A DWT-based crack image segmentation method, system and storage medium thereof

Abstract

This invention discloses a crack image segmentation method, system, and storage medium based on Discrete Wavelet Transform (DWT). The crack image segmentation method includes the following steps: Step 1, collecting and constructing a crack image segmentation dataset; Step 2, extracting image components using DWT; Step 3, establishing a DWT neural network; Step 4, training the DWTNet and obtaining the neural network weights; Step 5, inputting the crack image into the DWTNet and outputting the image segmentation result. The technical effect of this invention is that by utilizing Discrete Wavelet Transform (DWT) to extract components in different directions and frequency domains of the image, and introducing skip-layer connections and multi-stage fusion methods to construct the DWTNet for image segmentation, the clarity of cracks after image segmentation is improved compared with existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of image information processing technology, specifically relating to a method for crack image segmentation using discrete wavelet transform (DWT), its system, and storage medium. Background Technology

[0002] With the development of computer vision technology, automated crack detection using images has become an important method in the field of civil engineering. People typically perform semantic segmentation on crack images to extract relevant parameters such as crack length and width. However, during crack image acquisition, camera shake caused by external environment or forces can lead to motion blur. Currently, most research on semantic segmentation of crack images focuses only on normal, clean crack images, and the related image segmentation techniques are relatively mature. However, research on semantic segmentation of blurred crack images remains a blind spot. Blurred images contain shadows and noise, which can easily cause large-area gaps and breaks in the crack contours identified by semantic segmentation models. Therefore, it is necessary to remove the blurring effect of crack images using deblurring algorithms, and simultaneously utilize image segmentation algorithms to extract more accurate crack contours from the restored images.

[0003] Semantic segmentation algorithms are a mature technology, and in the field of civil engineering, they are used to extract crack contour information. Common deep learning semantic segmentation algorithms typically consist of CNNs (Convolutional Neural Networks) composed of convolutional layers, which can automatically extract crack feature information. To improve semantic segmentation performance, various underlying architectures have been developed, including deep decoder and encoder structures, generative adversarial networks (GANs), and multi-scale fusion network structures, each with different image segmentation effects. However, these deep learning semantic segmentation algorithms process clear crack images. For restored crack images after deblurring, the loss of some texture and color information increases the difficulty for the image segmentation network to learn effective feature information.

[0004] Discrete wavelet transform (DWT) can decompose low-frequency components, high-frequency horizontal components, high-frequency vertical components, and high-frequency diagonal components into graphs. These graphs contain rich image information that needs to be explored. Summary of the Invention

[0005] To address the problems existing in the semantic segmentation of restored images in the prior art, the technical problem to be solved by the present invention is to provide a crack image segmentation method and system based on discrete wavelet transform, as well as a storage medium, which can effectively handle the semantic segmentation problem of crack blurred images and restored images, and obtain the best image segmentation effect.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a crack image segmentation method based on discrete wavelet transform, which includes the following steps:

[0008] Step 1: Collect and construct a crack image segmentation dataset

[0009] Collect crack images taken by drones, divide the crack images into two categories: unlabeled and labeled, and randomly select an appropriate proportion of images as the training set and test set;

[0010] Step 2: Extract image components using DWT

[0011] The image is decomposed into components with different sizes, positions and orientations by using discrete wavelet transform, resulting in low-frequency component image, high-frequency horizontal component image and high-frequency vertical component image.

[0012] Step 3: Create DWTNet

[0013] The DWT neural network contains four feature extraction paths corresponding to the low-frequency component image, the high-frequency horizontal component image, the high-frequency vertical component image, and the original image. Each feature extraction path connects to a dual convolutional module of five feature extraction stages from top to bottom. The previous feature extraction stage uses a dual convolutional module to extract features, which are then input into the next feature extraction stage. Pooling layers are connected after the first, second, and third stage dual convolutional modules in the feature extraction paths for the low-frequency component image, the high-frequency horizontal component image, and the high-frequency vertical component image. Pooling layers are connected after the first, second, third, and fourth stage dual convolutional modules in the feature extraction path for the original image.

