] Furthermore, the ACRB module includes 3×3 convolutions, ACBlock blocks, batch normalization layers, LeakyReLU activation functions, and residual connections.

[0011] Furthermore, the data processing steps of the progressive contextual interaction mechanism include:

[0012] The input features are summed to obtain a fused feature map;

[0013] The fused feature map is non-linearly transformed by the GeLU activation function and then passed through a 3×3 convolutional layer and a GroupNorm layer to obtain the first intermediate feature.

[0014] The first intermediate feature is passed through another set of 3×3 convolutions and GroupNorm normalization layers, and finally a mask is generated by the Sigmoid activation function.

[0015] After multiplying the fused feature map element-wise with the mask, the result is then passed through a 1×1 convolution, a 3×3 convolution, and a 3×3 convolution with a dilation rate of 2. The three convolution results are summed, and finally passed through a batch normalization layer to complete the progressive context interaction mechanism.

[0016] Preferably, the data processing steps in the feature enhancement stage include:

[0017] The first shallow-layer output is processed by a triple attention module to obtain the first middle-layer output;

[0018] The second shallow layer output is processed by a hybrid module of vertical and horizontal attention to obtain the second middle layer output;

[0019] The third shallow layer output is processed by a hybrid module of vertical attention and horizontal attention to obtain the third middle layer output;

[0020] The fourth shallow-layer output is processed by a hybrid module of vertical and horizontal attention to obtain the fourth middle-layer output;

[0021] The fifth shallow-layer output is processed by a triple attention module to obtain the fifth middle-layer output.

[0022] Preferably, the data processing in the feature fusion stage includes: the first intermediate layer output is processed by 1×1 convolution to obtain the first deep layer output; the second intermediate layer output is processed by 1×1 convolution to obtain the second deep layer output; the first to third intermediate layer outputs are processed by the MADFM module to obtain the third deep layer output; the fourth intermediate layer output is processed by 1×1 convolution to obtain the fourth deep layer output; the fifth intermediate layer output is processed by 1×1 convolution to obtain the fifth deep layer output; the first to fifth deep layer outputs are spliced ​​together and then processed by 1×1 convolution to obtain the concrete pavement crack detection result.

[0023] Furthermore, the data processing procedure of the MADFM module is represented as follows:

[0024] X concat =Concat(X) side1 ,X side2 ,X side3 )

[0025] X height =AAP height (X concat )

[0026] X width =AAP width (X concat )

[0027] X global=HSwish(BN(C 1×1 (Concat(X height ,X width ))))

[0028] X global_height ,X global_width =Split(X) global )

[0029] A height =Sigmoid(C 1×1 (X global_height ))

[0030] A width =sigmoid(C 1×1 (X global_width ))

[0031] W dynamic =FC(GAP(X) concat ))

[0032] A′ height =W dynamic ·A height

[0033] A′ width =W dynamic ·A width

[0034] Y = SiLU(BN(C) 3×3 (X concat ·A′ height +X concat ·A′ width )))

[0035] Among them, X concat Indicates splicing features, X side1 ,X side2 ,X side3 These represent the first, second, and third middle-layer outputs, respectively; Concat indicates the concatenation operation; AAP height and AAP width These represent adaptive average pooling in the height and width directions, respectively; X height and X width C represents the height and width pooling features, respectively; 1×1 Represents a 1×1 convolution, C 3×3 This represents a 3×3 convolution, BN represents a batch normalized layer, HSwish represents the HSwish activation function, and X represents a 3×3 convolution. global Represents global features; Split represents a segmentation operation; X global_height and X global_widthThese represent the global features of height and width, respectively; Sigmoid represents the Sigmoid activation function; A height and A width These represent the height and width attention weights, respectively; GAP represents global average pooling; FC represents a fully connected layer; and W... dynamic Indicates dynamic weights; A′ height and A′ width represents the adjusted height and width attention weights, respectively; SiLU represents the SiLU activation function; and Y represents the output feature of the MADFM module.

[0036] Preferably, the loss function in the training process of the concrete pavement crack detection model is the sum of the total side output loss and the fusion output loss.

