Intelligent concrete crack detection method based on four-cardinal-number coordinate attention enhancement

Abstract

The invention discloses a concrete crack intelligent detection method based on four-cardinal-number coordinate attention enhancement, and belongs to the technical field of computer vision and civil engineering detection. According to the invention, a QuadCoord-YOLO network is constructed by improving a YOLO11 network architecture, and a CoT-CSP module and a QCE module are innovatively introduced: the CoT-CSP module deeply fuses a context converter mechanism and a CSP structure, so that the long-distance dependency relationship modeling capability is enhanced; the QCE module proposes a four-cardinal parallel channel attention mechanism for the first time, realizes three-path parallel multi-dimensional feature enhancement in combination with coordinate attention and deep convolution, and breaks through the information bottleneck of single-path compression of a traditional SE module. Experimental results show that according to the method, the bounding box detection mAP50 is improved from 72.6% to 78.1%, the segmentation mAP50 is improved from 58.9% to 66.6%, and only 3.2% of parameter quantity (82, 456 parameters) is increased. The method can be applied to an intelligent structure health monitoring system, and provides accurate technical support for concrete structure condition evaluation and maintenance decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision and civil engineering detection technology, and particularly relates to a concrete crack intelligent detection method based on four-base coordinate attention enhancement. BACKGROUND

[0002] With the rapid development of civil engineering construction in China, the scale of building structures is continuously expanding, and the demand for maintenance of concrete structures is increasing. As a major form of structural damage, timely and accurate detection of concrete cracks is of great significance to ensure structural safety and prolong the service life of buildings. Traditional manual inspection methods have low efficiency, high cost, and strong subjectivity of detection results, which cannot meet the detection needs of large-scale building structures.

[0003] In recent years, target detection technology based on deep learning has been widely applied in the field of concrete crack identification. The YOLO (You Only Look Once) series algorithm, as a mainstream single-stage target detection method, has the advantages of fast detection speed and strong real-time performance, and has shown good performance in crack detection tasks. YOLO11, as a new version, further optimizes the network structure and feature extraction capability. However, existing crack detection methods based on YOLO still have the following key technical bottlenecks: Firstly, concrete cracks have complex characteristics such as elongation, irregularity, and multi-scale, and traditional convolutional neural networks have limited ability to handle long-distance dependencies. Although the C3k2 module in YOLO11 uses cross-stage partial connection, it still has deficiencies in capturing global context information of cracks. Specifically, for elongated cracks longer than 100 pixels, existing methods are prone to produce broken detection; for curved cracks, it is difficult to accurately grasp their overall shape; for intersecting cracks, it is easy to miss the detailed features near the intersection points.

[0004] Secondly, existing methods have a serious information bottleneck problem in channel feature extraction. Traditional attention mechanisms mainly focus on the importance of channel features, and the most representative SE (Squeeze-and-Excitation) module realizes channel attention through global average pooling and two-layer fully connected networks. However, the SE module adopts a single-path feature compression structure ( , usually ), forming a serious information bottleneck at the middle layer. Taking the input channel number as an example, the SE module compresses it to dimension, and a large amount of feature information is lost in this compression process, severely limiting the diversity and richness of feature expression. For cracks, which have various morphological features (horizontal cracks, vertical cracks, cracks, and network cracks), single-path compression cannot fully retain the discriminative information of various features.

[0005] Thirdly, the utilization of spatial position information is insufficient. Traditional attention mechanisms mainly focus on channel dimensions, while cracks have obvious directionality and spatial distribution characteristics. For example, horizontal cracks mainly extend along the width direction, and vertical cracks mainly extend along the height direction. Existing methods lack explicit modeling of such directional characteristics, resulting in insufficient perception of crack direction and position information.

[0006] Finally, the multi-dimensional feature fusion mechanism needs to be optimized. Existing methods usually use simple addition or splicing to fuse different types of features, lacking a collaborative modeling mechanism for channel attention, spatial attention, and local convolutional features. In particular, in the small-scale feature detection branch, there is a lack of effective multi-dimensional feature enhancement mechanism, resulting in unsatisfactory detection results for small cracks. SUMMARY

[0007] The purpose of the present application is to provide a concrete crack intelligent detection method based on four-base coordinate attention enhancement, which solves the key problems of single-path feature compression bottleneck, insufficient long-distance dependency modeling, and weak spatial detail perception ability in the prior art by designing a four-base parallel attention mechanism and a context CSP fusion structure.