[0014] The shallow features extracted in the first feature extraction stage of the original image are concatenated with the features of the last stage of the decoder on the right after passing through the first-level double convolution module in the middle column.

[0015] The feature information from the upper stage of the feature extraction routes of the low-frequency component image, the high-frequency horizontal component image, and the high-frequency vertical component image is fused with the feature information from the lower stage of the feature extraction route of the original image across stages and input into the middle column of the corresponding order of the double convolution module. After passing through the double convolution module, the features are concatenated with the corresponding stages of the right decoder. The last feature extraction stage directly concatenates the features of the four feature extraction routes and passes them to the right decoder for upsampling through the middle column of the final double convolution module for feature concatenation.

[0016] The right-side decoder fuses the features transmitted from the left side while performing feature decoding. In the last upsampling layer of the decoder, it fuses with the shallow feature information extracted from the original image and outputs the segmentation result of the crack image through the convolutional layer at the output end.

[0017] Step 4: DWTNet training to obtain neural network weights

[0018] Images are fed into the DWT neural network in batches to obtain the corresponding image segmentation results. The total loss value of the neural network for that batch is calculated using the loss function and the labeled image. The weights of the neural network are updated using the gradient backpropagation method. The process continues until all images have been calculated, and then the next iteration cycle begins. This process is repeated until the iteration cycle meets the requirements. Each layer and module in the neural network is trained synchronously. The network weights are saved as a weight file after each training session.

[0019] Step 5: Input the crack image into DWTNet and output the image segmentation result.

[0020] Secondly, the present invention provides a crack image segmentation system based on discrete wavelet transform, comprising the following parts:

[0021] DWT module: used to decompose the original image into low-frequency component image, high-frequency horizontal component image and high-frequency vertical component image;

[0022] DWT Neural Network: The input consists of the original image before decomposition and the low-frequency component image, high-frequency horizontal component image, and high-frequency vertical component image after decomposition. This allows the neural network model to learn and acquire richer image textures and output more effective crack information images.

[0023] The DWT neural network includes four feature extraction paths corresponding to the low-frequency component image, the high-frequency horizontal component image, the high-frequency vertical component image, and the original image. Each feature extraction path connects to a dual convolutional module of five feature extraction stages from top to bottom. The previous feature extraction stage uses a dual convolutional module to extract features, which are then input into the next feature extraction stage. Pooling layers are connected after the first, second, and third stage dual convolutional modules in the feature extraction paths of the low-frequency component image, the high-frequency horizontal component image, and the high-frequency vertical component image. Pooling layers are connected after the first, second, third, and fourth stage dual convolutional modules in the feature extraction path of the original image.

[0024] The shallow features extracted in the first feature extraction stage of the original image are concatenated with the features of the last stage of the decoder on the right after passing through the first-level double convolution module in the middle column.

[0025] The feature information from the upper stage of the feature extraction routes of the low-frequency component image, the high-frequency horizontal component image, and the high-frequency vertical component image is fused with the feature information from the lower stage of the feature extraction route of the original image across stages and input into the middle column of the corresponding order of the double convolution module. After passing through the double convolution module, the features are concatenated with the corresponding stages of the right decoder. The last feature extraction stage directly concatenates the features of the four feature extraction routes and passes them to the right decoder for upsampling through the middle column of the final double convolution module for feature concatenation.

[0026] The right-side decoder fuses the features transmitted from the left side while performing feature decoding. In the last upsampling layer of the decoder, it fuses with the shallow feature information extracted from the original image and outputs the segmentation result of the crack image through the convolutional layer at the output end.

[0027] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a computing device, causes the computing device to implement the above-described crack image segmentation method.

