Bridge disease detection method based on semantic segmentation algorithm and digital image processing

Abstract

The invention discloses a bridge disease detection method based on a semantic segmentation algorithm and digital image processing, and belongs to the field of infrastructure health monitoring. The method comprises the following steps: carrying out disease detection on an image through an RT-DETR-R101 model, generating a bounding box and mask information of a disease region, reconstructing DeepLab v3 + and FCN into M-DeepNet and A-FCNet by utilizing a multi-scale feature complementation module (MFCM) and an adaptive mask optimization module (AMOM) respectively, and automatically identifying and extracting a bridge disease mask image in the image to be detected by combining an M-DeepLabNet semantic segmentation algorithm and an A-FCNet semantic segmentation algorithm. And then calculating the pixel size occupied by the disease by utilizing technologies such as digital image processing and the like, and measuring the real size of the disease by combining a pixel calibration method, thereby realizing qualitative and quantitative analysis on the bridge disease.

Description

Technical Field

[0001] This invention relates to a bridge defect detection method based on semantic segmentation and digital image processing, belonging to the field of infrastructure health inspection, and particularly to the defect detection of bridges and other structures. Background Technology

[0002] Highway bridges, as a vital transportation medium, have always played a crucial role in transportation networks. However, compared to the continuously increasing mileage of highway bridges, the management and maintenance level has not kept pace. If highway bridges are not properly and promptly maintained, various defects in bridge components, such as cracks, spalling, and steel corrosion, will severely weaken the structure's durability and adaptability, increasing the risk of bridge collapse and causing property damage and casualties. Therefore, how to promptly detect bridge defects and quantitatively analyze their severity is an urgent problem to be solved.

[0003] Traditional bridge defect detection methods primarily rely on experienced inspectors using visual inspection tools (such as scaffolding, kayaks, and bridge inspection vehicles). This method is extremely labor-intensive and time-consuming, and the safety of inspectors is compromised as bridge inspections are often conducted at heights or on water. With the development of computer vision, using computer vision technology for bridge defect detection and measurement has become a new trend. Computer vision technology is an image analysis technique that extracts defects from images through image processing operations such as threshold segmentation, edge detection, region growing, wavelet transform, and morphological processing. While computer vision technology has improved work efficiency, it still requires expert judgment and is subject to significant subjectivity. Furthermore, processing images using traditional computer vision techniques remains time-consuming. Therefore, deep learning technology has been introduced. Deep learning has developed powerful learning capabilities and proven its applicability in the image domain. The application of image classification network models, object detection network models, and semantic segmentation models has enabled the classification, localization, and pixel-level extraction of bridge defects, greatly improving detection efficiency. In practical bridge inspection work, qualitative classification and identification of defects alone, based solely on pixel size information, are insufficient to assess the extent of damage to the bridge structure. Only by obtaining the actual dimensions of the defects can the health status of the bridge structure be accurately evaluated. Therefore, how to obtain more accurate defect parameters and measure the true dimensions of the defects is an urgent problem that needs to be solved.

[0004] A typical prior art technology, CN118154864A, discloses a method, system, and storage medium for intelligent bridge defect identification based on semantic segmentation. This includes a training step and a recognition step. The training step includes: dataset preparation, dataset annotation, image preprocessing, dataset partitioning, network construction, and model training. After the training step is completed, the trained model is loaded, and the intelligent defect identification is performed using the recognition step. However, it lacks specific data processing, resulting in some shortcomings in recognition accuracy. CN116152697A discloses a three-dimensional model measurement method for concrete structure cracks, comprising: training an initial semantic segmentation model based on a labeled concrete crack training dataset to obtain a crack semantic segmentation model; using the crack semantic segmentation model to detect keyframes in acquired binocular video to obtain a crack segmentation mask; performing three-dimensional reconstruction based on the keyframes of the binocular video, the crack segmentation mask, and the camera poses corresponding to the keyframes to obtain a three-dimensional point cloud model with crack localization information; wherein, the camera poses are calculated by a binocular vision-based SLAM algorithm on the corresponding keyframes; and measuring cracks based on the three-dimensional point cloud model to obtain crack measurement data. This method aims to improve the efficiency of crack detection and measurement. However, this existing technology, based on three-dimensional modeling, also has corresponding deficiencies in accuracy.