[0008] To achieve the above purpose, the present application provides a concrete crack intelligent detection method based on four-base coordinate attention enhancement, comprising the following steps: S1: Construct a concrete crack dataset, accurately label the crack images, form complete labeling information containing boundary boxes and pixel-level segmentation masks, and divide the training set, validation set and test set; S2: Construct a QuadCoord-YOLO network architecture, innovatively introduce a CoT-CSP (Context Transformation based on CSP structure) module and a QCE (Quad-Cardinality Enhancement) module based on YOLO11; S3: Design the CoT-CSP module, which realizes efficient modeling of crack long-distance dependency and adaptive integration of local-global features by deeply integrating the context transformer mechanism with the CSP structure; S4: Construct the QCE module, innovatively propose a four-base parallel channel attention mechanism, combine coordinate attention to realize multi-dimensional and multi-path feature enhancement, and break through the information bottleneck of traditional single-path SE module; S5: Train the QuadCoord-YOLO network end-to-end using the training set, and optimize the network parameters through the carefully designed compound loss function; S6: Use the trained model to detect and segment cracks in concrete structure images, and output the crack bounding box, segmentation mask, confidence score, and quantitative analysis results.

[0009] Further preferably, in step S1, the concrete crack dataset adopts the YOLO polygon annotation format, with the training set containing 3717 images, the validation set containing 200 images, and the test set containing 112 images.

[0010] Further preferred, in step S2, the QuadCoord-YOLO network architecture is an innovative improvement based on YOLO11. The backbone network retains the initial two layers. Convolution (stride 2), replacing the C3k2 modules in layers 2, 4, 6, and 8 with CoT-CSP modules, with CoT-CSP repetition counts of 2, 2, 2, and 2 respectively, and setting the shortcut to... Retain the SPPF module at level 9 ( Multi-scale feature aggregation is performed, retaining the C2PSA module (repeated twice) in layer 10 to provide position-sensitive attention. The neck network adopts an FPN-PAN bidirectional feature fusion structure, achieving multi-scale feature transfer through upsampling and feature concatenation. The detection head innovatively integrates a QCE module after layer 16, with 64 input and output channels each, specifically enhancing small-scale crack features. Layer 23 uses a segment detection head with [number of output channels missing]. This enables joint prediction of bounding boxes and segments.

[0011] Further optimized, in step S3, the innovative design of the CoT-CSP module fully integrates the global modeling capability of the context transformer and the gradient flow advantages of the CSP structure. For the input feature map... Through two Convolutional branches cv1 and cv2 generate features respectively. and All dimensions Where e is the expansion factor (for layers 2 and 4) Floors 6 and 8 Apply n concatenated context attention blocks to X1, each block containing: key embeddings generated via 3×3 grouped convolutions ( Extracting local contextual features; value embedding generation uses... Convolution preserves the original features; dynamic attention computation fuses the key embeddings and the original features to generate... Spatial adaptive weights; feature enhancement applies dynamic weights to value embedding. Finally, the transformed... and splicing, through Convolutional CV3 fusion, when Add residual connections as needed.

[0012] In a further preferred embodiment, in step S4, the QCE module is the core innovation of this invention, which adopts a three-way parallel architecture to achieve multi-dimensional feature collaborative enhancement.

[0013] The coordinate attention submodule performs adaptive average pooling (H×1 and 1×W) on the input features to capture feature distributions in different directions. The pooled features from the two directions are concatenated and then reduced to max(8, C / 2) using a 1×1 convolution, followed by the h-swish activation function. Attention weights in the height and width directions are then generated through two separate 1×1 convolutions, activated by a sigmoid function, and multiplied element-wise to obtain the spatial attention map.

[0014] The four-cardinality channel attention submodule is a key innovation of this invention. It compresses the input into a C×1×1 global description vector through global average pooling. Four independent fully connected layer branches are designed, each with a compression ratio of 2, generating C / 2 dimensional features: the first branch focuses on low-frequency global information, the second branch on high-frequency detail information, the third branch on mid-frequency transition information, and the fourth branch on specific pattern information. The outputs of the four branches are concatenated to form a 2C-dimensional aggregated feature, which is then mapped back to C dimensions through a fully connected layer and activated by a sigmoid function. This four-path design significantly enhances the diversity of feature representation compared to traditional single-path attention modules.

