A pavement crack disease identification and segmentation method and system based on a UNet network

Abstract

The present disclosure belongs to the technical field of road crack detection, and provides a road crack disease identification and segmentation method and system based on a UNet network, comprising the following steps: collecting road surface data; constructing a UNet network model composed of a symmetrical encoder and a decoder, inputting a road surface image into the encoder, using residual hollow convolution blocks with different expansion rates and MobileOne efficient convolution blocks to perform feature extraction and output a high-dimensional feature layer, inputting the high-dimensional feature layer into the decoder, performing convolution and dense upsampling operations with the same number of layers to restore the high-dimensional effective feature layer into an image consistent with the size and dimension of the input image, and outputting the image. The present disclosure can realize high-precision identification of road crack diseases, calculate the quantitative value of the crack diseases, and solve the problems of traditional road crack detection methods, such as the need for a large amount of manual participation, a large amount of time consumption, and low efficiency.

Description

Technical Field

[0001] This disclosure relates to the field of pavement crack detection technology, specifically to a pavement crack defect identification and segmentation method and system based on UNet network. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Asphalt pavement is a common type of road surface. With increasing service life, various types of damage inevitably occur, and cracks are a common form of damage to asphalt pavement. Cracks accelerate pavement deterioration, and if not maintained promptly, they will severely impact the pavement's service life. Therefore, timely detection and maintenance of pavement cracks are extremely important.

[0004] Initially, pavement crack identification relied primarily on engineers and maintenance personnel visually observing the road surface to locate and record cracks. This process was not only time-consuming, but the quality of the results also depended heavily on the observer's experience and expertise. In the 1980s and 1990s, to improve efficiency and accuracy, various detection devices, such as cameras and laser sensors, were used to assist in manual observation and recording of cracks. These devices could continuously scan the road surface while vehicles were in motion and record the data. However, this data still required manual analysis and interpretation. With the development of computer vision technology, software for automatically identifying pavement cracks began to be developed from the 1990s to the early 2000s. This software used various image processing algorithms to automatically detect cracks from images obtained from cameras or laser scanners. However, these algorithms typically required manual parameter adjustment and had low tolerance for noise and lighting variations. From the early 2000s to the present, machine learning, especially deep learning, has been widely applied to pavement crack detection and identification. However, previous models suffer from poor segmentation capabilities and incomplete segmentation results, necessitating the development of a high-precision pavement crack defect identification and segmentation method to achieve accurate and complete pavement crack defect segmentation. Summary of the Invention

[0005] To address the aforementioned shortcomings, this disclosure proposes a method and system for identifying and segmenting pavement cracks based on the UNet network. Building upon the UNet algorithm, a high-precision optimization objective is proposed. A backbone feature extraction network is constructed by combining MobileOne efficient convolutional blocks and variable dilation rate dilated convolutional blocks to overcome image information loss caused by a fixed dilation rate. This solves the problems of poor segmentation ability and incomplete segmentation results found in previous models.

[0006] To achieve the above objectives, the present disclosure adopts the following technical solution:

[0007] The first aspect of this disclosure provides a method for identifying and segmenting pavement crack defects based on a UNet network, comprising the following steps:

[0008] Collect road surface data;

[0009] A UNet network model is constructed, which consists of a symmetrical encoder and decoder. The acquired road surface image is input into the encoder, and features are extracted using residual dilated convolutional blocks with different dilation rates and MobileOne efficient convolutional blocks. After multiple convolutions and downsampling operations, a high-dimensional feature layer is output. The high-dimensional feature layer is input into the decoder and undergoes the same number of convolutions and dense upsampling operations. Skip connections are used to establish direct connections between the same level feature layers of the encoder and decoder. The image obtained by the encoder is stacked onto the image channel of the corresponding layer of the decoder. The high-dimensional effective feature layer is restored to an image with the same size and dimensions as the input image and then output.

[0010] Preferably, the MobileOne high-efficiency convolutional block performs depthwise convolution and pointwise convolution sequentially, followed by batch normalization and activation operations to achieve upscaling and compression of the image.

[0011] Preferably, the residual dilated convolution block performs dilated convolutions on the input feature map sequentially with dilation rates of 1, 2, 5, and 7, and then adds it to the previous feature map.