[0037] Furthermore, the total output loss is expressed as:

[0038]

[0039]

[0040] Among them, L side (I,G,W,w) represents the total loss of the side output, I represents the total number of pixels in the original input image, G represents the total number of pixels in the label image, W represents the total parameters of the model, w represents the weights of each side output layer, N represents the number of side output layers, and α n This represents the loss weights of the nth side output layer. G represents the side output loss of the nth side output layer, w0 and w1 represent the weights of the cracked and non-cracked pixels, respectively. + With G _ P represents the sets of cracked pixels and non-cracked pixels, respectively; j G represents the prediction result for the j-th pixel. j This represents the actual label of the pixel; w (n) This represents the parameters of the nth side output layer.

[0041] Furthermore, the fusion output loss is expressed as:

[0042]

[0043] Among them, L fuse (I,G,W) represents the fusion output loss, where I represents the total pixels of the original input image, G represents the total pixels of the label image, W represents the total parameters of the model, N represents the number of side output layers, and w0 and w1 represent the weights of the cracked and non-cracked pixels, respectively. + With G _ P represents the sets of cracked pixels and non-cracked pixels, respectively; j G represents the prediction result for the j-th pixel.j This indicates the actual label of the pixel.

[0044] The beneficial effects of this invention are as follows: This invention proposes a concrete pavement crack detection method based on progressive contextual interaction and attention mechanisms. A concrete crack detection network based on progressive contextual interaction and attention mechanisms is proposed. First, an optimized UNet++ backbone is used, employing asymmetric convolutional blocks to enhance feature extraction capabilities. Second, a progressive contextual interaction mechanism is introduced to efficiently capture and fuse multi-scale features from adjacent stages. Then, in the feature enhancement stage, a triple attention module is combined, using attention combination to improve feature representation capabilities. Finally, in the feature fusion stage, a multi-semantic attention dynamic fusion module is used to enhance detail recovery and preservation. This invention improves crack detection accuracy while exhibiting excellent robustness and promising application prospects. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the concrete pavement crack detection model structure in this invention;

[0046] Figure 2 This is a structural diagram of the progressive context interaction mechanism in this invention;

[0047] Figure 3 This is a structural diagram of the multi-semantic attention dynamic fusion module in this invention;

[0048] Figure 4 This is a comparison diagram of the present invention and the comparative method. Detailed Implementation

[0049] 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.

[0050] This invention proposes a method for detecting cracks in concrete pavements based on progressive contextual interaction, the method comprising:

[0051] The road surface image to be processed is obtained and input into the trained concrete road surface crack detection model for processing to obtain the concrete road surface crack detection result.

[0052] like Figure 1As shown, the concrete pavement crack detection model designed in this invention adopts a progressive encoder-decoder structure, where the decoding input of each stage is the output of its corresponding encoder layer. It can be mainly divided into four parts: (1) The optimized UNet++ network is used as the backbone to achieve efficient feature extraction. (2) Different types of attention modules are embedded in the five side outputs of the first stage of the backbone network to improve the ability to capture the complex structural features of cracks. (3) The shallow side output adopts a multi-semantic attention dynamic fusion module to enhance and reconstruct the features, and better preserve the crack details. (4) The five feature maps are fused to obtain the final crack prediction map. In the feature extraction stage, asymmetric convolutional blocks are integrated into the original 3×3 convolutional blocks to form a new convolutional unit, namely ACRB, which is composed of 3×3 convolution, ACBlock block, batch normalization, LeakyReLU activation function and corresponding residual connection. At the same time, a progressive context interaction mechanism (PCIM) is introduced to continuously integrate high-level context information. In the feature enhancement stage, an attention combination (AC) approach is used, employing a triple attention module (TA) consisting of vertical attention (CWA, CWA), horizontal attention (HCAttention, HCA), and spatial attention (HWA, HWA). Vertical and horizontal attention are used on the first and fifth side outputs to capture vertical and horizontal features, preserving details and enhancing semantic information while reducing computational complexity. Spatial attention is added to the middle side output to comprehensively capture complex spatial features, ensuring accurate identification of irregular structures like cracks. This combination approach maximizes segmentation performance while maintaining computational efficiency. In the feature fusion stage, to help the network better recover and preserve crack details, a multi-semantic attention dynamic fusion module (MADFM) is proposed, fully utilizing features from shallow and intermediate layers for detail enhancement. Finally, the outputs from each side are concatenated, and the number of channels is adjusted using a 1×1 convolution to generate the final prediction map.