[0015] The convolutional enhancement submodule uses two layers of 3×3 convolutions combined with batch normalization and ReLU activation to extract deep local features and adds residual connections to maintain gradient flow.

[0016] Ultimately, through an adaptive fusion mechanism: This achieves synergistic enhancement of the three branch features.

[0017] Further preferred, the model training parameters are set as follows: input image size 640×640, batch size 16, initial learning rate 0.01, number of training epochs 100, optimizer SGD, momentum 0.937, weight decay 0.0005.

[0018] Therefore, the intelligent concrete crack detection method based on four-base coordinate attention enhancement using the above structure of the present invention has the following beneficial effects: (1) This invention proposes a four-radix parallel channel attention mechanism for the first time. Through the four independent squeeze paths in the QCE module, it breaks through the information bottleneck of the single-path compression of the traditional SE module and significantly enhances the diversity and richness of feature expression.

[0019] (2) The present invention achieves deep integration of context transformer and CSP structure. The CoT-CSP module effectively combines the global modeling capability of Contextual Transformer and the gradient flow advantage of CSP, solving the problem of long-distance dependency modeling of slender cracks.

[0020] (3) The optimized network architecture design of this invention achieves significant performance improvement by precisely deploying innovative modules at key locations, with only a 2.9% increase in the number of parameters.

[0021] (4) The detection performance of the present invention is improved. Experimental results show that the bounding box detection mAP50 is improved from 0.726 to 0.781 (an improvement of 7.6%), and the segmentation mAP50 is improved from 0.589 to 0.666 (an improvement of 13.1%), which fully verifies the effectiveness of the method.

[0022] (5) The present invention has high training efficiency. The entire training process can complete the training of 100 epchs in only 0.916 hours, demonstrating good training efficiency.

[0023] (6) The present invention has good engineering application value. The method has fast reasoning speed, good generalization ability and practicality, and can be widely used in intelligent structural health monitoring system.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] Figure 1 This is an overall flowchart of the QuadCoord-YOLO method of the present invention; Figure 2 This is a diagram of the overall architecture of the QuadCoord-YOLO network. Figure 3 A detailed structural diagram of the CoT-CSP module; Figure 4 A detailed structural diagram of the QCE module; Figure 5 This is a graph showing the loss function and accuracy changes during the training process. Detailed Implementation

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0028] Example like Figure 1 As shown, this invention provides a smart concrete crack detection method based on four-base coordinate attention enhancement, and the specific implementation steps are as follows: S1: Constructing a Concrete Crack Dataset Images of concrete structures from different locations, under varying lighting conditions, and with different crack types were collected to construct a dataset of 4029 images. Crack regions were precisely annotated using the YOLO polygon annotation format, with annotation information including the bounding box coordinates and pixel-level segmentation masks. The dataset was divided into a training set of 3717 images, a validation set of 200 images, and a test set of 112 images, in a 9:0.5:0.5 ratio. Data preprocessing included image size normalization, adjusting all images to 640×640 pixels; data augmentation using random horizontal flip probability of 0.5, random scaling ratio of 0.5, random translation ratio of 0.1, and HSV color space enhancement (H channel 0.015, S channel 0.7, V channel 0.4); mosaic enhancement by stitching four images together to form new training samples, enhancing the model's ability to detect targets at multiple scales; and random erasure with a probability of 0.4 to randomly occlude local areas of the image, improving the model's robustness.

[0029] S2: Building an improved YOLO11 network architecture Reference Figure 2As shown, the improved YOLO11 network architecture consists of three main parts. The backbone network improvement retains the 3×3 convolutional stride of layer 0 with a stride of 2 as the initial feature extraction layer, retains the 3×3 convolutional stride of layer 1 with a stride of 2 for downsampling, replaces the C3k2 modules in layers 2, 4, 6, and 8 with CoT-CSP modules to enhance long-range dependency modeling, retains the SPPF module in layer 9 for multi-scale feature aggregation, and retains the C2PSA module in layer 10 for position-sensitive attention processing. The neck network design adopts a structure combining a feature pyramid network and a path aggregation network, achieving multi-scale feature transfer through upsampling and feature fusion. Bilinear interpolation upsampling and feature concatenation operations are used in layers 11 to 15, and the C3k2 module is used for feature refinement in layers 13, 16, 19, and 22. The detection head enhancement integrates the QCE module after the 16th layer of the P3 detection branch, keeping the detection branch structure of the 19th layer of P4 and the 22nd layer of P5 unchanged. The 23rd layer adopts the Segment detection head to achieve the joint output of bounding box detection and instance segmentation.

