Concrete dam surface crack detection method and system based on diamond convolution unet

Abstract

The application discloses a kind of based on diamond convolution UNet's concrete dam surface crack detection method and system, the method in which first image acquisition is carried out to concrete dam surface;And the image collected is labeled, obtains image dataset;Then the image dataset is divided into training set and verification set;Again build diamond convolution Unet's image segmentation network, then the training set obtained by division is used to train diamond convolution Unet's image segmentation network, obtain training model, then the training model is verified using verification set, obtain algorithm model;Finally, using algorithm model is detected concrete surface and carries out crack feature identification, obtains crack detection result.The application solves the problem that the existing method is not complete to the crack shape feature identification of concrete dam surface, and improves the detection accuracy of the crack of concrete dam surface.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital image processing, and particularly relates to a concrete dam surface crack detection method and system based on a rhombus convolution UNet. BACKGROUND

[0002] During the construction and operation periods of a water conservancy and hydropower project concrete dam, the dam surface will inevitably produce various cracks under the action of complex conditions such as external temperature, humidity and load. Cracks are one of the most common diseases of concrete dams, which not only causes the decline of dam structure strength, stiffness and durability, but also even leads to the local or overall instability and damage of the dam structure. Therefore, it is of great significance to study the concrete dam surface crack detection technology to ensure the safe operation of the dam by timely discovering and repairing the cracks.

[0003] At present, the detection of the dam surface cracks of a water conservancy and hydropower project concrete dam mainly relies on artificial periodic inspection, and uses devices such as a ship, a basket, a telescope and the like as a detection platform to detect by using measuring instruments or geophysical prospecting equipment such as a vernier caliper, a crack width detector, an ultrasonic detector and a ground penetrating radar. However, the method has problems such as poor real-time performance, low reliability, time-consuming and labor-intensive. With the continuous development of digital image processing technology, the method of detecting concrete dam cracks by unmanned aerial vehicle (or other carrying platform) aerial photography and crack image recognition has the characteristics of high detection efficiency, flexible range selection and objective results, and has been widely applied.

[0004] A common digital image processing detection means is to collect images based on an optical camera, and then to obtain crack recognition and detection results by image target detection and image region segmentation and the like. For example, some documents establish a crack recognition system by using digital image technology, propose a multi-directional denoising algorithm according to a gradient proportion value range, and classify and count pavement crack images by using a support vector machine method. Other documents train cracks by using a BP neural network in view of the characteristics of many noises in concrete pavement crack images, to improve the recognition rate of pavement cracks. Some documents design a crack detection and positioning system based on a convolution deep learning method, to realize crack recognition, positioning and display. Some documents consider the differences of crack images under different light conditions, propose a Faster R-CNN crack detection network based on multi-task enhancement, and integrate a ResNet-50 feature extraction network to improve crack detection accuracy. Some documents design a multi-scale fusion module based on a classic SegNet image segmentation network, to better solve the problem of large crack scale span. In addition, other documents use deep supervision measurement to guide a deep learning method to obtain and fuse crack deep semantic and shallow feature characteristics, and model crack intercommunication characteristics by a conditional random field model, to obtain good crack detection results.

[0005] The existing dam surface crack detection method based on deep learning technology generally regards the concrete crack detection as image target detection or target segmentation, and improves the algorithm or adds an image preprocessing step to adapt to the crack detection. However, this kind of method ignores the morphological characteristics of the dam crack, and is prone to cause the problems of low detection result precision of crack length, width and other properties and crack missed detection. SUMMARY

[0006] The purpose of the present application is to provide a concrete dam surface crack detection method and system based on rhombus convolution UNet, which is used to solve the technical problem of low detection precision in the prior art.