[0028] The technical effect of this invention is that it uses Discrete Wavelet Transform (DWT) to extract components of the image in different directions and frequency domains, and introduces skip-layer connections and multi-stage fusion methods to construct DWTNet for image segmentation. Compared with the prior art, it improves the integrity of the crack contour after image segmentation. Attached Figure Description

[0029] The accompanying drawings of this invention are described below:

[0030] Figure 1 This is a flowchart of the present invention;

[0031] Figure 2 This is a schematic diagram of the structure of a crack image segmentation system based on DWT according to the present invention;

[0032] Figure 3 These are comparison images of the restored image segmentation results of embodiments of the present invention and existing benchmark models. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0034] To clearly describe the invention, this patent application uses the directional terms "up," "down," "left," and "right" for distinction. These terms are based on the appendix. Figure 2 The orientation of the arrangement is determined by the actual direction of use of the invention. If the actual direction of use of the invention changes, the name of the orientation will change accordingly. This should not be regarded as a limitation on the scope of patent protection.

[0035] The process of this invention is as follows Figure 1 As shown, it includes the following steps:

[0036] Step 1: Collect and construct a crack image segmentation dataset

[0037] Images of cracks captured by drones were collected, and a certain number of crack images were randomly selected and labeled using the Labelme annotation tool. This resulted in two crack image datasets: one with no labels and one with labels, with a 1:1 ratio of crack images to labeled crack images. A certain number of images were randomly selected as the training and test sets, with a ratio of 8:2.

[0038] Step 2: Extract image components using DWT

[0039] like Figure 2 As shown, a DWT module decomposes an image into components of different sizes, positions, and orientations using Discrete Wavelet Transform (DWT), resulting in image components of different resolutions. After one DWT decomposition, low-frequency component images, high-frequency horizontal component images, high-frequency vertical component images, and high-frequency diagonal component images are generated. The high-frequency diagonal component image represents diagonal details of the image and contains the least effective crack information; therefore, it is removed. The low-frequency component image, high-frequency horizontal component image, and high-frequency vertical component image are used as the output of the DWT.

[0040] Step 3: Build the DWT neural network (DWTNet for short)

[0041] like Figure 2 As shown in the example, a DWT neural network is provided, which includes four feature extraction routes: corresponding to the low-frequency component map, the high-frequency horizontal component map, the high-frequency vertical component map and the original image, respectively. The feature extraction routes corresponding to each image are connected to five stages of double convolution modules from top to bottom. The previous feature extraction stage uses the double convolution module to extract features and input them to the next feature extraction stage. Figure 2 As shown in the lower right corner, the dual convolutional module consists of, in sequence, a first convolutional layer, a first regularized normalization layer, a first ReLU activation layer, a second convolutional layer, a second regularized normalization layer, and a second ReLU activation layer. To reduce image resolution, pooling layers are connected after the first, second, and third stages of the dual convolutional modules in the feature extraction paths for the low-frequency component image, the high-frequency horizontal component image, and the high-frequency vertical component image; pooling layers are connected after the first, second, third, and fourth stages of the dual convolutional modules in the feature extraction path of the original image.

[0042] According to the literature “U-Net: Convolutional Networks for Biomedical Image Segmentation”, Ronneberger, O., Fischer, P. ed., Medical image computing and computer-assisted intervention – MICCAI 2015: 18th international conference, Part III, pp. 234-241, 2015: Skip-connection (referred to as copy and crop in the paper) can acquire more effective information and obtain more accurate segmentation results.

[0043] exist Figure 2 In this process, the feature information from the upper stage of the feature extraction routes for the low-frequency component image, high-frequency horizontal component image, and high-frequency vertical component image is fused with the feature information from the lower stage of the feature extraction route for the original image across stages and input into the corresponding order of the middle column of the double convolutional module. Skip-layer connections are used to enrich the contextual feature information. For example, for the low-frequency component image, high-frequency horizontal component image, and high-frequency vertical component image... Figure 3 In the first feature extraction stage corresponding to each different component image, feature information is extracted using a dual convolution module. This extracted feature information is then fused with the feature information obtained in the second feature extraction stage corresponding to the original image to extract contextual feature information, which is learned in different feature extraction stages. The same feature fusion strategy is used in subsequent feature extraction stages, except that the first feature extraction stage of the original image and all subsequent feature extraction stages do not employ cross-stage feature fusion strategies. The features from each stage are then input into the decoder (i.e.,...). Figure 2 The right-hand side (which includes the entire output module for upsampling) decodes the extracted feature information and converts it into an output result to restore the image resolution.

