Bridge concrete crack full-process intelligent detection method based on deep neural network

Abstract

The application discloses a bridge concrete crack full-process intelligent detection method based on a deep neural network, and comprises the following steps: step one: a filtering stage: inputting a to-be-detected image into an image filtering neural network Crack-Filter to remove false crack images in the to-be-detected image and obtain true crack images; step two: a positioning stage: inputting the true crack images into a semantic segmentation neural network Tres-net to obtain crack segmentation skeleton feature maps, realize pixel-level positioning of the cracks, and provide crack information without interference for subsequent stages; and step three: a classification stage: inputting the crack segmentation skeleton feature maps into a crack image classification neural network Crack-Class to obtain category information of the cracks and use the category information for crack diagnosis. The application provides a full-process automatic, intelligent and full-process solution for engineering detection of bridge cracks in view of the problem that most crack detection methods at present only focus on a single function of crack segmentation and cannot meet the needs of bridge crack detection in actual engineering, and has important engineering strategic significance.

Description

Technical Field

[0001] This invention belongs to the field of crack detection technology, and is a method for intelligent detection of bridge concrete cracks throughout the entire process based on deep neural networks. Background Technology

[0002] Concrete bridges constitute a large proportion of bridges and have long service lives. Government agencies worldwide allocate substantial funds annually to inspect and maintain the health of bridges. Concrete cracks are the starting point for many bridge defects, significantly reducing their lifespan. Effective early identification of cracks and timely implementation of appropriate maintenance measures based on their damage type can significantly reduce maintenance costs and extend the bridge's lifespan.

[0003] Currently, bridge crack detection mainly relies on manual inspection. This manual method has numerous drawbacks, including low efficiency, subjective results, and numerous safety risks. While some traditional detection methods have automated crack skeleton extraction, improving detection speed and eliminating the need for manual inspection, digital optics-based detection techniques such as threshold segmentation algorithms, crack edge-based detection algorithms, and median filtering image enhancement methods still require cumbersome manual operation and suffer from unstable performance in complex engineering environments, struggling to overcome interference from shadows, stains, and other noise. In recent years, the rapid development of machine learning in image processing has provided valuable experience for crack detection. Deep neural networks have begun to be applied to solve various crack detection problems; networks such as Region-based CNN (R-CNN) and You Only Look Once (YOLO) use sliding window techniques to locate cracks in images, selecting rectangular regions containing crack pixels. In practical engineering, extracting the crack skeleton from the background can more directly obtain crack information than locating the crack image bounding box. Semantic segmentation deep neural networks such as U-net and SegNet have been applied to crack detection, realizing pixel-level learning of crack features, which significantly improves the accessibility, computing speed and prediction accuracy of crack detection methods.

[0004] Currently, the crack detection process widely used in actual engineering is mainly divided into three stages: (1) Crack detection: Select images containing true cracks (distinguishing them from false cracks such as uneven shadows, stains, deformation joints, and expansion joints) for subsequent analysis of the location of the defective components in the structure; (2) Crack location: Identify and locate the skeleton of true cracks for subsequent research on the degree of the defect; (3) Crack diagnosis: Assess the degree of crack development and specific category and take targeted maintenance measures. At present, bridge crack research based on deep learning focuses on the "crack location" stage, which is far from sufficient for actual engineering. The images to be processed in the location stage are often mixed with a large number of highly interfering false crack images. The analysis in the "crack diagnosis" stage requires information support on the type of crack. Most crack detection methods only focus on a single function and cannot fully meet the engineering needs of different stages of actual projects. Actual bridge concrete crack detection engineering requires a fully intelligent processing method. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a fully intelligent detection method for bridge concrete cracks based on deep neural networks. This method addresses the problem that most current crack detection methods only focus on the single function of crack segmentation, which cannot meet the needs of bridge crack detection in actual engineering. It provides a fully automated, intelligent, and end-to-end solution for the engineering detection of bridge cracks, and has important strategic engineering significance.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention first proposes a method for intelligent detection of bridge concrete cracks throughout the entire process based on deep neural networks, including the following steps:

[0008] Step 1: Filtering Stage

[0009] The image to be detected is input into the image filtering neural network Crack-Filter to remove false crack images from the image to be detected and obtain the real crack image;

[0010] Step Two: Positioning Phase

[0011] The real crack image is input into the semantic segmentation neural network Tres-net to obtain the crack segmentation skeleton feature map, which realizes the pixel-level localization of cracks and provides interference-free crack information for subsequent stages.