[0030] S3: Design CoT-CSP Module Reference Figure 3 As shown, the CoT-CSP module is designed based on the context transform mechanism and includes three components: channel separation mechanism, context transform unit, and feature fusion.

[0031] S31, Channel Separation Mechanism: S311, Input Features Through two Convolutional branches cv1 and cv2; S312, Generate features respectively and All dimensions ; S313, among which As the expansion factor, layers 2 and 4 Floors 6 and 8 .

[0032] S32, Context Transformation Unit: S321, Apply to the first branch X1 A series of CoTBottleneck modules; S322, Each CoTBottleneck contains: S33, Key Embedding Generation: Through Grouped convolution ( Extract local contextual features; S34, Value Embedding Generation: Using Convolution preserves the original feature information; S35, Dynamic Attention Calculation: Concatenate the key embedding and the original features, and generate k×k×C spatial adaptive convolutional kernel weights through the attention embedding layer; S36. Feature Enhancement: Applying dynamic weights to value embedding to achieve context-aware feature enhancement; S37, Feature Fusion: S371, Transformed With unchanged bypass branches splicing along the channel dimension; S372, Through convolution Combined into the final output; S373, when Add residual connections as needed.

[0033] S4: Building the QCE module For reference Figure Four As shown, the QCE (Quad-Cardinality Enhancement) module uses a three-way parallel architecture to achieve multi-dimensional feature collaborative enhancement.

[0034] S41, Coordinate Attention Branch: S411, Input Features Perform adaptive average pooling separately: Horizontal direction: ; Vertical direction: ; S412. Combine the features from the two directions in the spatial dimension: ; S413, Reduce dimensionality to 1×1 convolution. Use activation; S414. Generate attention weights in the height and width directions respectively: ; ; S415, Spatial Attention Map: .

[0035] S42, Four-Radius Channel Attention Branch (Core Innovation): S421, Compressed using global average pooling to The global feature vector; S422. Innovatively designed four parallel fully connected layer branches, each performing feature compression independently (compression ratio r=2): Branch 1 Focus on low-frequency global information to capture overall structural features; Branch 2 Focus on high-frequency detail information and capture texture edge features; Branch 3 Focus on mid-frequency transition information to capture intermediate-scale features; Branch 4 Focus on specific pattern information and capture domain-related features.

[0036] S423, The four branches Dimensional output is formed by splicing together the channels. Dimensional aggregation features; S424. Map back to the C-dimensional dimension through a fully connected layer and generate channel attention weights after Sigmoid activation: ; Advantages compared to traditional SE modules: SE modules employ a single-path compression structure ( This presents a serious information bottleneck. The four-radix structure of this invention employs four parallel paths (each path C → C / 2), with a total compression dimension of 2C, avoiding single-point bottlenecks and greatly enhancing the diversity and richness of feature representation.

[0037] S43, Convolutional Feature Enhancement Branch: S431, First Layer: ; S432, Second Layer: ; S433, Residual Connection: ; S44, Three-way adaptive fusion: S441, Final Output: ; S442, where ⊙ represents element-wise multiplication (Hadamard product), achieving synergistic enhancement across three dimensions.

[0038] Step S5: Network Training Process Training parameters were set as follows: input image size 640×640 pixels, batch size 16, training epochs 100, initial learning rate 0.01, learning rate decay using cosine annealing, final learning rate 0.01, optimizer SGD momentum 0.937, weight decay 0.0005, mixed precision training enabled with AMP acceleration, and a warm-up strategy for the first 3 epochs for learning rate preheating. The loss function was designed as a weighted sum of four loss terms. The bounding box regression loss uses the CIoU loss function, which considers the overlap of the bounding boxes, the distance between their center points, and their aspect ratio. The classification loss uses binary cross-entropy loss. The segmentation mask loss combines Dice loss and cross-entropy loss. Distribution focus loss is used to optimize the distribution modeling of bounding box regression. Training strategies include enabling mosaic enhancement for the first 90 epochs and disabling it for the next 10 epochs, updating gradients every 64 samples through gradient accumulation, stopping training when validation set performance shows no improvement for 100 consecutive epochs, and saving the model weights that best perform on the validation set.