[0044] During the skip connection, the obtained feature fusion information is input into the middle column of double convolutional modules, which enhances the feature fusion effect of the neural network. The obtained features are then concatenated with different upsampling stages of the decoder, allowing the model to learn more crack detail information. To improve the computational efficiency of the model, the decoder adopts a simple upsampling plus double convolution structure. In the last upsampling layer of the decoder, the shallow feature information extracted from the original image is fused to further deepen the crack contour details. Finally, the output convolutional layer ( Figure 2The last convolutional layer before the output in the upper right corner adjusts the number of channels and outputs the segmentation result of the crack image.

[0045] Figure 2 The concatenation shown refers to dimensional concatenation, also called concatenation, which connects two feature vectors along a certain dimension.

[0046] The DWTNet data stream transmission and processing process:

[0047] The low-frequency component image, high-frequency horizontal component image, high-frequency vertical component image, and the original image are input into four feature extraction paths of DWTNet. For the feature extraction path containing the three component images, features are extracted using a double-layer convolutional module in the first feature extraction stage, and simultaneously input into the next stage and the feature concatenation stage on the right. For the feature extraction path containing the original image, shallow features extracted in the first feature extraction stage using a double-layer convolutional module are simultaneously input into the next stage and, after passing through the first-level double convolutional module in the middle column, are concatenated with the last stage of the right-side decoder. Then, a max-pooling layer is used to reduce the image feature size. In the second feature extraction stage, features are extracted using a double-layer convolutional module, and its output is simultaneously input into the next stage and the upsampling and feature concatenation stage on the right. It is then concatenated with the features obtained in the first feature extraction stage of the feature extraction paths containing the three component images, completing cross-stage feature fusion. After passing through a double-layer convolutional module, it is concatenated with the corresponding stages of the right-side decoder. The subsequent three feature extraction stages all employ the same feature fusion strategy. In the final feature extraction stage, features from each route are directly concatenated and passed to the right-side decoder for upsampling via a middle column of final-level double-layer convolutional modules. The right-side decoder fuses the features transmitted from the left side while performing feature decoding, essentially a skip-layer connection, ultimately outputting the image segmentation result.

[0048] Step 4: DWTNet training to obtain neural network weights

[0049] The training process of DWTNet is as follows: Images are fed into the DWT neural network in batches to obtain the corresponding image segmentation results. The total loss value of the neural network for that batch is calculated using the loss function and the labeled images. The weights of the neural network are updated using the gradient backpropagation method. This process continues until all images have been processed, at which point the next iteration cycle begins. This update and calculation is repeated until the iteration cycle meets the requirements. Each layer and module in the neural network is trained synchronously. The network weights are saved as a weight file after each training session. Training ends when the predetermined number of iterations or accuracy requirements are met.

[0050] Convolutional layers are used to extract shallow and deep features from the image. The convolution stride is 1, 2, and 4. The specific changes in image size before and after the convolution operation are shown in the following formula:

[0051] (1)

[0052] In Equation (1), H1 is the size of the input feature map, H2 is the size of the convolution kernel, H3 is the size of the output feature map, P is the number of padding pixels, and S is the convolution kernel stride.

[0053] The feature data of the image is normalized using a batch normalization layer, and the specific formula used is as follows:

[0054] (2)

[0055] (3)

[0056] (4)

[0057] (5)

[0058] In equations (2)-(5), x i Given the input image feature map, y i The output image feature map is given by m, where m is the number of input feature channels, γ and β are additional variables involved in weight updates, and μ is the input feature map. B and σ B These are the mean and standard deviation, respectively. A very small constant to prevent the denominator from being zero.