[0012] Step 3: Classification Stage

[0013] The crack segmentation skeleton feature map is input into the crack image classification neural network Crack-Class to obtain crack category information for crack diagnosis.

[0014] Furthermore, in step one, the image filtering neural network Crack-Filter includes:

[0015] 1×1 convolutional layer: used to perform dimensionality upscaling on the input image;

[0016] Bottleneck block: Used to extract crack features;

[0017] SE attention mechanism: Added to the bottleneck block to focus on more important crack feature information;

[0018] The SE attention mechanism includes:

[0019] 1×1 convolutional layer: used to calculate local feature statistics without changing the spatial structure of the image;

[0020] The Squeeze part is used to generate a feature representation for each channel in the spatial dimension by global average pooling.

[0021] The excitation section includes two 1×1 convolutional layers. The first 1×1 convolutional layer is followed by a Swish-B activation function, and the second 1×1 convolutional layer is followed by a Sigmoid function. These two 1×1 convolutional layers are used to map the dimensionality-reduced representation back to the original number of channels. The Swish-B activation function is expressed as follows:

[0022]

[0023] Where: β is a learnable and trainable parameter;

[0024] The Sigmoid function is used to generate a weight vector, which is then multiplied by the input to enhance the effective feature information in the channel.

[0025] Furthermore, the training method for the image filtering neural network Crack-Filter is as follows:

[0026] A crack presence dataset, including both real and fake crack images, was constructed to train the image filtering neural network Crack-Filter. The loss function used during training was the binary cross-entropy function.

[0027]

[0028] Among them, L BCE Represents the binary cross-entropy loss function; x n Indicates the predicted pixel output, y n Indicates with x n The corresponding label; n represents the total number of samples;

[0029] The training of the image filtering neural network Crack-Filter is completed when the loss function converges, and the optimal weight coefficients obtained from the training are used as the weight coefficients of the image filtering neural network Crack-Filter.

[0030] Furthermore, the semantic segmentation neural network Tres-net includes an encoder and a decoder;

[0031] The encoder includes a structured residual block Red-X and an MVB module; the structured residual block Red-X is used to linearly combine the three input channels, which utilizes the high computational performance of dense matrices while ensuring the sparsity of the network structure; the MVB module is used to encode global information and autonomously learn the key regions of the crack.

[0032] The encoder performs downsampling feature extraction during encoding; the final result of the encoder's downsampling is fused using an attention-drilled pyramid module for multi-scale feature fusion; after each step of the encoder's downsampling result is passed sequentially through a channel attention module and a spatial attention module, it is fused with the result of each step of the decoder's upsampling result through a gated attention unit to determine the importance of each pixel, locate the region to be paid attention to, and suppress the feature response of irrelevant regions; the final result of the feature fusion of the encoder's downsampling and the decoder's upsampling is used to generate a crack segmentation skeleton feature map through two convolution operations.

[0033] Furthermore, the training method for the semantic segmentation neural network Tres-net is as follows:

[0034] Real-world crack images were acquired and crack locations were marked to obtain training labels for the semantic segmentation neural network Tres-net;

[0035] The original image was processed using the Clahe image enhancement method to enhance local details of the cracks, suppress noise introduced by image enhancement, and improve the before-and-after contrast.

[0036] The crack and label images are standardized to form a crack segmentation dataset;

[0037] The semantic segmentation neural network Tres-net was trained using a crack segmentation dataset, with the Crack_BD_Loss loss function used during training.

[0038] Crack_BD_Loss = 0.2 × L Dice (x n ,y n )+0.8×L BCE (x n ,y n )

[0039]

[0040] Where: Crack_BD_Loss is the Crack_BD_Loss loss function; x n Indicates the predicted pixel output; y n Indicates with x n Corresponding label; L Dice It is the Dice coefficient loss function; L BCE It is a binary cross-entropy loss function;

[0041] When the loss function converges, the training of the semantic segmentation neural network Tres-net is completed, the optimal weight coefficients are obtained, and the optimal weight coefficients are used as the weight coefficients of the semantic segmentation neural network Tres-net.