[0005] However, the above methods have problems such as insufficient accuracy, insufficient performance, and low computational efficiency. Based on these problems, the inventors proposed a bridge defect detection method based on semantic segmentation algorithm and digital image processing. Summary of the Invention

[0006] To address the problem of obtaining the true dimensions of bridge defects, this invention provides a method, system, terminal device, and computer-readable storage medium for bridge defect detection and measurement based on semantic segmentation algorithms and computer vision technology.

[0007] In the first aspect, this invention proposes a method for bridge defect detection and measurement based on semantic segmentation algorithms and digital image processing, comprising the following steps:

[0008] Step 1: Collect images of bridge defects and damage.

[0009] Preferably, in step 1, images of bridge defects are collected, and the collection of bridge defects images is carried out using at least one or more of the following methods: using a drone to collect distant bridge images, or using manual collection of bridge deck structure images.

[0010] Preferably, manual methods mainly involve using cameras to photograph the bridge deck and the structures adjacent to it, while drones mainly capture distant views and images of the higher structures beneath the bridge deck.

[0011] Step 2: Use the RT-DETR-R101 model to detect diseases in the images acquired in Step 1, output the bounding box and mask information of the diseased area, and generate preliminary disease detection results;

[0012] Preferably, the RT-DETR-R101 model is used to detect defects in the original images captured by the UAV. The specific process is as follows: the bridge image captured by the UAV is input into the RT-DETR-R101 model, and multi-scale feature maps are extracted through the ResNet-101 backbone network to capture the local details and global context information of the defects; then, the feature maps are decoded using the Transformer decoder to generate the bounding boxes and mask information of the defect area, including the bounding box coordinates and the binary image of the mask; finally, the detection results are directly used as the training input for the subsequent M-DeepLabNet and A-FCNet semantic segmentation models, replacing the traditional manual annotation steps, reducing subjective errors and improving processing efficiency.

[0013] Step 3: Employ a class weighting strategy to balance the contribution of pixels of each class to model training, and train two semantic segmentation network models, M-DeepLabNet and A-FCNet.

[0014] Preferably, in step 3, the training data for the semantic segmentation model is directly derived from the detection results of step 2. The mask information generated by RT-DETR-R101 already includes disease region annotations, eliminating the need for additional manual annotation. The class-weight strategy addresses the low proportion of diseased pixels by dynamically adjusting the weights of the loss function using the following formula: Where, p c This represents the percentage of defective pixels in the training set.

[0015] In some embodiments, when the percentage of defective pixels is 2%, its weight ω c It is 28 times the size of the background pixels, effectively improving the model's attention to rare categories.

[0016] In some embodiments, the class weighting strategy calculates weights based on the inverse of the class pixel frequency. For classes containing only 2% to 5% of diseased pixels, their weights are set to 20 to 50 times that of background pixels to balance the model's focus on rare classes.

[0017] Preferably, during training, weight parameters are introduced into the cross-entropy loss function, as shown in the formula: Among them, w c For the weight of the category, y c For real labels, p cTo predict probabilities, M-DeepLabNet and A-FCNet employ different learning rates and batch sizes to optimize training results. M-DeepLabNet, based on DecpLab v3+, uses ResNet-101 as its backbone. It utilizes dilated spatial pyramid pooling to obtain multi-scale context, and then the Multi-Scale Feature Complementation (MFCM) module aligns deep and shallow feature channels before spatial attention fusion to highlight detailed edges such as cracks. The decoder then progressively upsamples and fuses with low-level features to output a high-detail mask. A-FCNet, on the other hand, is based on a fully convolutional network (FCN), with a backbone of either VGG16 or ResNet50. It preserves spatial information through skip connections and incorporates an adaptive mask optimization (AMOM) module at the output. This module enhances mask continuity and global consistency through boundary enhancement and gradient adjustment, suppressing oversegmentation and isolated noise. During training, the two networks are trained with different learning rates and batch sizes. M-DeepLabNet emphasizes edge accuracy, while A-FCNet prioritizes region integrity; subsequent fusion allows for complementary advantages.