[0007] In order to solve the above technical problems, the technical scheme of the present application is as follows:

[0008] The first aspect provides a concrete dam surface crack detection method based on rhombus convolution UNet, comprising:

[0009] S1: image acquisition is performed on the concrete dam surface;

[0010] S2: the collected images are labeled to obtain an image dataset;

[0011] S3: the image dataset is divided into a training set and a validation set;

[0012] S4: a rhombus convolution Unet image segmentation network is built, wherein the segmentation network comprises an encoding part and a decoding part, the encoding part comprises four same down-sampling modules, each down-sampling module comprises once down-sampling, twice rhombus convolution operation, once rhombus convolution result splicing operation and once convolution feature fusion operation, the decoding part comprises four same up-sampling modules, each up-sampling module comprises once up-sampling, twice rhombus convolution operation, once rhombus convolution result splicing operation and once convolution feature fusion operation; there is a jump connection between the encoding part and the decoding part, and the output features of the down-sampling modules at the same level are spliced with the output features of the up-sampling modules;

[0013] S5: the training set obtained by the division is used to train the rhombus convolution Unet image segmentation network to obtain a training model;

[0014] S6: the training model is verified by using the validation set to obtain an algorithm model;

[0015] S7: the crack feature recognition is performed on the concrete surface to be detected by using the algorithm model to obtain a crack detection result.

[0016] In one embodiment, step S1 comprises:

[0017] An RGB visible light camera is carried by a UAV to perform image acquisition on the concrete dam surface.

[0018] In an embodiment, step S2 comprises:

[0019] The collected concrete dam surface images are determined manually using image processing software, the image pixels identified as cracks are marked, and the marking results are stored.

[0020] In an embodiment, in the encoding part and the decoding part of the image segmentation network of the diamond convolution Unet in S4, the two diamond convolution operations are a 45° 9x9 pixel diamond convolution and a 135° 9x9 pixel diamond convolution, the diamond convolution result splicing operation is to splice the two obtained diamond convolution results, and the convolution feature fusion operation is to perform 1x1 convolution fusion on the spliced result.

[0021] In an embodiment, step S5 uses the Adam algorithm to train the image segmentation network of the diamond convolution Unet.

[0022] In an embodiment, step S6 comprises:

[0023] On the validation set, the crack results detected by the image segmentation network of the diamond convolution Unet are compared with the manual annotation results, and it is determined whether the actual requirements are met. If the actual requirements are met, the training of the network is completed, otherwise the training model in step S5 is iterated for training.

[0024] Based on the same inventive concept, the second aspect of the present application provides a concrete dam surface crack detection system based on diamond convolution UNet, comprising:

[0025] An image acquisition module is configured to acquire images of the concrete dam surface.

[0026] A labeling module is configured to label the acquired images to obtain an image dataset.

[0027] A dataset division module is configured to divide the image dataset into a training set and a validation set.

[0028] A network building module is configured to build an image segmentation network of diamond convolution Unet, wherein the segmentation network comprises an encoding part and a decoding part, the encoding part comprises four identical down-sampling modules, each down-sampling module comprises a down-sampling, two diamond convolution operations, a diamond convolution result splicing operation and a convolution feature fusion operation, the decoding part comprises four identical up-sampling modules, each up-sampling module comprises an up-sampling, two diamond convolution operations, a diamond convolution result splicing operation and a convolution feature fusion operation; there is a skip connection between the encoding part and the decoding part, and the output features of the down-sampling module and the output features of the up-sampling module at the same level are spliced.

[0029] a training module configured to train the image segmentation network of the rhombus convolution UNet using the training set obtained by the dividing, and obtain a training model;

[0030] a verification module configured to verify the training model using a verification set, and obtain an algorithm model;

[0031] a detection module configured to use the algorithm model to identify crack features of a concrete surface to be detected, and obtain a crack detection result.

[0032] Based on the same inventive concept, the third aspect of the present application provides a computer readable storage medium, which stores a computer program, and the program is executed to implement the method of the first aspect.

[0033] Based on the same inventive concept, the fourth aspect of the present application provides a computer device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the program to implement the method of the first aspect.