[0053] The concrete pavement crack detection model designed in this invention specifically includes the following structure:

[0054] The first stage of the network maintains the original resolution while continuously integrating higher-level contextual information. Unlike traditional encoder-decoder structures, the encoding and decoding processes in this invention are performed in parallel, rather than encoding first and then decoding. Specifically, the network uses a progressive contextual interaction mechanism to perform gradual feature fusion during feature extraction and feature recovery.

[0055] Under this mechanism, the shallow layers (lower resolution stages) of the network are mainly responsible for extracting fine-grained semantic information, while the deeper layers (higher resolution stages) provide richer contextual information and expand the local receptive field of the feature maps. Unlike traditional architectures that compress the features extracted by the encoder and then restore them through the decoder, the network avoids the problem of spatial information loss by directly fusing feature maps from adjacent stages (including downsampled feature maps from the encoding stage and upsampled feature maps from the decoding stage).

[0056] The road surface image is processed by the ACRB module and downsampling to obtain feature A; feature A is then processed by the ACRB module to obtain feature B; feature B is upsampling and then processed with the first shallow side output through a progressive context interaction mechanism to obtain the second shallow side input; the second shallow side input is processed by the ACRB module to obtain the second shallow side output.

[0057] Feature B and the first shallow side output, after being downsampled, are then processed by a progressive context interaction mechanism to obtain feature C; feature C is then processed by the ACRB module to obtain feature D; feature D is upsampled and then processed by the second shallow side output through a progressive context interaction mechanism to obtain the third shallow side input; the third shallow side input is processed by the ACRB module to obtain the third shallow side output.

[0058] Feature B is downsampled to obtain feature I; feature I is processed by the ACRB module to obtain feature J; the downsampled features of the second shallow side output, feature D, and feature J are upsampled and then processed by a progressive context interaction mechanism to obtain feature E; feature E is processed by the ACRB module to obtain feature F; feature F is upsampled and then processed by the progressive context interaction mechanism with the third shallow side output to obtain the fourth shallow side input; the fourth shallow side input is processed by the ACRB module to obtain the fourth shallow side output.

[0059] Feature D is downsampled and then processed with feature J through a progressive context interaction mechanism to obtain feature K; feature K is processed by the ACRB module to obtain feature L; the downsampled features of the third shallow side output, feature F, and feature L are upsampled and then processed by a progressive context interaction mechanism to obtain feature G; feature G is processed by the ACRB module to obtain feature H; feature H is upsampled and then processed with the fourth shallow side output through a progressive context interaction mechanism to obtain the fifth shallow side input; the fifth shallow side input is processed by the ACRB module to obtain the fifth shallow side output.

[0060] Progressive Contextual Interaction Mechanism (PCIM):

[0061] To address the issue of spatial information loss caused by multiple downsampling operations, inspired by the high-resolution network HRNet, this invention removes the dense connection operations at each stage in the UNet++ network and replaces them with a progressive context interaction mechanism designed in this paper. Unlike HRNet, this proposed mechanism does not employ fully parallel connections but only fuses feature maps from adjacent stages, thereby reducing feature redundancy and improving operational efficiency. This mechanism progressively fuses contextual information by gradually expanding in both the horizontal and vertical directions through upsampling, downsampling, and convolution.

[0062] like Figure 2 As shown, the progressive contextual interaction mechanism first sums the feature map from the previous residual block with the upsampled and / or downsampled feature maps from adjacent stages, fusing them into a single composite feature map. Then, the fused feature map undergoes a non-linear transformation using the GeLU activation function to enhance its expressive power. Next, the feature map passes through a 3×3 convolutional layer to change the number of channels, and the convolutional feature map is then normalized using GroupNorm. Subsequently, the processed feature map passes through another set of 3×3 convolutional and GroupNorm normalization layers, and finally, a mask is generated using the Sigmoid activation function. This mask is used to weight the feature map, highlighting important regions and suppressing secondary information, thus implementing a local attention mechanism. Next, the previously fused integrated feature map is multiplied element-wise with the mask, and then passed through a 1×1 convolution, a 3×3 convolution, and a 3×3 convolution with a dilation rate of 2 to further enhance the expressive power of the features and the fusion of contextual information. Finally, the three convolution results are summed and then passed through a batch normalization layer to complete the progressive contextual interaction mechanism.