[0018] Step 4: Apply high-threshold binarization to the M-DeepNet mask and low-threshold binarization to the A-FCNnet mask, then fuse them using the dot product formula: M fusion (i, j) = M M-DeepNet (i,j)·M A-FCN (i, j), where M M-DeepNet and M A-FCN These are binary masks for the two models. After fusion, the disease features of the overlapping area are preserved and isolated noise points are eliminated.

[0019] Preferably, the M-DeepLabNet mask uses a high threshold to filter edge noise, while the A-FCNet mask uses a low threshold to preserve potential defect areas. Morphological opening is then performed on the former to eliminate discrete noise, and edge erosion shrinkage is applied to the latter.

[0020] Step 5: Skeletonize the fusion result, fill the contour gaps using morphological closing operations, and calculate the pixel values ​​of the area, length, and width of the lesion through connected component analysis to obtain the true size of the lesion.

[0021] Preferably, in order to calculate the area of ​​the defects, the mask of the defect identification result is first binarized, and then the number of bridge defects and their respective area size are determined by using connected component extraction.

[0022] Preferably, to calculate the length and width of the lesion, the binarized image after recognition is skeletonized using the Zhang-Suen algorithm. Edge pixels are iteratively deleted until a lesion centerline with a single pixel width is obtained, and morphological thinning is alternately applied to smooth the skeleton. Subsequently, connected component analysis is performed on the skeletonized image to remove noise regions with an area less than 50 pixels, and contour gaps are filled by morphological closing operations. The first image after recognition is fused with the skeletonized second image of the first image to obtain a third image. Then, the lesion length is calculated based on the connected component statistics of the lesion area (total number of pixels) and the accumulated skeletonized pixels, and the distance of the maximum coverage pixels is measured along the normal direction of the skeleton line to determine the lesion width.

[0023] The above method can only obtain the numerical value of the lesions in pixels, but cannot directly obtain the actual size of the lesions. Therefore, a reference object method is used to convert the pixel size of the lesions in the image into the corresponding actual size. The target of the same image should not only include the lesions, but also a reference object of known size (such as a calibration plate with a side length of 10cm). The corresponding expression is:

[0024] W = λw i

[0025]

[0026] Where: W is the target disease calibration size;

[0027] w i - The pixel value of the target lesion in the image;

[0028] λ - The scaling factor of the target lesion in the image;

[0029] D - Reference object calibration dimensions;

[0030] d - The pixel value of the reference object in the image.

[0031] Secondly, the present invention also provides a bridge defect detection system, which performs bridge detection based on the bridge defect detection and measurement method described in the first aspect using semantic segmentation algorithm and digital image processing.

[0032] Thirdly, the present invention also provides a terminal device, which uses the bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing described in the first aspect to perform bridge detection.

[0033] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing as described in the first aspect.

[0034] The beneficial effects of this invention are as follows:

[0035] This invention introduces the RT-DETR-R101 detection model to achieve automated detection and mask generation of bridge defect areas, replacing the traditional manual annotation steps and significantly reducing data processing time. At the same time, through the efficient collaboration of ResNet-101 multi-scale feature extraction and Transformer decoder, it ensures the accurate output of defect bounding boxes and mask information, providing high-quality annotation input for subsequent semantic segmentation models.

[0036] Building upon this foundation, the traditional DeepLab v3+ and FCN models are improved into M-DeepLabNet and A-FCNet respectively using the Multi-Scale Feature Complementary Module (MFCM) and Adaptive Mask Optimization Module (AMOM) to enhance detail boundary extraction and global structure awareness. Furthermore, a class-balanced strategy is employed to train the M-DeepLabNet and A-FCNet models separately. M-DeepLabNet excels at extracting edge details, while A-FCNet optimizes global mask continuity; the fusion of the two compensates for the over-segmentation and under-detection defects inherent in single models.