[0012] Preferably, the UNet network model introduces a Ladder-ASPP structure between the encoder and decoder, which superimposes the high-dimensional feature layer after global pooling with the high-dimensional feature layer after trapezoidal dilated convolution module to achieve feature fusion, forming a new high-dimensional feature layer as the input feature map of the decoder.

[0013] Preferably, an SE attention mechanism is added after each convolutional block to build the dependency relationship between the channels.

[0014] Preferably, the following steps are also included:

[0015] The collected dataset is labeled to obtain pixel-level binary masked images of road surface cracks;

[0016] The labeled image is preprocessed to obtain a labeled image in VOC format;

[0017] The labeled images in VOC format are divided into training, validation, and test sets according to a certain ratio.

[0018] Preferably, the dense upsampling first convolves the original feature map with h×w×c into h×w×r. 2 Then reshape the output feature map to H×W;

[0019] Where H×W is the original feature map size, h×w×c is the size after convolution, and r is the dimensionality reduction ratio.

[0020] The second aspect of this disclosure provides a pavement crack defect identification and segmentation system based on a UNet network, comprising:

[0021] The data acquisition module is configured to collect road surface data.

[0022] The model building module is configured to: build a UNet network model, which consists of a symmetrical encoder and decoder. The acquired road surface image is input into the encoder, and features are extracted using residual dilated convolutional blocks with different dilation rates and MobileOne efficient convolutional blocks. After multiple convolutions and downsampling operations, a high-dimensional feature layer is output. The high-dimensional feature layer is input into the decoder and subjected to the same number of convolutions and dense upsampling operations. Skip connections are used to establish direct connections between the same level feature layers of the encoder and decoder. The image obtained by the encoder is stacked onto the image channel of the corresponding layer of the decoder. The high-dimensional effective feature layer is restored to an image with the same size and dimensions as the input image and output.

[0023] The third aspect of this disclosure provides a medium having a program stored thereon, which, when executed by a processor, implements the steps of the pavement crack defect identification and segmentation method based on UNet network described in the first aspect of this disclosure.

[0024] The fourth aspect of this disclosure provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the pavement crack defect identification and segmentation method based on UNet network described in the first aspect of this disclosure.

[0025] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0026] This disclosure combines MobileOne's efficient convolutional blocks and variable dilation rate dilated convolutional blocks to construct a pavement crack identification and segmentation network model to overcome the loss of image information caused by a fixed dilation rate. An SE attention mechanism is added after each convolutional block to build dependencies between channels and enhance the model's representational ability. Furthermore, a Ladder-ASPP structure is introduced into the model, employing a dense trapezoidal connection approach. By sequentially connecting dilated convolutional layers with different dilation rates, ASPP achieves better feature fusion. The importance of global context information in ASPP is strengthened by introducing global pooling to highlight the priority of channels containing important information. During feature extraction in the network model, dense upsampling convolutions are used instead of traditional interpolation upsampling to avoid the loss of detailed information during upsampling. This enables high-precision identification of pavement cracks and the calculation of quantitative values ​​for crack damage. Attached Figure Description

[0027] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0028] Figure 1 This is a flowchart of the first embodiment of this disclosure;

[0029] Figure 2 This is a diagram of the UNet network structure according to Embodiment 1 of this disclosure;

[0030] Figure 3 This is a network structure diagram of the high-efficiency convolutional block MobileOne of Embodiment 1 of this disclosure;

[0031] Figure 4 This is a diagram of the variable dilation rate convolutional block network structure of Embodiment 1 of this disclosure;

[0032] Figure 5 This is a network structure diagram of the SE attention mechanism in Embodiment 1 of this disclosure;

[0033] Figure 6 This is a structural diagram of the Ladder-ASPP module according to Embodiment 1 of this disclosure;

[0034] Figure 7 This is a diagram of the dense upsampling convolution structure of Embodiment 1 of this disclosure. Detailed Implementation

[0035] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0036] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0037] Where there is no conflict, the embodiments and features described herein can be combined with each other.