[0063] Through a progressive contextual interaction mechanism, multi-resolution feature information is progressively enhanced and fused. While preserving detailed information, the receptive field is expanded to capture a wider range of contextual information, ultimately outputting richer and more expressive feature maps. Assuming the three inputs are X1, X2, and X3, which are features from different network layers, this process can be represented by the following formula:

[0064] X fused =U(X1)+ACB(X2)+D(X3)

[0065] A = Sigmoid(G(C) 3×3 (G(C 3×3 (X fused )))))

[0066]

[0067] Where A, G, and C represent attention weights, GeLU, and convolution, respectively.

[0068] The decoder structure of the concrete pavement crack detection model designed in this invention specifically includes:

[0069] The output of the first decoding layer is processed by a triple attention module to obtain the first middle layer output; the first middle layer output is then processed by a 1×1 convolution to obtain the first deep layer output.

[0070] The output of the second decoding layer is processed by a hybrid vertical attention and horizontal attention module to obtain the second middle layer output; the second middle layer output is then processed by a 1×1 convolution to obtain the second deep layer output.

[0071] The output of the third decoding layer is processed by a hybrid vertical attention and horizontal attention module to obtain the third middle layer output; the outputs of the first to third middle layers are processed by the MADFM module to obtain the third deep layer output.

[0072] The output of the fourth decoding layer is processed by a hybrid vertical attention and horizontal attention module to obtain the fourth middle layer output; the fourth middle layer output is then processed by a 1×1 convolution to obtain the fourth deep layer output.

[0073] The output of the fifth decoding layer is processed by a triple attention module to obtain the fifth middle layer output; the fifth middle layer output is processed by a 1×1 convolution to obtain the fifth deep layer output.

[0074] The output features of the decoder are obtained by concatenating the outputs of the first to fifth deep layers and then performing a 1×1 convolution.

[0075] Semantic Attention Dynamic Fusion Module (MADFM):

[0076] To address the issues of feature redundancy and semantic ambiguity in the middle-layer output side3, this invention proposes a multi-semantic attention dynamic fusion module, the structure of which is as follows: Figure 3 As shown, the core idea of ​​this module is multi-semantic feature fusion and dynamic feature enhancement. Specifically, since the first three side outputs come from different network layers, they contain multi-level feature information. Shallower side outputs often contain more details and edge information, while deeper side outputs contain more semantic information. By fusing these feature maps, information at different scales can be integrated, ensuring that the model can capture details while understanding the global context. This module generates a comprehensive feature representation by fusing these features. This unified feature representation can more comprehensively describe the complex morphology of the crack. At the same time, this dynamic weighting mechanism allows the model to pay more attention to the key feature regions of the crack, such as the crack edges and slender structures.

[0077] The module operates as follows: First, the three side outputs, side1, side2, and side3, are concatenated to obtain a new input feature map. Next, this feature map undergoes adaptive average pooling along both the height and width directions, generating two global feature maps. These two feature maps are then concatenated and subjected to channel compression via a 1×1 convolutional layer, followed by batch normalization and the introduction of non-linearity through the HSwish activation function. After these steps, the concatenated feature map is segmented into two directions: height and width. The feature map in each direction is first processed through a 1×1 convolutional layer, and then a Sigmoid activation function is applied to each feature map to generate corresponding attention weights. These attention weights aim to enhance regions in the feature map that are important in a specific direction (height or width). Simultaneously, the original input feature map undergoes a global average pooling operation to reduce it to a global representation, which is then input into a fully connected layer to generate dynamic weights. These dynamic weights adaptively adjust the importance of the attention weights in the height and width directions, ensuring the model focuses on the most critical spatial information. Finally, the original input feature maps are element-wise multiplied with the generated attention weights in the height and width directions, respectively, applying the attention weights to the feature maps. These weighted feature maps are then further processed through a 3×3 convolutional layer for channel integration, followed by batch normalization and the SiLU activation function to generate the final output feature map, thus obtaining the optimized side output side3. The entire process not only preserves multi-scale features but also strengthens key feature regions through a dynamic attention mechanism, significantly improving the model's ability to segment complex crack structures.