[0059] The ReLU function is used to activate the feature layer and perform non-linear processing. The formula used is as follows:

[0060] (6)

[0061] In equation (6), x i is the input image feature map, 0 is the lower limit of the required value, and ƒ(x) is the activated image feature map.

[0062] DWTNet uses cross-entropy loss and Dice loss as the training loss functions for the DWT neural network.

[0063] The formula for calculating cross-entropy loss is:

[0064] (7)

[0065] In equation (7), It is the prediction result of the neural network, y c This represents the corresponding label value, and Loss is the training loss value.

[0066] The formula for calculating Dice loss is:

[0067] (8)

[0068] In equation (8), y i Y represents the label value corresponding to pixel i. i This represents the predicted value corresponding to pixel i, and N is the total number of pixels.

[0069] Step 5: Input the crack image into DWTNet and output the image segmentation result.

[0070] Input a dataset of crack images, train DWTNet, and record the weight parameters for each training stage. After training, load the optimal neural network weight file and input it into a test set of crack images for result evaluation. Use DWTNet for image segmentation and output the image segmentation results.

[0071] Example

[0072] The crack image dataset collected in this embodiment contains a total of 200 images, with a crack image to crack label image ratio of 1:1. The dataset is divided into training and testing sets according to an 8:2 ratio.

[0073] DWTNet is trained using the training set. In this embodiment, the Adam algorithm is used to optimize the parameters of DWTNet, with an initial learning rate of η = 0.0001. The batch size is 1, and the training lasts for 3000 epochs.

[0074] Four randomly selected blurred images containing tiny cracks were used for experimental analysis and comparison. MIMO-Unet, MIMO-Unet-Plus, MPRNet, and DeblurGAN-v2 were selected to deblur the blurred images respectively.

[0075] Image segmentation was performed on four restored images. The invention was compared with U-Net in experiments, and the image segmentation capability of DWTNet on the restored images was evaluated. Figure 3 The image of the present invention is shown in the figure. Figure 3 (b), (d), (f), (h), the images of U-Net are Figure 3 (a), (c), (e), (g). In Figure 3 The comparison shows that after image segmentation using the DWTNet of this invention, more small cracks were successfully identified.

[0076] For image 1, MIMO-Unet and MIMO-Unet-Plus perform poor deblurring, resulting in incomplete crack information and affecting subsequent image segmentation. Image segmentation using DWTNet is then performed as follows: Figure 3 As shown in (b) and (d), with U-Ne Figure 3Comparing (a) and (c), the missing part of the horizontal main crack outline in Image 1 has been filled in; Figure 3 Comparing (e) and (f), after image segmentation of the image restored by MPRNet using U-Net, the narrow crack in the red circle in Image 1 could not be successfully identified.

[0077] For the images of small cracks in Images 2 and 4, DWTNet successfully identified most of the narrow crack outlines. Crack junctions and blurred crack outlines were identified to the greatest extent possible, while U-Net struggled to identify and segment regions with sparse crack information.

[0078] After using DWTNet to segment the images, the corresponding narrow cracks were preserved to the greatest extent, and the crack outlines in images 2, 3 and 4 were more complete. The crack area on the right side of image 3 was also identified with richer outlines.

[0079] As can be seen from this embodiment, DWTNet can utilize discrete wavelet transform to obtain richer image information and achieve the best image segmentation results.

[0080] like Figure 2 As shown, the crack image segmentation system based on discrete wavelet transform provided by this invention includes the following parts:

[0081] DWT module: used to decompose the original image into low-frequency component image, high-frequency horizontal component image and high-frequency vertical component image;

[0082] DWT Neural Network: Input the original image and the decomposed low-frequency component image, high-frequency horizontal component image, and high-frequency vertical component image, so that the neural network model can learn to acquire richer image textures and output more effective crack information images.