[0038] Example 1

[0039] like Figure 1 As shown, Embodiment 1 of this disclosure provides a method for identifying and segmenting pavement crack defects based on a UNet network, including the following steps:

[0040] Collect road surface data;

[0041] A UNet network model is constructed, which consists of a symmetrical encoder and decoder. The acquired road surface image is input into the encoder, and features are extracted using residual dilated convolutional blocks with different dilation rates and MobileOne efficient convolutional blocks. After multiple convolutions and downsampling operations, a high-dimensional feature layer is output. The high-dimensional feature layer is input into the decoder and undergoes the same number of convolutions and dense upsampling operations. Skip connections are used to establish direct connections between the same level feature layers of the encoder and decoder. The image obtained by the encoder is stacked onto the image channel of the corresponding layer of the decoder. The high-dimensional effective feature layer is restored to an image with the same size and dimensions as the input image and then output.

[0042] The collected data is crucial for deep learning algorithms. In deep learning, the model's performance and generalization ability depend on the diversity and quality of the training data. Therefore, for a specific road crack segmentation task, it is necessary to collect road crack image data. Select the road sections to be collected and utilize vehicle-mounted cameras, handheld sensors, etc., to perform high-precision image and data acquisition of the road.

[0043] The collected road surface data was labeled using Labelme software, and binary mask label images were exported.

[0044] To improve data quality, increase data diversity, reduce model complexity, accelerate training speed, and enhance model performance and effectiveness, all images were uniformly and adaptively resized to 512×512 pixels before being input into the model. Data augmentation was performed using operations such as flipping, symmetry, scaling, aspect ratio twisting, and random adjustments to brightness and contrast. Simultaneously, the binary mask labels were adjusted to VOC data format, meaning the background pixels in the label image were 0, and the crack target pixels were 1.

[0045] The preprocessed dataset is then divided into training, validation, and test sets in a 6:1:3 ratio using programming.

[0046] The specific steps for constructing the UNet network model are as follows: A pavement crack identification and segmentation network model is constructed based on residual dilated convolutions with variable dilation rates and MobileOne efficient convolutional blocks. The calculation process is shown below:

[0047] Y1 = C d＝7 (C d＝5 (C d＝2 (C d＝1 (M(X)))))+M(X) (1)

[0048] In the formula: Y1 is the output feature map, C is the dilated convolution operation, d is the dilation rate, M is the MobileOne convolution, and X is the input feature map.

[0049] The structure of Mobile One Block's high-efficiency convolutional blocks is as follows: Figure 3 As shown, the process includes sequentially performing depthwise convolution and pointwise convolution, followed by batch normalization and activation operations after depthwise convolution, and then performing batch normalization and activation operations after pointwise convolution, to achieve dimensionality upscaling and compression of the image.

[0050] like Figure 3 As shown, depthwise convolution includes three branches during training: a concatenated batch normalized deep convolution layer with a kernel size of 1, a concatenated batch normalized deep convolution layer with a kernel size of 3, and a batch normalization layer. During inference, the refactored parameter is a concatenated batch normalized deep convolution layer with a kernel size of 3. Pointwise convolution includes two branches during training: a concatenated batch normalized pointwise convolution layer with a kernel size of 1 and a batch normalization layer. During inference, the refactored parameter is a single concatenated batch normalized pointwise convolution layer with a kernel size of 1×1. During training, the three branches of the depthwise convolution are summed, passed through an activation function, and used as the input to the pointwise convolution part, which is then activated again. During inference, the depthwise convolution part is activated and then passed through an activation function as the input to the pointwise convolution part, which is then activated again.

[0051] The structure of a variable dilation rate residual dilated convolution block is as follows: Figure 4 As shown, the previous feature map is sequentially subjected to dilated convolutions with dilation rates of 1, 2, 5, and 7, and then added to the previous feature map.

[0052] An SE attention mechanism is added after each convolutional block to build dependencies between channels. The calculation formula is as follows:

[0053] Y2=σ(W2(σ(W1G(Y1)))) (2)

[0054] In the formula: Y2 is the output feature map, σ is the Sigmoid operation, W is the fully connected layer, G is the global pooling operation, and Y1 is the input feature map.