[0078] Assume the outputs of the three sides are X side1 X side2 X side3This process can then be represented by the following formula:

[0079] X concat =Concat(X) side1 ,X side2 ,X sid )

[0080] X height =AAP height (X concat )

[0081] X width =AAP width (X concat )

[0082] X global =HSwish(BN(C 1×1 (Concat(X height ,X width ))))

[0083] X global_height ,X global_width =Split(X) global )

[0084] A height =Sigmoid(C 1×1 (X global_height ))

[0085] A width =Sigmoid(C 1×1 (X global_width ))

[0086] W dynamic =FC(GAP(X) concat ))

[0087] A′ height =W dynamic ·A height

[0088] A′ width =W dynamic ·A width

[0089] Y = SiLU(BN(C) 3×3 (X concat ·A′ height +X concat ·A′ width )))

[0090] Among them, X concat Indicates splicing features, X sid ,X side ,Xside3 These represent the first, second, and third middle-layer outputs, respectively; Concat indicates the concatenation operation; AAP height and AAP width These represent adaptive average pooling in the height and width directions, respectively; X height and X width C represents the height and width pooling features, respectively; 1×1 Represents a 1×1 convolution, C 3×3 This represents a 3×3 convolution, BN represents a batch normalized layer, HSwish represents the HSwish activation function, and X represents a 3×3 convolution. global Represents global features; Split represents a segmentation operation; X global_height and X global_width These represent the global features of height and width, respectively; Sigmoid represents the Sigmoid activation function; A height and A width These represent the height and width attention weights, respectively; GAP represents global average pooling; FC represents a fully connected layer; and W... dynamic Indicates dynamic weights; A′ height and A′ width represents the adjusted height and width attention weights, respectively; SiLU represents the SiLU activation function; and Y represents the output feature of the MADFM module.

[0091] Binary cross-entropy loss function:

[0092] This invention employs deep supervision to monitor each side output layer and assigns different weights to each side output layer, with the corresponding weights being w = {(w (1) ,…,w (N) )}, where N is the number of side output layers. Meanwhile, W represents all parameters of the network proposed in this invention.

[0093] Therefore, the formula for the total loss function of the side output layer is as follows:

[0094]

[0095] Where G represents the total pixels of the label image, I represents the total pixels of the original input image, G represents the total pixels of the label image, W represents the total parameters of the model, and α n It is a hyperparameter that represents the loss weights of each side output layer; The side-output loss of the nth side-output layer is expressed as:

[0096]

[0097] Because the input images in the dataset exhibit uneven distribution of cracked and non-cracked pixel samples, the loss function of this invention requires different weighted losses. Therefore, class balance is achieved by assigning different weights to positive and negative samples. Here, w0 and w1 represent the weights of cracked and non-cracked pixels, respectively, and G... + With G - P represents the sets of crack (positive) pixels and non-crack (negative) pixels, respectively. j G represents the prediction result for the j-th pixel. j This represents the true label of the pixel, w (n) This represents the parameters of the nth side output layer.

[0098] Since each side output layer can generate a crack prediction map, the network model proposed in this invention finally connects the prediction maps obtained from the five side output layers together and fuses them to form the final fused prediction map. The resulting fusion loss is as follows:

[0099]

[0100] The total network loss function proposed in this invention, namely the binary classification cross-entropy loss function, is the sum of the fusion loss and the total side output loss, expressed as:

[0101] L = L side (I,G,W,w)+L fuse (I,G,W)

[0102] During model training, the model parameters are adjusted based on the total loss. Training stops when the total loss converges or reaches the maximum preset number of iterations, and the model parameters are saved, resulting in a trained concrete pavement crack detection model. The pavement image to be processed is then acquired and input into the trained concrete pavement crack detection model for further processing to obtain the crack detection results.