[0042] Furthermore, the training method for the crack image classification neural network Crack-Class is as follows:

[0043] Strip cracks, block cracks, and mesh cracks were selected from the crack segmentation skeleton feature map to obtain a crack classification dataset, and cross-validation was used to ensure the accuracy of crack classification.

[0044] Image enhancement techniques were used to expand the crack classification dataset;

[0045] The Crack-Class neural network for crack image classification was trained using the expanded crack classification dataset. The loss function used during training was the multivariate cross-entropy loss function.

[0046]

[0047] Among them, L CE This represents the multivariate cross-entropy loss function; N is the number of categories; y i It is the one-hot encoding of the real label, representing the real situation of the i-th category, p i It is the probability that the model predicts for the i-th category;

[0048] When the loss function converges, the training of the crack image classification neural network Crack-Class is completed, the optimal weight coefficients are obtained, and the optimal weight coefficients are used as the weight coefficients of the crack image classification neural network Crack-Class.

[0049] The beneficial effects of this invention are as follows:

[0050] This invention provides an intelligent detection method for the entire process of bridge concrete cracks based on deep neural networks, which has the following technical advantages:

[0051] (1) It can accurately filter, locate, and classify cracks in images;

[0052] (2) A Crack-Filter image filtering neural network based on MobileNetV2 was built, and an SE attention mechanism was added. While being lightweight, it can distinguish and filter various misleading non-crack images.

[0053] (3) A semantic segmentation neural network Tres-net with fast processing speed, strong generalization ability and robustness was built, which can realize the segmentation of cracks of various materials and widths in different complex scenarios.

[0054] (4) A method for classifying cracks in concrete bridges was summarized and proposed. A crack image classification neural network, Crack-Class, was built to achieve accurate and automated classification of bridge cracks.

[0055] This invention presents a fully automated, intelligent, and end-to-end intelligent method for detecting bridge concrete cracks based on deep neural networks. It addresses the problem that most current crack detection methods only focus on the single function of crack segmentation, which cannot meet the needs of bridge crack detection in actual engineering. This invention provides a fully automated, intelligent, and end-to-end solution for the engineering detection of bridge cracks, and has significant strategic engineering implications. Attached Figure Description

[0056] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0057] Figure 1 This is a flowchart of the intelligent detection method for bridge concrete cracks based on deep neural networks according to the present invention.

[0058] Figure 2 This is a diagram of the architecture of a three-stage deep neural network.

[0059] Figure 3 A schematic diagram of the SE attention mechanism proposed for the filtering stage;

[0060] Figure 4 A schematic diagram of the module structure proposed for the segmentation stage; (a) is the structured residual block Red-X; (b) is the attention void pyramid ASEP module;

[0061] Figure 5 A schematic diagram of the gated attention unit technology used in the segmentation phase;

[0062] Figure 6 Images from the crack existence dataset;

[0063] Figure 7 The image is processed using the ClAHE image enhancement method;

[0064] Figure 8(a) is the pixel-level semantic segmentation result of the crack; (b) is the original crack image; (c) is the crack label image; (d) is the model recognition result image.

[0065] Figure 9 Images from the expanded crack classification dataset;

[0066] Figure 10 The results of the crack filtering and classification tests are shown in Figure 1; (a) shows the results of the crack filtering test; (b) shows the results of the crack classification test. Detailed Implementation

[0067] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0068] like Figure 1 As shown in the figure, the intelligent detection method for the entire process of bridge concrete cracks based on deep neural networks in this embodiment includes the following steps.

[0069] Step 1: Filtering Stage

[0070] The image to be detected is input into the image filtering neural network Crack-Filter to remove false crack images from the image to be detected and obtain the true crack image.

[0071] Specifically, in this embodiment, the image filtering neural network Crack-Filter combines the MobileNetV2 network and an improved SE attention mechanism, featuring an inverse residual structure that first increases dimensionality, performs convolution, and then reduces dimensionality. Specifically, as... Figure 2 As shown, the image filtering neural network Crack-Filter in this embodiment includes:

[0072] 1×1 convolutional layer: used to perform dimensionality upscaling on the input image;

[0073] Bottleneck block: Extract crack features through convolution operations;

[0074] SE attention mechanism: Added to the bottleneck block to focus on more important crack feature information, effectively suppressing the influence of interference noise;

[0075] In this embodiment, there are 15 bottleneck blocks.