[0083] The crack image segmentation system based on discrete wavelet transform provided by this invention can execute the crack image segmentation method provided by this invention, and has the corresponding functional modules and beneficial effects of executing this method. For details not described in the system embodiments of this invention, please refer to the description in the method invention.

[0084] The present invention also provides a computer program product comprising a computer program or computer program instructions, which, when executed by a computing device, cause the computing device to implement the crack image segmentation method provided by the present invention. The computer program product can be written in any combination of one or more programming languages ​​to perform program code for carrying out the operations of the embodiments of the present invention.

[0085] The present invention may also provide a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a computing device, cause the computing device to implement the crack image segmentation method provided by the present invention.

[0086] The crack image method provided in this embodiment of the invention may include: decomposing the original image to obtain a low-frequency component image, a high-frequency horizontal component image, and a high-frequency vertical component image; inputting the original image and the decomposed low-frequency component image, high-frequency horizontal component image, and high-frequency vertical component image, so that the neural network model learns to acquire richer image textures and outputs more effective crack information images.

[0087] Computer-readable storage media can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

Claims

1. A crack image segmentation method based on discrete wavelet transform, characterized in that, Includes the following steps: Step 1: Collect and construct a crack image segmentation dataset Collect crack images taken by drones, divide the crack images into two categories: unlabeled and labeled, and randomly select an appropriate proportion of images as the training set and test set; Step 2: Extract image components using discrete wavelet transform The image is decomposed into components with different sizes, positions and orientations by using discrete wavelet transform, resulting in low-frequency component image, high-frequency horizontal component image and high-frequency vertical component image. Step 3: Create DWTNet The DWT neural network contains four feature extraction routes corresponding to the low-frequency component map, the high-frequency horizontal component map, the high-frequency vertical component map, and the original image. The feature extraction routes of each image are connected to the dual convolutional modules of the five feature extraction stages from top to bottom. The previous feature extraction stage uses the dual convolutional module to extract features and input them into the next feature extraction stage. Pooling layers are connected after the first, second, and third stage double convolutional modules in the feature extraction routes of the low-frequency component map, the high-frequency horizontal component map, and the high-frequency vertical component map; pooling layers are connected after the first, second, third, and fourth stage double convolutional modules in the feature extraction route of the original image. The shallow features extracted in the first feature extraction stage of the original image are concatenated with the features of the last stage of the decoder on the right after passing through the first-level double convolution module in the middle column. The feature information from the upper stage of the feature extraction routes of the low-frequency component image, the high-frequency horizontal component image, and the high-frequency vertical component image is fused with the feature information from the lower stage of the feature extraction route of the original image across stages and input into the middle column of the corresponding order of the double convolution module. After passing through the double convolution module, the features are concatenated with the corresponding stages of the right decoder. The last feature extraction stage directly concatenates the features of the four feature extraction routes and passes them to the right decoder for upsampling through the middle column of the final double convolution module for feature concatenation. The right-side decoder fuses the features transmitted from the left side while performing feature decoding. In the last upsampling layer of the decoder, it fuses with the shallow feature information extracted from the original image and outputs the segmentation result of the crack image through the convolutional layer at the output end. Step 4: DWTNet training to obtain neural network weights Images are fed into the DWT neural network in batches to obtain the corresponding image segmentation results. The total loss value of the neural network for that batch is calculated using the loss function and the labeled image. The weights of the neural network are updated using the gradient backpropagation method. The process continues until all images have been calculated, and then the next iteration cycle begins. This process is repeated until the iteration cycle meets the requirements. Each layer and module in the neural network is trained synchronously. The network weights are saved as a weight file after each training session. Step 5: Input the crack image into DWTNet and output the image segmentation result.

2. The crack image segmentation method based on discrete wavelet transform according to claim 1, characterized in that: in In step 3, the dual convolutional module consists of a first convolutional layer, a first regularized normalization layer, a first ReLU activation layer, a second convolutional layer, a second regularized normalization layer, and a second ReLU activation layer, in sequence.