[0039] Step S6: Crack Detection and Segmentation After training, the optimal model is used for inference on concrete structure images. Image preprocessing resizes the input image to 640×640 pixels and normalizes the pixel values ​​by dividing them by 255. Forward inference extracts features using an improved YOLO11 network, generating prediction results including bounding boxes, confidence scores, and segmentation masks. Post-processing includes non-maximum suppression (IoU) with an IoU threshold of 0.7 and a confidence threshold of 0.25. The segmentation mask is refined and upsampled to the original image size. The output includes crack location, category, confidence score, and pixel-level segmentation results.

[0040] Example 2: Performance Verification Experiment To verify the effectiveness of the method of this invention, comparative experiments were conducted. The experimental setup included an NVIDIA GeForce RTX 3090 GPU with 24GB of video memory, and PyTorch 2.5.1 and CUDA 12.4 software. Evaluation metrics included the average accuracy at mAP50 (IoU threshold of 0.5) and the average accuracy at mAP50-95 (IoU thresholds of 0.5 to 0.95). The ablation experiment results are shown in Table 1.

[0041] Table 1 Ablation Experiment Results

[0042] Experimental results show that the method of this invention achieves 78.1% mAP50 for bounding box detection, a 5.5 percentage point improvement compared to the YOLO11 benchmark of 72.6%. It achieves 58.3% mAP50-95 for bounding box detection, a 5.7 percentage point improvement compared to the benchmark of 52.6%. In the segmentation task, the mask mAP50 reaches 66.6%, a 7.7 percentage point improvement compared to the benchmark of 58.9%, and the mask mAP50-95 reaches 24.1%, a 3.2 percentage point improvement compared to the benchmark of 20.9%. The model has 2.92M parameters, only a 3.2% increase compared to the benchmark of 2.83M, and the computational cost is 12.1 GFLOPs, an 18.6% increase compared to the benchmark of 10.2 GFLOPs, maintaining good efficiency.

[0043] Ablation experiments showed that using the CoT-CSP module alone achieved a bounding box mAP50 of 74.9%, an improvement of 2.3 percentage points, demonstrating the effectiveness of context transformation. Using the QCE module alone achieved a bounding box mAP50 of 76.2%, an improvement of 3.6 percentage points, demonstrating the role of coordinate-aware enhancement. Using both modules together yielded the best performance, demonstrating their complementarity. The training process was completed within 100 epochs, taking 0.916 hours, demonstrating good training efficiency.

[0044] This invention constructs a high-precision concrete crack detection and segmentation network by integrating context-aware transformation and coordinate-aware feature enhancement techniques. The introduction of a CoT-CSP module into the backbone network enhances the modeling capability of long-distance dependencies, while the integration of a QCE module into the detection head improves the accuracy of detail feature extraction. Experimental results demonstrate that this method significantly improves the accuracy of crack detection and segmentation while maintaining low computational overhead, providing reliable technical support for intelligent structural health monitoring systems and holding significant importance for promoting the application of computer vision technology in the field of civil engineering structural inspection.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A four-base coordinate attention enhanced intelligent concrete crack detection method, characterized in that, The method comprises the following steps: S1, constructing a concrete crack dataset, labeling crack images in the dataset to generate labeling information containing boundary box information and pixel-level segmentation mask, and dividing the dataset into a training set, a validation set and a test set; S2, constructing a QuadCoord-YOLO network architecture based on a YOLO series network, replacing at least part of the original feature extraction module in the backbone network of the YOLO series network with a CoT-CSP module, integrating a QCE module in the detection head of the YOLO series network, the CoT-CSP module being a module deeply integrating a context transformer mechanism and a CSP structure, the QCE module being a module integrating a four-base parallel channel attention mechanism, a coordinate attention mechanism and deep convolution enhancement, and the QuadCoord-YOLO network architecture further comprising a feature fusion subnetwork in the neck; S3, designing the CoT-CSP module, dividing input features into a first branch and a second branch through a double-path feature separation unit, applying at least one context attention block in series to the first branch for context transformation processing, and then splicing and convolving the first branch features after context transformation processing with the second branch features through a feature fusion unit; S4, constructing the QCE module, adopting a three-way parallel architecture, respectively generating spatial attention weights through a coordinate attention submodule, generating channel attention weights through a four-base channel attention submodule, extracting deep local features through a convolution feature enhancement submodule, and then weighting and fusing the outputs of the three submodules through an adaptive fusion unit; S5, using the training set to perform end-to-end training on the QuadCoord-YOLO network, and optimizing the parameters of the QuadCoord-YOLO network through a compound loss function; S6, using the trained QuadCoord-YOLO network model to detect and segment cracks in a concrete structure image, and outputting boundary boxes, segmentation masks and confidence information of the cracks.