[0076] like Figure 3 As shown, the SE attention mechanism in this embodiment includes:

[0077] 1×1 convolutional layer: used to calculate local feature statistics without changing the spatial structure of the image;

[0078] The Squeeze part is used to generate a feature representation for each channel in the spatial dimension by global average pooling.

[0079] The excitation section includes two 1×1 convolutional layers. Considering the deep stacking of the 15 bottleneck blocks in the Crack-Filter image filtering neural network, this embodiment connects the Swish-B activation function after the first 1×1 convolutional layer to replace the ReLU activation function. The Swish-B activation function can achieve smoother gradient changes at deeper network layers and does not involve complex exponential operations, thus improving the efficiency and accuracy of the model. This embodiment connects the Sigmoid function after the second 1×1 convolutional layer; the two 1×1 convolutional layers are used to map the dimensionality-reduced representation back to the original number of channels; specifically, the Swish-B activation function is expressed as:

[0080]

[0081] Where β is a learnable and trainable parameter.

[0082] In this embodiment, the Sigmoid function is used to generate a weight vector, which is then multiplied by the input to enhance the effective feature information on the channel.

[0083] In this embodiment, the training method for the image filtering neural network Crack-Filter is as follows:

[0084] A crack presence dataset, including images of real and fake cracks (such as uneven shadows, stains, expansion joints, and deformation joints), was constructed to train the image filtering neural network Crack-Filter. Specifically, the images of real and fake cracks can be collected in the field or derived from open-source datasets online. The loss function used during training is the binary cross-entropy function.

[0085]

[0086] Among them, L BCE Represents the binary cross-entropy loss function; x n Indicates the predicted pixel output, y n Indicates with x n The corresponding label; n represents the total number of samples;

[0087] The training of the image filtering neural network Crack-Filter is completed when the loss function converges, and the optimal weight coefficients obtained from the training are used as the weight coefficients of the image filtering neural network Crack-Filter.

[0088] Step Two: Positioning Phase

[0089] The true crack image is input into the semantic segmentation neural network Tres-net to obtain the crack segmentation skeleton feature map, which realizes the pixel-level localization of cracks and provides interference-free crack information for subsequent stages.

[0090] In this embodiment, the semantic segmentation neural network Tres-net borrows the encoder-decoder structure from the U-net network. Specifically, as shown... Figure 2 As shown, the semantic segmentation neural network Tres-net in this embodiment includes an encoder and a decoder;

[0091] The encoder includes structured residual blocks (Red-X) and MVB modules; such as Figure 4 As shown in (a), the structured residual block Red-X is used to linearly combine the three input channels, which utilizes the high computational performance of dense matrices while ensuring the sparsity of the network structure; the MVB module is used to encode global information and autonomously learn the key areas of the crack.

[0092] The encoder performs downsampling feature extraction during encoding. For example... Figure 2 As shown, the final result of downsampling performed on the encoder is fused using an attention-drilled pyramid module for multi-scale feature fusion. After each step of downsampling by the encoder is passed sequentially through a channel attention module and a spatial attention ASEP module, the result is then concatenated with the result of each step of upsampling by the decoder using a gated attention unit. Specifically, the principle of the gated attention unit is as follows: Figure 5 As shown, it contains semantic information of the context to determine the importance of each pixel through gating attention units, locate the region to which attention is applied, and suppress feature responses of irrelevant regions.

[0093] Finally, the final result of feature fusion, achieved by downsampling by the encoder and upsampling by the decoder, is used to generate a slit segmentation skeleton feature map through two convolution operations.

[0094] Specifically, the training method for the semantic segmentation neural network Tres-net is as follows:

[0095] Real-world crack images were collected using drones, cameras, etc., and crack locations were marked using traditional crack annotation methods to obtain training labels for the semantic segmentation neural network Tres-net.