[0037] By employing a differentiated thresholding strategy, dot-multiplication fusion, Zhang-Suen skeletonization, and closing operation repair, the accuracy of defect size extraction is significantly improved. The error is significantly reduced compared to traditional methods, and the true defect size is obtained by combining reference calibration technology. This method improves the efficiency of defect detection, yields more accurate pixel parameters for bridge defects, and thus achieves more reliable measurement accuracy of the true defect size. Attached Figure Description

[0038] Figure 1 A flowchart of an embodiment of the present invention;

[0039] Figure 2 One embodiment of the present invention involves the use of a drone to collect images of diseased areas;

[0040] Figure 3 One embodiment of the present invention utilizes the RT-DETR-R101 model for preliminary detection of diseases;

[0041] Figure 4 One embodiment of the present invention uses a class-balanced strategy for network training;

[0042] Figure 5 A network identification result according to an embodiment of the present invention;

[0043] Figure 6In one embodiment of the present invention, the mask images identified by M-DeepLabNet and A-FCNet are used to obtain the fusion result:

[0044] Figure 7 Image preprocessing for crack width calculation according to one embodiment of the present invention. Detailed Implementation

[0045] The specific embodiments of the present invention are described in detail below with reference to the technical solutions and accompanying drawings.

[0046] Example 1

[0047] Reference Appendix Figure 1-7 A bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing includes the following implementation steps: Step 1, collecting images of bridge defects using a combination of manual and drone methods. Apparent defects in bridges may appear in various parts of bridge components, therefore comprehensive image collection of bridge components is essential. Manual methods mainly use cameras to photograph the bridge deck and structures adjacent to it, while drones mainly capture distant views and images of the higher structures below the bridge deck.

[0048] Step 2: Use the RT-DETR-R101 model to detect diseases in the images acquired in Step 1, output the bounding box and mask information of the diseased area, and generate preliminary disease detection results.

[0049] The RT-DETR-R101 model is used to detect bridge defects in raw images captured by a drone. The specific process is as follows: The bridge image captured by the drone is input into the RT-DETR-R101 model. Multi-scale feature maps are extracted using the RcsNet-101 backbone network to capture local details and global contextual information of the defects. Subsequently, a Transformer decoder is used to decode the feature maps, generating bounding boxes and mask information for the defect areas (e.g., ...). Figure 3 The detection results include bounding box coordinates and mask binary images. Finally, the detection results are directly used as the training input for subsequent M-DeepLabNet and A-FCNet semantic segmentation models, replacing the traditional manual annotation step, reducing subjective errors and improving processing efficiency.

[0050] Step 3: Employ a class weighting strategy to balance the contribution of pixels of each class to model training, and train two semantic segmentation network models, M-DeepLabNet and A-FCNet.

[0051] In step 3, the training data for the semantic segmentation model is directly derived from the detection results of step 2. The mask information generated by RT-DETR-R101 already includes disease region annotations, eliminating the need for additional manual annotation. The class-weighted strategy addresses the low proportion of diseased pixels by dynamically adjusting the weights of the loss function using the following formula: Where, p c This represents the percentage of defective pixels in the training set.

[0052] The class weighting strategy is based on calculating weights using the reciprocal of the class pixel frequency. For classes containing only 2%–5% of diseased pixels, their weights are set to 20–50 times that of background pixels to balance the model's focus on rare classes. During training, weight parameters are introduced into the cross-entropy loss function, as shown in the formula: Among them, w c For the weight of the category, y c For real labels, p c To predict probabilities, M-DeepLabNet and A-FCNet employ different learning rates and batch sizes to optimize training results. M-DeepLabNet, based on DeepLab v3+, uses ResNet-101 as its backbone. It utilizes dilated spatial pyramid pooling to obtain multi-scale context, and then the Multi-Scale Feature Complementation Module (MFCM) aligns deep and shallow feature channels before spatial attention fusion to highlight detailed edges such as cracks. The decoder then progressively upsamples and fuses with low-level features to output a high-detail mask. A-FCNet, on the other hand, is based on a fully convolutional network (FCN), with a backbone of either VGG16 or ResNet50. It preserves spatial information through skip connections and incorporates an adaptive mask optimization module (AMOM) at the output. This module enhances mask continuity and global consistency through boundary enhancement and gradient adjustment, suppressing oversegmentation and isolated noise. During training, the two networks are trained with different learning rates and batch sizes. M-DeepLabNet emphasizes edge accuracy, while A-FCNet prioritizes region integrity; subsequent fusion allows for complementary advantages.