[0055] The Ladder-ASPP structure is introduced into the model, which adopts the idea of ​​dense trapezoidal connection to connect dilated convolutional layers with different dilation rates step by step to help ASPP achieve better feature fusion, strengthen ASPP's emphasis on global context information, and introduce global pooling to highlight the priority of channels containing important information.

[0056] To better understand L-ASPP, let Y3 be the output feature map, y1 be the result of the input feature map after global pooling, and y2 be the result of the input feature map after passing through the trapezoidal dilated convolution module. Then we have Y3 = y1 + y2, where y1 has the following meaning:

[0057] y1=B[C1[GP S (Y2)]]×Y2 (3)

[0058] In the formula: Y2 is the input feature map, GP S This indicates global pooling, where C1 represents the weight of each feature channel after a 1×1 convolution, and B represents the normalization of the feature weights.

[0059] Let G (K,D) For densely connected outputs, where K is the horizontal level of the pyramid and D represents the dilation rate of the dilated convolution, we have the following equation:

[0060]

[0061] In the formula:

[0062]

[0063] This indicates a channel overlay operation.

[0064] In enhancing feature extraction networks, dense upsampling convolutions are used instead of traditional interpolation upsampling to avoid the loss of detail information during the upsampling process. Assuming the original feature map size is H×W, after convolution it becomes h×w×c, where h = H / r, w = W / r, and r is the dimensionality reduction ratio. First, the original feature map is convolved from h×w×c to h×w×r. 2 Then reshape the output to H×W. The calculation formula is:

[0065] Y = R(C r (Y3)) (6)

[0066] In the formula: Y is the output feature map, R is the reshape operation, C is the convolution operation, r is the dimensionality reduction ratio, and Y3 is the input feature map.

[0067] Focal Dice Loss is used to validate the model's loss, thereby increasing the weight of positive samples and balancing the imbalance in data classification. The calculation formula is shown below:

[0068]

[0069] In the formula, |P| and |T| represent the number of elements in the segmentation result P and the true label T, respectively, |P∩T| represents the number of elements in their intersection, and Smooth is the smoothness exponent. The smaller the Dice coefficient, the larger the Dice loss, indicating that the segmentation result is less similar to the true label. Conversely, the larger the Dice coefficient, the smaller the Dice loss, indicating that the segmentation result is more similar to the true label.

[0070]

[0071] In the formula, p is the predicted probability of the model, α is the weight of the sample, and γ is the hyperparameter of Focal Loss.

[0072] By fusing Dice Loss and Focal Loss, a Focal Dice Loss model is constructed to validate the prediction results. The formula is as follows:

[0073] FocalDiceLoss=FocalLoss-βDiceLoss (9)

[0074] Where β is the weight of Dice Loss.

[0075] Upload the dataset and the built model to the server, train the model, and save the optimal parameters after iterative training to the local machine.

[0076] The original images from the test set are input into the model. The automatically identified binary crack images and the original label images are compared using a performance evaluation system that includes Intersection over Union (IoU), Pixel Accuracy (PA), Precision, Recall, and F1 Score to validate the model's performance.

[0077] Intersection over Union (IoU): IoU is the ratio of the area of ​​intersection to the area of ​​union between the predicted and actual segmentation results. It measures the degree of overlap between the model's predicted segmentation and the actual segmentation. The calculation formula is shown below:

[0078]

[0079] In the formula:

[0080] TP: True Positive Count, the number of actual positives predicted as positive;

[0081] FP: False positives, the number of cases that were actually negative but predicted to be positive;

[0082] FN: False Negative Number, the number of cases that are actually positive but predicted to be negative.

[0083] Pixel accuracy: Pixel accuracy refers to the proportion of correctly classified pixels out of the total number of pixels in a predicted segmented image. The formula for calculation is as follows:

[0084]

[0085] In the formula:

[0086] TP: True Positive Count, the number of actual positives predicted as positive;

[0087] FP: False positives, the number of cases that were actually negative but predicted to be positive;

[0088] TN: True Negative Number, the number of actual negatives predicted to be negative;

[0089] FN: False Negative Number, the number of cases that are actually positive but predicted to be negative.