[0096] The original image was processed using the Clahe image enhancement method to enhance local details of the cracks and suppress noise introduced by the enhancement, thereby improving the before-and-after contrast. Figure 7 As shown;

[0097] The crack and label images are standardized to form a crack segmentation dataset;

[0098] The semantic segmentation neural network Tres-net was trained using a crack segmentation dataset, with the Crack_BD_Loss loss function used during training.

[0099] Crack_BD_Loss = 0.2 × L Dice (x n ,y n )+0.8×L BCE (x n ,y n )

[0100]

[0101] Where: Crack_BD_Loss is the Crack_BD_Loss loss function; x n Indicates the predicted pixel output; y n Indicates with x n Corresponding label; L Dice It is the Dice coefficient loss function, L BCE It is a binary cross-entropy loss function, and the Crack_BD_Loss loss function can be obtained by using L... BCE Considering pixel-level classification accuracy, and simultaneously through L Dice Considering the similarity of the segmentation results, a more global pixel distribution can be derived. In this embodiment, L Dice Loss and L BCE The loss is a linear combination of the two.

[0102] When the loss function converges, the training of the semantic segmentation neural network Tres-Net is complete, and the optimal weight coefficients are obtained. These optimal weight coefficients are then used as the weight coefficients for Tres-Net. After using the optimal weight coefficients and undergoing testing, Tres-Net can achieve pixel-level crack localization and output a crack segmentation skeleton feature map. The generated crack segmentation skeleton feature map, free from environmental interference, reflects the location and features of the crack in the form of a binary image. It can directly distinguish the geometric features of the crack from irrelevant backgrounds, making it easier for the neural network to learn how to classify cracks and provide interference-free crack information for subsequent stages. The segmentation results of Tres-Net are shown below. Figure 8 As shown, (a) is the original image to be segmented, (b) is the label image (i.e., the correct answer that the network needs to learn), and (c) is the segmentation result, which generates a crack segmentation skeleton feature map with location information and crack features.

[0103] Step 3: Classification Stage

[0104] The crack segmentation skeleton feature map is input into the crack image classification neural network Crack-Class to obtain crack category information for crack diagnosis.

[0105] In this embodiment, the training method for the crack image classification neural network Crack-Class is as follows:

[0106] We manually selected strip cracks, block cracks, and mesh cracks from the crack segmentation skeleton feature map to obtain a crack classification dataset, and used cross-validation to ensure the accuracy of crack classification.

[0107] Image enhancement techniques are used to expand the crack classification dataset, such as... Figure 9 As shown; image enhancement methods include adding flip, salt and pepper noise, Gaussian noise, and blurring;

[0108] The Crack-Class neural network for crack image classification was trained using the expanded crack classification dataset. The loss function used during training was the multivariate cross-entropy loss function.

[0109]

[0110] Among them, I CE This represents the multivariate cross-entropy loss function; N is the number of categories; y i It is the one-hot encoding of the real label, representing the real situation of the i-th category, p i It is the probability that the model predicts for the i-th category;

[0111] When the loss function converges, the training of the crack image classification neural network Crack-Class is complete, and the optimal weight coefficients are obtained. These optimal weight coefficients are then used as the weight coefficients for the Crack-Class neural network. After testing with the optimal weight coefficients, the test results for the Crack-Class neural network are as follows. Figure 10 As shown in (b), the automatic classification of the crack segmentation skeleton feature map can be realized, and then mapped to the original bridge crack, providing information for bridge crack diagnosis.