[0034] Compared with the prior art, the present application has the following advantages and beneficial technical effects:

[0035] The present application discloses a concrete dam surface crack detection method based on rhombus convolution UNet, which is different from the prior art using square convolution operation to easily introduce more non-crack area background interference, thereby causing incomplete crack morphology feature recognition and low detection precision. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0037] Figure 1 is the flow chart of the concrete dam surface crack detection method based on rhombus convolution UNet provided in the present application;

[0038] Figure 2is a convolution kernel form diagram of a rhombus convolution UNet in the implementation of the present application, wherein (a) is a 9x9 pixel rhombus convolution of 45°, and (b) is a 9x9 pixel rhombus convolution of 135°;

[0039] Figure 3 is a general network structure of the rhombus convolution UNet in the implementation of the present application;

[0040] Figure 4 is an operation flow diagram of the rhombus convolution in the implementation of the present application. DETAILED DESCRIPTION

[0041] The purpose of the present application is to provide a concrete dam surface crack detection method and system based on a rhombus convolution UNet, which aims to identify more real crack morphological features by using a rhombus convolution operation on the basis of a classic UNet, thereby improving the concrete dam surface crack detection accuracy.

[0042] To make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] Embodiment one

[0044] The present embodiment provides a concrete dam surface crack detection method based on a rhombus convolution UNet, please refer to Figure 1 , which comprises the following steps:

[0045] S1: image acquisition is performed on the concrete dam surface;

[0046] S2: the collected images are labeled to obtain an image dataset;

[0047] S3: the image dataset is divided into a training set and a validation set;

[0048] S4: a rhombus convolution UNet image segmentation network is built, wherein the segmentation network comprises an encoding part and a decoding part, the encoding part comprises four same down-sampling modules, each down-sampling module comprises one down-sampling, two rhombus convolution operations, one rhombus convolution result splicing operation and one convolution feature fusion operation, the decoding part comprises four same up-sampling modules, each up-sampling module comprises one up-sampling, two rhombus convolution operations, one rhombus convolution result splicing operation and one convolution feature fusion operation; there is a skip connection between the encoding part and the decoding part, and the output features of the down-sampling modules at the same level are spliced with the output features of the up-sampling modules;

[0049] S5: training the image segmentation network of the rhombus convolution Unet using the training set obtained by the division to obtain a training model;

[0050] S6: verifying the training model using the verification set to obtain an algorithm model;

[0051] S7: using the algorithm model to perform crack feature recognition on the concrete surface to be detected to obtain a crack detection result.

[0052] Specifically, the division ratio of the training set and the verification set can be set according to actual conditions. Unet is a semantic segmentation model, and its main execution process is similar to that of other semantic segmentation models. First, convolution is used for down-sampling, then features are extracted layer by layer, and then each layer of features is used for up-sampling, and finally an image corresponding to each pixel point is obtained.

[0053] In the prior art, a deep learning algorithm (for example, UNet) using square convolution operation is generally used, but the crack shape of the concrete dam surface is generally irregular linear, and the use of square convolution operation can easily introduce more interference of the background of non-crack areas, thereby causing the problem of incomplete crack shape feature recognition and low detection accuracy. The present application proposes a concrete dam surface crack detection method based on rhombus convolution UNet, which aims to use rhombus convolution operation to identify more real crack shape features, thereby improving the detection accuracy of the concrete dam surface cracks.

[0054] In one embodiment, step S1 comprises:

[0055] An RGB visible light camera is carried by a UAV to collect images of the concrete dam surface.

[0056] In the specific implementation process, an RGB visible light camera can be carried by a UAV or other carrying platform to collect images.

[0057] If a UAV is used, the UAV platform should have the functions of controlling the flight direction and height, and the image collection device should have the functions of collecting dam surface images, real-time image transmission or offline copying. The resolution of the collected images is related to the camera shooting distance and camera performance parameters, and generally requires that the effective spatial resolution of a single pixel in the collected image is 0.1 mm to meet the detection requirements of the concrete surface crack width ≥0.15 mm.

[0058] In one embodiment, step S2 comprises:

[0059] The collected concrete dam surface image is determined by manual means using image processing software, the image pixels identified as cracks are marked, and the marking results are stored.

[0060] In the implementation process, the image processing software can be Adobe Photoshop.

[0061] In addition, each picture in the data sample can be data-augmented by a method randomly performed in four ways of vertical flipping, mirroring, changing brightness and Gaussian noise, to obtain a concrete crack data set.

[0062] In an embodiment, in the encoding part and the decoding part of the image segmentation network of the diamond convolution UNet in S4, the two diamond convolution operations are a 45° 9×9 pixel diamond convolution and a 135° 9×9 pixel diamond convolution, the diamond convolution result splicing operation is to splice the two obtained diamond convolution results, and the convolution feature fusion operation is to perform 1×1 convolution fusion on the spliced result.