2. The method according to claim 1, wherein, In step S1, the concrete crack dataset adopts a YOLO polygon labeling format, the training set contains 3717 images, the validation set contains 200 images, and the test set contains 112 images.

3. The method of claim 1, wherein the method is based on a four-base coordinate attention enhanced concrete crack intelligent detection method. In step S2, the YOLO series network is YOLOv1, the original feature extraction module replaced in the backbone network is the C3k2 module of the 2nd, 4th, 6th and 8th layers, the number of repetitions of the CoT-CSP module is 2, and the shortcut is set to ; the neck feature fusion subnetwork is a FPN-PAN bidirectional feature fusion structure; in the detection head, the QCE module is integrated after the 16th layer of the P3 detection branch, the input and output channel numbers of the QCE module are both 64, and the output channel number of the detection head is .

4. The method of claim 1, wherein the method is based on a four-base coordinate attention enhanced concrete crack intelligent detection method. In step S3, the dual-path feature separation unit is two Convolution; the context attention block comprises a key embedding generation layer, a value embedding generation layer, a dynamic attention calculation layer and a feature enhancement layer, the key embedding generation layer adopts Grouped convolution, group number The value embedding generation layer adopts Convolution.

5. The method of claim 1, wherein the method is based on a four-base coordinate attention enhanced concrete crack intelligent detection method. In step S4, the coordinate attention submodule processes the input features respectively. and Adaptive average pooling is used to obtain pooling features in two directions; these pooling features are then concatenated and processed... Convolution dimensionality reduction to And activated by h-swish, then through two Convolution generates height-oriented attention weights Attention weights in the width direction The and After activation by the Sigmoid algorithm, the elements are multiplied one by one to obtain the spatial attention weights.

6. The method of claim 1, wherein the method is based on a four-base coordinate attention enhanced concrete crack intelligent detection method. In step S4, the four-base channel attention submodule obtains a global feature vector of C×1×1 by performing global average pooling on the input feature; the four-base channel attention submodule includes four parallel fully connected branches, each branch has a compression ratio of 2, and the input feature is compressed from to ; the outputs of the four fully connected branches are concatenated in the channel dimension to form -dimensional aggregated features, which are then mapped back to -dimensional features by a fully connected layer and passed through a Sigmoid activation function to obtain channel attention weights.

7. The method of claim 1, wherein the method is based on a four-base coordinate attention enhanced concrete crack intelligent detection method. In step S4, the convolution feature enhancer module includes two layers of convolutions, each of the convolutions is sequentially connected with a batch normalization layer and a ReLU activation function; the convolution feature enhancer module further includes a residual connection, the residual connection is two layers of convolutions, and outputs of the two layers of convolutions are directly superimposed on input features; a fusion formula of the adaptive fusion unit is: ； wherein, is a spatial attention weight, is an output of the convolutional feature enhancement sub-module, is a channel attention weight, is an element-wise multiplication.

8. The method of claim 1, wherein the method is based on a four-base coordinate attention enhanced concrete crack intelligent detection method. In step S5, the expression of the compound loss function is: ； wherein, is a CIoU loss, is a binary cross-entropy loss, is a combination of Dice loss and cross-entropy loss, is a distributed focal loss.

9. The method of claim 1, wherein, In step S5, the parameter settings of the end-to-end training are as follows: input image size of 640x640 pixels, batch size of 16, initial learning rate of 0.01, training rounds of 100, optimizer of SGD, momentum of 0.937, weight decay of 0.0005; during the training process, AMP mixed precision training is enabled, and the learning rate is preheated for the first 3 training rounds. In step S6, before performing crack detection and segmentation on the concrete structure image, the concrete structure image is first adjusted to 640x640 pixels and normalized.

10. The method of claim 1, wherein, ​ After detection and segmentation, a post-processing step is further included: repeated bounding boxes are removed by non-maximum suppression with an IoU threshold of 0.7 and a confidence threshold of 0.25; the segmentation mask is up-sampled to the original size of the concrete structure image, and the position, category, confidence and pixel-level segmentation result of the crack are output.

Images