[0112] The embodiments described above are merely preferred embodiments for fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for intelligent detection of bridge concrete cracks throughout the entire process based on deep neural networks, characterized in that: Includes the following steps: Step 1: Filtering Stage The image to be detected is input into the image filtering neural network Crack-Filter to remove false crack images from the image to be detected and obtain the real crack image; Step Two: Positioning Phase The real crack image is input into the semantic segmentation neural network Tres-net to obtain the crack segmentation skeleton feature map, which realizes the pixel-level localization of cracks and provides interference-free crack information for subsequent stages. Step 3: Classification Stage The crack segmentation skeleton feature map is input into the crack image classification neural network Crack-Class to obtain crack category information for crack diagnosis. In step one, the image filtering neural network Crack-Filter includes: 1×1 convolutional layer: used to perform dimensionality upscaling on the input image; Bottleneck block: Used to extract crack features; SE attention mechanism: Added to the bottleneck block to focus on more important crack feature information; The SE attention mechanism includes: 1×1 convolutional layer: used to calculate local feature statistics without changing the spatial structure of the image; The Squeeze part is used to generate a feature representation for each channel in the spatial dimension by global average pooling. The excitation section includes two 1×1 convolutional layers. The first 1×1 convolutional layer is followed by a Swish-B activation function, and the second 1×1 convolutional layer is followed by a Sigmoid function. These two 1×1 convolutional layers are used to map the dimensionality-reduced representation back to the original number of channels. The Swish-B activation function is expressed as follows: in: It is a learnable and trainable parameter; The Sigmoid function is used to generate a weight vector, which is then multiplied by the input to enhance the effective feature information in the channel. The semantic segmentation neural network Tres-net includes an encoder and a decoder; The encoder includes a structured residual block Red-X and an MVB module; the structured residual block Red-X is used to linearly combine the three input channels, which utilizes the high computational performance of dense matrices while ensuring the sparsity of the network structure; the MVB module is used to encode global information and autonomously learn the key regions of the crack. The encoder performs downsampling feature extraction during encoding; the final result of the encoder's downsampling is fused using an attention-drilled pyramid module for multi-scale feature fusion; after each step of the encoder's downsampling result is passed sequentially through a channel attention module and a spatial attention module, it is fused with the result of each step of the decoder's upsampling result through a gated attention unit to determine the importance of each pixel, locate the region to be paid attention to, and suppress the feature response of irrelevant regions; the final result of the feature fusion of the encoder's downsampling and the decoder's upsampling is used to generate a crack segmentation skeleton feature map through two convolution operations.

2. The intelligent detection method for bridge concrete cracks based on deep neural networks according to claim 1, characterized in that: The training method for the image filtering neural network CrackFilter is as follows: A crack presence dataset, including both real and fake crack images, was constructed to train the image filtering neural network Crack-Filter. The binary cross-entropy loss function was used during training. in, Represents the binary cross-entropy loss function; Indicates the predicted pixel output. Indicates and Corresponding tags; Indicates the total number of samples; The training of the image filtering neural network Crack-Filter is completed when the loss function converges, and the optimal weight coefficients obtained from the training are used as the weight coefficients of the image filtering neural network Crack-Filter.

3. The intelligent detection method for bridge concrete cracks based on deep neural networks according to claim 1, characterized in that: The training method for the semantic segmentation neural network Tres-net is as follows: Real-world crack images were acquired and crack locations were marked to obtain training labels for the semantic segmentation neural network Tres-net; The original image was processed using the Clahe image enhancement method to enhance local details of the cracks, suppress noise introduced by image enhancement, and improve the before-and-after contrast. The crack and label images are standardized to form a crack segmentation dataset; The semantic segmentation neural network Tres-net was trained using a crack segmentation dataset, with the Crack_BD_Loss loss function used during training. in: The loss function is Crack_BD_Loss; Indicates the predicted pixel output; Indicates and The corresponding label; It is the Dice coefficient loss function; It is a binary cross-entropy loss function; When the loss function converges, the training of the semantic segmentation neural network Tres-net is completed, the optimal weight coefficients are obtained, and the optimal weight coefficients are used as the weight coefficients of the semantic segmentation neural network Tres-net.

4. The intelligent detection method for bridge concrete cracks based on deep neural networks according to claim 1, characterized in that: The training method for the crack image classification neural network Crack-Class is as follows: Strip cracks, block cracks, and mesh cracks were selected from the crack segmentation skeleton feature map to obtain a crack classification dataset, and cross-validation was used to ensure the accuracy of crack classification. Image enhancement techniques were used to expand the crack classification dataset; The Crack-Class neural network for crack image classification was trained using the expanded crack classification dataset. The loss function used during training was the multivariate cross-entropy loss function. in, Represents the multivariate cross-entropy loss function; It is the number of categories; It is the one-hot encoding of the real label, representing the first... The actual situation of each category The model predicts the first The probability of each category; When the loss function converges, the training of the crack image classification neural network Crack-Class is completed, the optimal weight coefficients are obtained, and the optimal weight coefficients are used as the weight coefficients of the crack image classification neural network Crack-Class.

Images