[0063] As shown in Figure 2 , it is a convolution kernel form diagram of the diamond convolution UNet in the implementation of the present application, wherein part (a) is a 45° 9×9 pixel diamond convolution, and part (b) is a 135° 9×9 pixel diamond convolution.

[0064] As shown in Figure 4 , it is an operation flow diagram of the diamond convolution in the implementation of the present application.

[0065] Please refer to Figure 3 , it is a general network structure diagram of the diamond convolution UNet in the implementation of the present application. Based on the classical UNet deep learning image segmentation network framework, the encoder (Encoder) end includes four down-sampling modules. After one down-sampling operation in each down-sampling module, two diamond convolution operations are performed. The diamond convolution operation includes a 45° 9×9 pixel diamond convolution and a 135° 9×9 pixel diamond convolution. The two convolution results are spliced, and then fused by 1×1 convolution. The decoder (Decoder) end includes four up-sampling modules. After one up-sampling operation in each up-sampling module, two diamond convolution operations are performed. The diamond convolution operation includes a 45° 9×9 pixel diamond convolution and a 135° 9×9 pixel diamond convolution. The two convolution results are spliced, and then fused by 1×1 convolution. The encoder end and the decoder end have a jump connection to directly splice the encoder features (output features of the down-sampling module) and the decoder features (output features of the up-sampling module) at the same level.

[0066] In an implementation, step S5 adopts the Adam algorithm to train the image segmentation network of the diamond convolution Unet.

[0067] In an implementation, step S6 includes:

[0068] On the validation set, the crack results detected by the image segmentation network of the diamond convolution Unet are compared with the artificial labeling results to determine whether the actual requirements are met. If yes, the training of the network is completed, otherwise the training model in step S5 is iterated for training.

[0069] The method proposed in the application is described and verified below through specific examples.

[0070] 1. Collecting a concrete dam body surface crack data set, specifically:

[0071] (1) using a UAV (or other mounting platform) to mount a visible light camera, and manually controlling the UAV platform to collect concrete surface images;

[0072] (2) using a general image processing software (such as Adobe Photoshop, etc.) to manually plot cracks in the images;

[0073] (3) randomly performing data augmentation on each picture in the data sample through four methods of vertical flipping, mirroring, changing brightness and Gaussian noise, to obtain a concrete crack data set;

[0074] (4) randomly dividing the data set into two categories as a training set and a validation set, with a ratio of 70:30.

[0075] 2. Building a diamond convolution UNet crack detection network, as shown in Figure 3 , specifically:

[0076] (1) based on the classical UNet deep learning image segmentation network framework, the encoder (Encoder) end contains 4 times of downsampling, Figure 3 each convolution operation contains a 45° diamond convolution and a 135° diamond convolution, and the two convolution results are spliced and fused by 1x1 convolution;

[0077] (2) based on the classical UNet deep learning image segmentation network framework, the decoder (Decoder) end contains 4 times of upsampling, Figure 3 each convolution operation contains a 45° diamond convolution and a 135° diamond convolution, and the two convolution results are spliced and fused by 1x1 convolution;

[0078] (3) the encoder end and the decoder end have a skip connection to splice the encoder features and the decoder features;

[0079] 3. Training the target detection network model using the training set obtained by division to obtain a training model;

[0080] 4. Verifying the training model using the verification set obtained by division to obtain an algorithm model;

[0081] 5. Using the algorithm model to extract and detect the crack features of the concrete dam surface image data to be detected to obtain a crack extraction result.

[0082] As shown in Table 1, compared with the square convolution UNet method, the application has achieved remarkable results in the detection accuracy of concrete dam surface cracks.

[0083] (1) Experimental data

[0084] The experimental test data used in Table 1 is obtained from a certain dam concrete dam surface image acquired by M300RTK equipped with P1 single-lens reflex camera, a total of 524 images, each image has a resolution of 8192 pixels x 5460 pixels, the camera shooting distance is about 10 meters, and the effective spatial resolution of a single pixel in the collected image is about 0.1 mm. Adobe Photoshop software is used to manually plot the cracks in the image. All images are randomly divided into training set, verification set and test set, with a ratio of 70:20:10.