[0053] Step 4: Extract the mask images output by M-DeepLabNet and A-FCNet, perform differential binarization processing on each, and fuse them through pixel-by-pixel dot multiplication. The M-DeepLabNet mask uses a high threshold to filter edge noise, while the A-FCNet mask uses a low threshold to preserve potential defect areas. Then, morphological opening is performed on the former to eliminate discrete noise, and edge erosion and shrinkage processing is applied to the latter. Finally, the two mask images are multiplied bitwise to obtain the fused image, where the dot multiplication formula is: M... fusion (i, j) = M M-DeepNet (i,j)·M A-FCNet (i, j), where M M-DeepNet and M A-FCNetThese are binary masks for the two models. After fusion, the disease features of the overlapping areas of the two models are preserved, and isolated noise points are eliminated.

[0054] Depend on Figure 4 The detection results clearly show that M-DeepLabNet excels in both general outline and detail, but some interference points around the image affect the recognition results. While A-FCNet identifies the lesion as a whole, the lesion outlines are smoothly transitioned, and the recognition range is significantly larger than the original image. To avoid the shortcomings of both while retaining their respective advantages, mask images of the recognition results of both can be extracted and subjected to differentiated binarization processing: a high threshold is used to filter edge noise for the M-DeepLabNet mask, and a low threshold is used for the FCN mask to retain potential lesion areas. Then, morphological opening is performed on the M-DeepLabNet mask to eliminate discrete noise; edge shrinkage (erosion operation) is performed on the A-FCNet mask to reduce over-recognition areas. Finally, the processed masks are combined according to formula M... fusion (i, j) = M M-DeepNet (i,j)·M A-FCNet (i, j) are multiplied pixel by pixel, preserving the disease features of overlapping areas and eliminating isolated points to obtain the fusion result, such as... Figure 6 .

[0055] Step 5: Skeletonize the fusion result, fill the contour gaps using morphological closing operations, and calculate the pixel values ​​of the area, length, and width of the lesion through connected component analysis. Then, use the reference method to calibrate the pixels and obtain the true size of the lesion.

[0056] To calculate the area of ​​bridge defects, the defect identification result is first binarized using a mask. Then, connected component extraction is used to determine the number of defects and their respective areas. To calculate the length and width of defects, the binarized image is skeletonized using the Zhang-Suen algorithm. Edge pixels are iteratively removed until a single-pixel-width centerline of the defect is obtained, and morphological thinning is alternately applied to smooth the skeleton. Subsequently, connected component analysis is performed on the skeletonized image to remove noise regions smaller than 50 pixels, and morphological closing operations are used to fill in the contour gaps. Figure 7 (b) Image 7(d) is obtained by fusing the skeletonized image 7(c) with the image. Then, the disease length is calculated by statistically analyzing the disease area (total number of pixels) based on the connected components and accumulating the skeletonized pixels. The distance of the maximum covered pixels is measured along the normal direction of the skeleton line to determine the disease width.

[0057] The above method can only obtain the numerical value of the lesions in pixels, but cannot directly obtain the actual size of the lesions. Therefore, a reference object method is used to convert the pixel size of the lesions in the image into the corresponding actual size. The target of the same image should not only include the lesions, but also a reference object of known size (such as a calibration plate with a side length of 10cm). The corresponding expression is:

[0058] W = λw i

[0059]

[0060] Where: W - the target disease calibration size;

[0061] w i - The pixel value of the target lesion in the image;

[0062] λ - The scaling factor of the target lesion in the image;

[0063] D - Reference object calibration dimensions;

[0064] d - The pixel value of the reference object in the image.

[0065] In some embodiments, the present invention also claims protection for a bridge defect detection and measurement system based on semantic segmentation algorithms and digital image processing.

[0066] In some embodiments, the present invention also claims protection for a terminal device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing in some embodiments.