[0103] Evaluation of the present invention:

[0104] To verify the effectiveness of the proposed model, this invention is compared with several other deep learning crack detection techniques (UNet, UNet++, HED, DeepCrack, CrackFormer, CrackW-Net, HACNE, PAF-Net). To ensure impartiality, all methods use the same training parameters and are fully trained on the DeepCrack dataset. The comparison results are as follows: Figure 4 As shown, from Figure 4As can be seen, the detection results of this invention are clearer, with less noise, and it can better identify small cracks and capture more key details. In summary, this invention designs a concrete pavement crack detection network based on progressive contextual interaction and attention mechanisms. The study optimized three stages: feature extraction, enhancement, and fusion. In the feature extraction stage, by introducing asymmetric convolutional blocks and residual structures, not only is the effective receptive field expanded, but the stability of gradient propagation is also enhanced. The progressive contextual interaction mechanism enables the exchange and fusion of features at different resolutions, comprehensively integrating multi-scale contextual information, which helps to more accurately capture the global and local features of cracks. In the feature enhancement stage, different types of attention modules are embedded into multiple side outputs, and the feature representation capability is improved through the combination of vertical, horizontal, and spatial attention. In the feature fusion stage, a multi-semantic attention dynamic fusion module is used to achieve dynamic fusion of shallow and intermediate layer features, thereby enhancing the ability to recover and preserve details. Overall, this invention shows significant improvement in feature capture and detail restoration in crack detection, providing an effective and innovative solution for crack detection tasks.

[0105] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting cracks in concrete pavement based on progressive contextual interaction, characterized in that, include: The road surface image to be processed is obtained and input into the trained concrete road surface crack detection model for processing to obtain the concrete road surface crack detection result. The concrete pavement crack detection model adopts a progressive encoding-decoding structure, and the model is divided into three stages: feature extraction stage, feature enhancement stage, and feature fusion stage. The feature extraction stage includes the ACRB module and a progressive context interaction mechanism; The feature enhancement stage includes a triple attention module and a hybrid module of vertical and horizontal attention; The feature fusion stage includes convolutional layers and the Multi-Semantic Attention Dynamic Fusion (MADFM) module; The ACRB module includes a 3×3 convolution, an asymmetric convolution block ACBlock, a batch normalization layer, a LeakyReLU activation function, and residual connections; The data processing steps of the progressive context interaction mechanism include: The input features are summed to obtain a fused feature map; The fused feature map is non-linearly transformed by the GeLU activation function and then passed through a 3×3 convolutional layer and a GroupNorm normalization layer to obtain the first intermediate feature. The first intermediate feature is passed through another set of 3×3 convolutions and GroupNorm normalization layers, and finally a mask is generated by the Sigmoid activation function. After multiplying the fused feature map element-wise with the mask, the result is then passed through a 1×1 convolution, a 3×3 convolution, and a 3×3 convolution with a dilation rate of 2. The three convolution results are summed, and finally passed through a batch normalization layer to complete the progressive context interaction mechanism. The data processing steps in the feature enhancement stage include: The first shallow-layer output is processed by a triple attention module to obtain the first middle-layer output; The second shallow layer output is processed by a hybrid module of vertical attention and horizontal attention to obtain the second middle layer output; The third shallow layer output is processed by a hybrid module of vertical attention and horizontal attention to obtain the third middle layer output; The fourth shallow-layer output is processed by a hybrid module of vertical and horizontal attention to obtain the fourth middle-layer output; The fifth shallow-layer output is processed by a triple attention module to obtain the fifth middle-layer output; The data processing procedure of the MADFM module is represented as follows: ； ； ； ； ； ； ； ； ； ； ； in, Indicates splicing characteristics, These represent the first, second, and third middle-layer outputs, respectively. Indicates a splicing operation; and These represent adaptive average pooling in the height and width directions, respectively; and These represent the height and width pooling features, respectively. Represents a 1×1 convolution. This represents a 3×3 convolution; Indicates the batch normalization layer. This represents the HSwish activation function. Represents global features; Indicates a splitting operation; and These represent the global features of height and width, respectively. This represents the Sigmoid activation function; and These represent the attention weights for height and width, respectively. Indicates global average pooling. Indicates a fully connected layer. Indicates dynamic weights; and These represent the adjusted height and width attention weights, respectively. This represents the SiLU activation function. This indicates the output characteristics of the MADFM module.