[0085] (2) Evaluation index

[0086] The precision index, also known as the precision P, is a measure of accuracy, which is equal to the number of correctly detected crack pixels divided by the number of all crack pixels detected by the algorithm.

[0087] The recall rate index, also known as the recall R, is a measure of coverage, which is represented as the proportion of actual crack pixels correctly detected.

[0088] The F1 score is a comprehensive index of precision and recall, as shown in the following formula:

[0089]

[0090] (3) Experimental results

[0091] As shown in Table 1, compared with the square convolution UNet method, the application has achieved higher detection accuracy and F1 score in the detection of concrete dam surface cracks.

[0092] Table 1 Comparison of concrete dam surface crack detection accuracy between the method of the application and the square convolution UNet method

[0093]

[0094] Embodiment Two

[0095] Based on the same inventive concept, the application discloses a concrete dam surface crack detection system based on a rhombus convolution UNet, comprising:

[0096] An image acquisition module is configured to acquire images of the concrete dam surface.

[0097] A labeling module is configured to label the acquired images to obtain an image dataset.

[0098] A dataset division module is configured to divide the image dataset into a training set and a validation set.

[0099] A network building module is configured to build an image segmentation network of the rhombus convolution UNet, wherein the segmentation network comprises an encoding part and a decoding part, the encoding part comprises four identical down-sampling modules, each down-sampling module comprises one down-sampling, two rhombus convolution operations, one rhombus convolution result splicing operation and one convolution feature fusion operation, the decoding part comprises four identical up-sampling modules, each up-sampling module comprises one up-sampling, two rhombus convolution operations, one rhombus convolution result splicing operation and one convolution feature fusion operation, there is a skip connection between the encoding part and the decoding part, and the output features of the down-sampling modules at the same level are spliced with the output features of the up-sampling modules.

[0100] A training module is configured to train the image segmentation network of the rhombus convolution UNet using the training set to obtain a training model.

[0101] A verification module is configured to verify the training model using the validation set to obtain an algorithm model.

[0102] A detection module is configured to use the algorithm model to identify crack features of a concrete surface to be detected to obtain a crack detection result.

[0103] Since the system introduced in Embodiment Two of the application is the system used in the method for detecting cracks on the surface of a concrete dam based on the rhombus convolution UNet in Embodiment One of the application, the specific structure and deformation of the system can be understood by those skilled in the art based on the method introduced in Embodiment One of the application, and thus will not be described here again. Any system used in the method in Embodiment One of the application belongs to the scope of the application.

[0104] Embodiment Three

[0105] Based on the same inventive concept, the application further provides a computer readable storage medium having a computer program stored thereon, the program being executed to implement the method described in Embodiment One.

[0106] Since the computer readable storage medium introduced in the embodiment three of the present application is the computer readable storage medium used for implementing the concrete dam surface crack detection method based on the rhombus convolution UNet in the embodiment one of the present application, therefore, based on the method introduced in the embodiment one of the present application, the person skilled in the art can understand the specific structure and deformation of the computer readable storage medium, and thus it is not repeated here. The computer readable storage medium used by the method in the embodiment one of the present application belongs to the scope of the present application.

[0107] Embodiment four

[0108] Based on the same inventive concept, the present application also provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the above-mentioned program to realize the method in the embodiment one.

[0109] Since the computer device introduced in the embodiment four of the present application is the computer device used for implementing the concrete dam surface crack detection method based on the rhombus convolution UNet in the embodiment one of the present application, therefore, based on the method introduced in the embodiment one of the present application, the person skilled in the art can understand the specific structure and deformation of the computer device, and thus it is not repeated here. The computer device used by the method in the embodiment one of the present application belongs to the scope of the present application.

[0110] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0111] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The device for implementing the functions specified in one or more flows or one or more blocks. Figure 1 The device for implementing the functions specified in one or more flows or one or more blocks.

[0112] While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the present disclosure without departing from the spirit and scope of the application. Accordingly, it is intended that the appended claims include all such variations and modifications as fall within the scope of the application. It is also to be understood that the application can be practiced with more than one embodiment. Therefore, if there is any conflict between the language of the prior art incorporated herein by reference and the language of the present application, then the language of the present application controls.