[0067] In some embodiments, the present invention also claims protection for a computer-readable storage medium having computer instructions stored thereon that, when executed, perform a bridge defect detection and measurement method based on semantic segmentation algorithms and digital image processing in some embodiments, for detecting and measuring structural defects.

[0068] The above embodiments are merely preferred exemplary embodiments of this application, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any development or substitution that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention shall fall within the protection and disclosure scope of the present invention.

Claims

1. A method for bridge defect detection and measurement based on semantic segmentation algorithm and digital image processing, characterized in that: Includes the following steps: Step 1: Acquire images of bridge defects and damage; Step 2: Use the RT-DETR-R101 model to detect diseases in the images acquired in Step 1, output the bounding boxes and mask information of the diseased areas, and generate preliminary disease detection results; Step 3: Employ a class weighting strategy to balance the contribution of pixels from each class to model training, and train two semantic segmentation network models: M-DeepLabNet and A-FCNet. Step 4: Apply high-threshold binarization to the M-DeepNet mask and low-threshold binarization to the A-FCNnet mask, then fuse them using the dot product formula: M fusion (i, j) = M M-DeepNet (i,j)·M A-FCN (i, j), where M M-DeepNet and M A-FCNet These are binary masks for the two models. After fusion, the disease features of the overlapping area are preserved and isolated noise points are eliminated. Step 5: Skeletonize the fusion result, fill the contour gaps using morphological closing operations, and calculate the pixel values ​​of the area, length, and width of the lesion through connected component analysis. Then, use the reference method to calibrate the pixels and obtain the true size of the lesion.

2. The bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing according to claim 1, characterized in that: In step 1, images of bridge defects are collected. The collection of bridge defects images can be done in at least one or more of the following ways: using drones to collect distant bridge images, or using manual collection of bridge deck structure images.

3. The bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing according to claim 2, characterized in that: In step 2, the bridge image captured by the UAV is input into the RT-DETR-R101 model, and multi-scale feature maps are extracted through the ResNet-101 backbone network to capture the local details and global context information of the disease. The feature map is decoded using the Transformer decoder to generate the bounding box and mask information of the diseased area, including the bounding box coordinates and the binary mask image. The detection results are directly used as the training input for the subsequent semantic segmentation model.

4. The bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing according to claim 1, characterized in that: In step 3, during training, weight parameters are introduced into the cross-entropy loss function, with the following formula: Among them, w c For the weight of the category, y c For real labels, p c To predict probabilities.

5. The bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing according to claim 1, characterized in that: In step 4, the M-DeepLabNet mask uses a high threshold to filter edge noise and performs morphological opening operations to eliminate discrete noise, while the A-FCNet mask uses a low threshold to preserve potential defect areas for edge corrosion shrinkage processing.

6. The bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing according to claim 1, characterized in that: In step 5, in order to calculate the area of ​​the defects, the mask of the defect identification result is first binarized, and then the number of bridge defects and their respective area size are determined by using connected component extraction.

7. The bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing according to claim 6, characterized in that: In step 5, to calculate the length and width of the lesion, the identified binarized image is skeletonized using the Zhang-Suen algorithm. Edge pixels are iteratively removed until a lesion centerline with a single pixel width is obtained, and morphological thinning is alternately applied to smooth the skeleton. Subsequently, connected component analysis is performed on the skeletonized image to remove noise regions with an area less than 50 pixels, and contour gaps are filled through morphological closing operations. The identified first image is fused with the skeletonized second image of the first image to obtain a third image. The lesion area is calculated based on the connected component statistics, the lesion length is calculated by accumulating the skeletonized pixels, and the maximum coverage pixel distance is measured along the normal direction of the skeleton line to determine the lesion width.

8. The bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing according to claim 7, characterized in that: The pixel size of the lesion is obtained by using a reference method for pixel calibration.

9. A bridge defect detection system, characterized in that... A bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing, as described in any one of claims 1-8, is used for bridge detection.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the bridge defect detection and measurement method based on semantic segmentation algorithm and digital image processing as described in any one of claims 1 to 8.

Images