2. The method for detecting cracks in concrete pavement based on progressive contextual interaction according to claim 1, characterized in that, The data processing steps in the feature extraction stage include: The road surface image is processed by the ACRB module and downsampling to obtain feature A; feature A is then processed by the ACRB module to obtain feature B; feature B is upsampling and then processed with the first shallow side output through a progressive context interaction mechanism to obtain the second shallow side input; the second shallow side input is processed by the ACRB module to obtain the second shallow side output. Feature B and the first shallow side output, after being downsampled, are then processed by a progressive context interaction mechanism to obtain feature C; feature C is then processed by the ACRB module to obtain feature D; feature D is upsampled and then processed by the second shallow side output through a progressive context interaction mechanism to obtain the third shallow side input; the third shallow side input is processed by the ACRB module to obtain the third shallow side output. Feature B is downsampled to obtain feature I; feature I is processed by the ACRB module to obtain feature J; the downsampled features of the second shallow side output, feature D, and feature J are upsampled and then processed by a progressive context interaction mechanism to obtain feature E; feature E is processed by the ACRB module to obtain feature F; feature F is upsampled and then processed by the progressive context interaction mechanism with the third shallow side output to obtain the fourth shallow side input; the fourth shallow side input is processed by the ACRB module to obtain the fourth shallow side output. Feature D is downsampled and then processed with feature J through a progressive context interaction mechanism to obtain feature K; feature K is processed by the ACRB module to obtain feature L; the downsampled features of the third shallow side output, feature F, and feature L are upsampled and then processed by a progressive context interaction mechanism to obtain feature G; feature G is processed by the ACRB module to obtain feature H; feature H is upsampled and then processed with the fourth shallow side output through a progressive context interaction mechanism to obtain the fifth shallow side input; the fifth shallow side input is processed by the ACRB module to obtain the fifth shallow side output.

3. The method for detecting cracks in concrete pavement based on progressive contextual interaction according to claim 1, characterized in that, The data processing in the feature fusion stage includes: the first intermediate layer output is processed by 1×1 convolution to obtain the first deep layer output; the second intermediate layer output is processed by 1×1 convolution to obtain the second deep layer output; the first to third intermediate layer outputs are processed by the MADFM module to obtain the third deep layer output; the fourth intermediate layer output is processed by 1×1 convolution to obtain the fourth deep layer output; the fifth intermediate layer output is processed by 1×1 convolution to obtain the fifth deep layer output; the first to fifth deep layer outputs are spliced ​​together and then processed by 1×1 convolution to obtain the concrete pavement crack detection result.

4. The method for detecting cracks in concrete pavement based on progressive contextual interaction according to claim 1, characterized in that, The loss function in the training process of the concrete pavement crack detection model is the sum of the total side output loss and the fusion output loss.

5. The method for detecting cracks in concrete pavement based on progressive contextual interaction according to claim 4, characterized in that, The total output loss is expressed as: ； ； in, Let I represent the total output loss, and G represent the total pixels of the original input image and the label image, respectively. Represents the total parameters of the model. This represents the weights of each side of the output layer. Indicates the number of side output layers. This represents the loss weights of the nth side output layer. This represents the side-output loss of the nth side-output layer. and These represent the weights of the pixels with and without cracks, respectively. and These represent the sets of cracked pixels and non-cracked pixels, respectively. This represents the prediction result for the j-th pixel. This indicates the actual label of the pixel; This represents the parameters of the nth side output layer.

6. The method for detecting cracks in concrete pavement based on progressive contextual interaction according to claim 4, characterized in that, The fusion output loss is expressed as: ； in, The output loss represents the fusion loss, where I represents the total pixels of the original input image and G represents the total pixels of the label image. Represents the total parameters of the model; Indicates the number of side output layers. and These represent the weights of the pixels with and without cracks, respectively. and These represent the sets of cracked pixels and non-cracked pixels, respectively. This represents the prediction result for the j-th pixel. This indicates the actual label of the pixel.

Images