[0090] Accuracy: Accuracy refers to the percentage of samples that the model predicts as positive, i.e., the proportion of true positives among those predicted as positive. A higher accuracy value indicates a higher accuracy rate in predicting positives and a lower probability of misclassifying negatives. The formula is shown below:

[0091]

[0092] In the formula:

[0093] TP: True Positive Count, the number of actual positives predicted as positive;

[0094] FP: False positives, the number of cases that were actually negative but predicted to be positive.

[0095] Recall: Recall refers to a model's ability to identify positive examples; that is, the proportion of positive examples the model successfully identifies out of all true positives. A higher recall value indicates a stronger ability to identify positive examples and a lower probability of missing them. The formula for calculating recall is shown below:

[0096]

[0097] In the formula:

[0098] TP: True Positive Count, the number of actual positives predicted as positive;

[0099] FN: False Negative Number, the number of cases that are actually positive but predicted to be negative.

[0100] F1-score: The F1-score is the harmonic mean of precision and recall, used to measure the performance of a model in classification. Its calculation formula is shown below:

[0101]

[0102] The trained model is used to segment cracks in new road surface images, and the classification result for each pixel is output.

[0103] Furthermore, it includes post-processing of the classification results for each output pixel to obtain quantified values ​​of the crack damage. For binary crack images, quantized parameters such as crack length, width, and area are calculated. The calculation formula is as follows:

[0104] Length calculation:

[0105] N = μ·A (15)

[0106] In the formula, N is the crack length, A is the number of pixels describing the crack skeleton, and μ represents the actual length represented by each pixel in the image, which is called the scaling factor.

[0107] The scaling factor μ is the actual length represented by each pixel. Given that the camera's height above the ground is h and the camera's focal length is f, the formula for calculating μ is:

[0108] μ=h / f (16)

[0109] Width calculation:

[0110] S=L×n (17)

[0111] In the formula: S is the maximum width of the crack, L is the width that shrinks after each corrosion operation, and the crack completely disappears after n corrosion operations.

[0112] Calculation of the area of ​​network cracks:

[0113] A coordinate system is constructed with the upper left corner of the binary crack as the origin, the x-axis extending horizontally to the right, and the y-axis extending vertically downwards.

[0114] Find the upper (lower) boundary by traversing all pixels from top to bottom (up). The ordinate of the first pixel with a value of 1 is the upper (lower) boundary of the crack.

[0115] Find the left (right) boundary. Traverse all pixels from left (right) to right (left). The ordinate of the first pixel with a value of 1 is the left (right) boundary of the crack.

[0116] The boundary coordinates (·, y) of the circumscribed rectangle of the network cracks are obtained. min ), (·, y max ), (x min ,·), (x max The formula for calculating the minimum bounding rectangle is as follows:

[0117] S=(y max -y min )×(x max -x min )×μ 2 (18)

[0118] Example 2

[0119] Embodiment 2 of this disclosure provides a pavement crack defect identification and segmentation system based on a UNet network, including:

[0120] The data acquisition module is configured to collect road surface data.

[0121] The model building module is configured to: build a UNet network model, which consists of a symmetrical encoder and decoder. The acquired road surface image is input into the encoder, and features are extracted using residual dilated convolutional blocks with different dilation rates and MobileOne efficient convolutional blocks. After multiple convolutions and downsampling operations, a high-dimensional feature layer is output. The high-dimensional feature layer is input into the decoder and subjected to the same number of convolutions and dense upsampling operations. Skip connections are used to establish direct connections between the same level feature layers of the encoder and decoder. The image obtained by the encoder is stacked onto the image channel of the corresponding layer of the decoder. The high-dimensional effective feature layer is restored to an image with the same size and dimensions as the input image and output.

[0122] Example 3

[0123] This disclosure provides a medium on which a program is stored. When executed by a processor, the program implements the steps in the road surface crack defect identification and segmentation method based on UNet network as described in this disclosure, in embodiment one.

[0124] The more detailed steps are the same as in Example 1, and will not be repeated here.

[0125] Example 4

[0126] This disclosure provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the road surface crack defect identification and segmentation method based on UNet network as described in this disclosure.

[0127] The more detailed steps are the same as in Example 1, and will not be repeated here.