Claims

1. A method for detecting surface cracks in concrete dams based on rhombic convolution UNet, characterized in that, include: S1: Image acquisition of the concrete dam surface; S2: Annotate the collected images to obtain an image dataset; S3: Divide the image dataset into a training set and a validation set; S4: Construct a diamond convolution UNet image segmentation network, which includes an encoding part and a decoding part. The encoding part includes four identical downsampling modules. Each downsampling module contains one downsampling operation, two diamond convolution operations, one diamond convolution result concatenation operation, and one convolution feature fusion operation. The decoding part includes four identical upsampling modules. Each upsampling module contains one upsampling operation, two diamond convolution operations, one diamond convolution result concatenation operation, and one convolution feature fusion operation. There are skip connections between the encoding and decoding parts. The output features of the downsampling modules at the same level are concatenated with the output features of the upsampling modules. S5: Use the partitioned training set to train the diamond convolutional Unet image segmentation network to obtain the trained model; S6: Validate the trained model using the validation set to obtain the algorithm model; S7: Use an algorithm model to identify crack features on the concrete surface to be inspected, and obtain crack detection results; In the encoding and decoding parts of the Unet image segmentation network in S4, the two diamond convolution operations are a 45° 9×9 pixel diamond convolution and a 135° 9×9 pixel diamond convolution. The diamond convolution result concatenation operation is to concatenate the two diamond convolution results. The convolution feature fusion operation is to perform 1×1 convolution fusion on the concatenated results.

2. The method for detecting surface cracks in concrete dams based on rhombic convolution UNet as described in claim 1, characterized in that, Step S1 includes: Using a drone equipped with an RGB visible light camera, images of the concrete dam surface were collected.

3. The method for detecting surface cracks in concrete dams based on rhombic convolution UNet as described in claim 1, characterized in that, Step S2 includes: Image processing software was used to manually assess the collected images of the concrete dam surface, and the pixels identified as cracks were marked and the marking results were stored.

4. The method for detecting surface cracks in concrete dams based on rhombic convolution UNet as described in claim 1, characterized in that, Step S5 uses the Adam algorithm to train the diamond convolution Unet image segmentation network.

5. The method for detecting surface cracks in concrete dams based on rhombic convolution UNet as described in claim 1, characterized in that, Step S6 includes: On the validation set, the crack detection results obtained by the diamond convolutional Unet image segmentation network are compared with the manually labeled results to determine whether they meet the actual requirements. If they do, the network training is completed; otherwise, the training model in step S5 is iterated and trained.

6. A concrete dam surface crack detection system based on rhombic convolution UNet, characterized in that, include: Image acquisition module, used to acquire images of the surface of concrete dam; The annotation module is used to annotate the acquired images to obtain an image dataset; The dataset partitioning module is used to divide the image dataset into a training set and a validation set; The network construction module is used to build the diamond convolution UNet image segmentation network. The segmentation network includes an encoding part and a decoding part. The encoding part includes four identical downsampling modules. Each downsampling module contains one downsampling operation, two diamond convolution operations, one diamond convolution result concatenation operation, and one convolution feature fusion operation. The decoding part includes four identical upsampling modules. Each upsampling module contains one upsampling operation, two diamond convolution operations, one diamond convolution result concatenation operation, and one convolution feature fusion operation. There are skip connections between the encoding and decoding parts. The output features of the downsampling modules at the same level are concatenated with the output features of the upsampling modules. The training module is used to train the diamond convolutional Unet image segmentation network using the partitioned training set to obtain the training model; The validation module is used to validate the trained model using a validation set to obtain the algorithm model; The detection module is used to identify crack features on the concrete surface to be inspected using an algorithm model, and obtain crack detection results. In the image segmentation network encoding and decoding parts of the diamond convolution Unet in the network construction module, the two diamond convolution operations are a 45° 9×9 pixel diamond convolution and a 135° 9×9 pixel diamond convolution. The diamond convolution result stitching operation is to stitch the two diamond convolution results together. The convolution feature fusion operation is to perform 1×1 convolution fusion on the stitched results.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1 to 5.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 5.

Images