3. The crack image segmentation method based on discrete wavelet transform according to claim 2, characterized in that: The convolution stride of each convolutional layer is 1, 2, and 4. The image size change before and after the convolution operation is as follows: (1) In equation (1), H1 is the size of the input feature map, H2 is the size of the convolution kernel, H3 is the size of the output feature map, P is the number of padding pixels, and S is the stride of the convolution kernel. The normalization layer normalizes the feature data of the image using the following formula: (2) (3) (4) (5) In equations (1)-(5), x i Given the input image feature map, y i The output image feature map is given by m, where m is the number of input feature channels, γ and β are additional variables involved in weight updates, and μ is the input feature map. B and σ B ε represents the mean and standard deviation, respectively; ϵ is a very small constant. The ReLU function activation layer performs non-linear processing on the feature layer, using the following formula: (6) In equation (6), x i is the input image feature map, 0 is the lower limit of the required value, and ƒ(x) is the activated image feature map.

4. The crack image segmentation method based on discrete wavelet transform according to claim 1, 2 or 3, characterized in that: in In step 4, DWTNet uses cross-entropy loss and Dice loss as the training loss functions for the DWT neural network; The formula for calculating cross-entropy loss is: (7) In equation (7), It is the prediction result of the neural network, y c This represents the corresponding label value, and Loss is the training loss value; the Dice loss calculation formula is: (8) In equation (8), y i Y represents the label value corresponding to pixel i. i This represents the predicted value corresponding to pixel i, and N is the total number of pixels.

5. A crack image segmentation system based on discrete wavelet transform, characterized in that, Includes the following parts: DWT module: used to decompose the original image into low-frequency component image, high-frequency horizontal component image and high-frequency vertical component image; DWT Neural Network: Input the original image before decomposition and the low-frequency component image, high-frequency horizontal component image, and high-frequency vertical component image after decomposition, so that the neural network model can learn to acquire richer image textures and output more effective crack information images. The DWT neural network contains four feature extraction routes corresponding to the low-frequency component map, the high-frequency horizontal component map, the high-frequency vertical component map, and the original image. The feature extraction routes of each image are connected to the dual convolution modules of the five feature extraction stages from top to bottom. The previous feature extraction stage uses the dual convolution module to extract features and input them into the next feature extraction stage. Pooling layers are connected after the first, second, and third stage double convolutional modules in the feature extraction routes of the low-frequency component map, the high-frequency horizontal component map, and the high-frequency vertical component map; pooling layers are connected after the first, second, third, and fourth stage double convolutional modules in the feature extraction route of the original image. The shallow features extracted in the first feature extraction stage of the original image are concatenated with the features of the last stage of the decoder on the right after passing through the first-level double convolution module in the middle column. The feature information from the upper stage of the feature extraction routes of the low-frequency component image, the high-frequency horizontal component image, and the high-frequency vertical component image is fused with the feature information from the lower stage of the feature extraction route of the original image across stages and input into the middle column of the corresponding order of the double convolution module. After passing through the double convolution module, the features are concatenated with the corresponding stages of the right decoder. The last feature extraction stage directly concatenates the features of the four feature extraction routes and passes them to the right decoder for upsampling through the middle column of the final double convolution module for feature concatenation. The right-side decoder fuses the features transmitted from the left side while performing feature decoding. In the last upsampling layer of the decoder, it fuses with the shallow feature information extracted from the original image and outputs the segmentation result of the crack image through the convolutional layer at the output end.

6. The crack image segmentation system based on discrete wavelet transform according to claim 5, characterized in that: The dual convolutional module consists of, in sequence, a first convolutional layer, a first regularized normalization layer, a first ReLU activation layer, a second convolutional layer, a second regularized normalization layer, and a second ReLU activation layer.

7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a computing device, causes the computing device to implement the crack image segmentation method completed in steps 2 to 5 of claim 1.

Images