[0128] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for pavement crack disease identification and segmentation based on a UNet network, characterized in that, Includes the following steps: Collect road surface data; A UNet network model is constructed, which consists of a symmetrical encoder and decoder. The acquired road surface image is input into the encoder, and features are extracted using residual dilated convolutional blocks with different dilation rates and MobileOne efficient convolutional blocks. After multiple convolutions and downsampling operations, a high-dimensional feature layer is output. The high-dimensional feature layer is input into the decoder and undergoes the same number of convolutions and dense upsampling operations. Skip connections are used to establish direct connections between the same level feature layers of the encoder and decoder. The image obtained by the encoder is stacked onto the image channel of the corresponding layer of the decoder. The high-dimensional effective feature layer is restored to an image with the same size and dimensions as the input image and then output. A SE attention mechanism is added after each convolutional block to build the dependencies between the channels; A road crack disease identification and segmentation network model is constructed based on a variable dilation rate residual hole convolution and a MobileOne efficient convolution block, and the calculation process is: , In the formula: Y1 is the output feature map, C is the dilated convolution operation, d is the dilation rate, M is the MobileOne convolution, and X is the input feature map; Upload the road surface dataset and the constructed UNet network model to the server, train the model, and save the optimal parameters after iterative training to the local machine. The trained UNet network model is used to segment cracks and defects in the new road surface image, and the classification result of each pixel is output. It also includes post-processing of the classification results for each pixel in the output to obtain the quantitative value of the crack disease, and for the binary crack image, calculating the quantitative parameters of crack length, width and area. The formula for calculating dense upsampling convolution is: , where Y is the output feature map, R is the reshape operation, C is the convolution operation, r is the dimension reduction ratio, and Y3 is the input feature map. The UNet network model introduces a Ladder-ASPP structure between the encoder and decoder, which superimposes the high-dimensional feature layer after global pooling with the high-dimensional feature layer after trapezoidal dilated convolution module to achieve feature fusion, forming a new high-dimensional feature layer, which serves as the input feature map of the decoder.

2. The pavement crack defect identification and segmentation method based on UNet network as described in claim 1, characterized in that, The MobileOne high-efficiency convolutional block sequentially performs depthwise convolution and pointwise convolution, followed by batch normalization and activation operations to achieve upscaling and compression of the image.

3. The pavement crack defect identification and segmentation method based on UNet network as described in claim 1, characterized in that, The residual dilated convolution block performs dilated convolutions on the input feature map sequentially with dilation rates of 1, 2, 5, and 7, and then adds it to the previous feature map.

4. The pavement crack defect identification and segmentation method based on UNet network as described in claim 1, characterized in that, It also includes the following steps: The collected dataset is labeled to obtain pixel-level binary masked images of road surface cracks; The labeled image is preprocessed to obtain a labeled image in VOC format; The labeled images in VOC format are divided into training, validation, and test sets according to a certain ratio.

5. The pavement crack defect identification and segmentation method based on UNet network as described in claim 1, characterized in that, The dense up-sampling first convolves the original feature map from h x w x c to h x w x r 2 and reshapes the output feature map into H x W. Where H×W is the original feature map size, h×w×c is the size after convolution, and r is the dimensionality reduction ratio.

6. A pavement crack defect identification and segmentation system based on a UNet network, employing the pavement crack defect identification and segmentation method based on a UNet network as described in any one of claims 1-5, characterized in that, include: The data acquisition module is configured to collect road surface data. The model building module is configured to: build a UNet network model, which consists of a symmetrical encoder and decoder. The acquired road surface image is input into the encoder, and features are extracted using residual dilated convolutional blocks with different dilation rates and MobileOne efficient convolutional blocks. After multiple convolutions and downsampling operations, a high-dimensional feature layer is output. The high-dimensional feature layer is input into the decoder and subjected to the same number of convolutions and dense upsampling operations. Skip connections are used to establish direct connections between the same level feature layers of the encoder and decoder. The image obtained by the encoder is stacked onto the image channel of the corresponding layer of the decoder. The high-dimensional effective feature layer is restored to an image with the same size and dimensions as the input image and output.

7. A medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the pavement crack defect identification and segmentation method based on UNet network as described in any one of claims 1-5.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the road surface crack defect identification and segmentation method based on UNet network as described in any one of claims 1-5.

Images