Building crack detection method, system and device based on deep learning

Abstract

The building crack detection method, system and equipment based on deep learning belong to the technical field of image processing and target detection. In order to solve the problem that the existing crack detection method has low precision and slow detection speed, the building crack image is collected first, then the light crack detection network model is used for building crack detection on the collected building crack image, the input layer is first subjected to a 3*3 convolution block, then a maximum layer is connected to the input image for down-sampling, and a feature image after down-sampling is obtained; then, for the feature image after down-sampling, three groups of ShuffleNetV2 networks are used to extract network features, then the feature aggregation network is used for feature aggregation, finally, three parallel prediction branches of the prediction network are used for common processing, and cracks of different sizes are detected.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing and target detection, and relates to a building crack detection method, system and equipment. BACKGROUND

[0002] With the rapid development of the construction industry, the volume of houses and bridge buildings increases every year, and the maintenance work of the buildings is also increased. Cracks, as one of the most common hazards in buildings, are an important indicator for judging the safety of building structures. Cracks can affect the performance of building structures, damage the integrity of building structures, and reduce the service life of buildings, which is a major safety hazard. If not maintained in time, it may even cause serious safety accidents.

[0003] Traditional image processing algorithms process and analyze collected images in order to achieve the purpose of identifying building cracks. However, the clarity of the image, the influence of the shadow, the image distortion and other factors will affect the detection result. In addition, the background of the image is complex and there is a lot of noise, so the traditional algorithm cannot effectively and quickly identify the crack target in the image.

[0004] Although the deep learning algorithm greatly improves the detection accuracy and speed of cracks, the calculation amount required by the deep learning algorithm is very large and is not easy to be mounted on some mobile devices. Therefore, the existing technology still needs to be improved and promoted. SUMMARY

[0005] The application is to solve the problems of low precision and slow detection speed of the existing crack detection method, and the problem that the crack detection using the deep learning algorithm is mostly not suitable for mobile devices due to the large amount of calculation.

[0006] A building crack detection method based on deep learning, first collects building crack images, and then uses a lightweight crack detection network model to detect building cracks in the collected building crack images.

[0007] The lightweight crack detection network model comprises a feature extraction network, a feature aggregation network and a prediction network.

[0008] The feature extraction network processing process is as follows:

[0009] The input layer first passes through a 3x3 convolution block, and then connects to a maximum value layer to down-sample the input image to obtain a down-sampled feature image. Then, three groups of improved ShuffleNetV2 networks are used to extract network features from the down-sampled feature image.

[0010] Each improved ShuffleNetV2 network includes one basic unit SN_2 and multiple basic units SN_1;

[0011] SN_2: the input is respectively connected to two branches, the first branch sequentially passes through a 3*3 depth separable convolution and a 1*1 group convolution; the second branch sequentially passes through a 1*1 group convolution, a 3*3 depth separable convolution and a 1*1 group convolution, and then the two branches are concatenated and then channel integration is performed;

[0012] SN_1: the input is channel split and becomes two branches after the Channel Split operation, the first branch is connected to a 3*3 maximum pooling operation; the second branch is sequentially connected to three convolutions with a step of 1, which are a 1*1 group convolution, a 3*3 depth separable convolution and a 1*1 group convolution; then the two branches are concatenated and then channel integration is performed;

[0013] The three improved ShuffleNetV2 networks for feature extraction are respectively referred to as the second layer to the fourth layer, and the process of feature extraction by the second layer to the fourth layer is as follows:

[0014] The down-sampled feature image is input into the second layer, the second layer is first connected to SN_2 and then connected to a1 SN_1, and the output is recorded as the output of the second layer; the features output by the second layer are transmitted to the third layer, and the output is transmitted to the first branch of the feature aggregation network, and the first branch of the feature aggregation network is first processed by the improved Coordinate Attention model;

[0015] The third layer is first connected to SN_2 and then connected to a2 SN_1 to obtain the output of the third layer; the output of the third layer is connected to the fourth layer, and the output is transmitted to the second branch of the feature aggregation network, and the second branch of the feature aggregation network is first processed by the improved Coordinate Attention model;

[0016] The fourth layer is first connected to SN_2 and then connected to a3 SN_1, and the output is then connected to SN_1 and then connected to SN_1, and the output of the fourth layer is connected to the SPPCSPC module of the third branch of the feature aggregation network and then processed by the improved Coordinate Attention model;

[0017] The processing process of the feature aggregation network is as follows:

[0018] The improved Coordinate Attention model of the first to third branches of the feature aggregation network is respectively connected with a convolution layer; wherein, the input features of the improved Coordinate Attention model are respectively subjected to maximum pooling operation in the h direction and the w direction, then the pooled features are subjected to Concat splicing, after splicing, 1x1 2d convolution operation is performed on the row, then normalization and non-linear processing are performed, the processed features are then divided into the h direction and the w direction and subjected to 2d convolution operation respectively, then sigmoid processing is performed in the respective directions to obtain the weight coefficients of the coordinates, the coordinate weight coefficients are multiplied with the original input features to obtain the weighted features;

[0019] The output of the convolution layer corresponding to the improved Coordinate Attention model in the third branch of the feature aggregation network is input into the first upsampling layer, after upsampling, Cat processing is performed with the output of the convolution layer corresponding to the improved Coordinate Attention model in the second branch of the feature aggregation network, and then the output is input into the first ELAN-H layer;

[0020] The output of the first ELAN-H layer is input into the second upsampling layer after being processed by the improved Coordinate Attention model and the convolution layer, after upsampling, Cat processing is performed with the output of the convolution layer corresponding to the improved Coordinate Attention model in the first branch of the feature aggregation network, and then the output is input into the second ELAN-H layer, the output of the second ELAN-H layer is input into the first MP-2 layer for processing and also input into the first prediction branch of the prediction network for processing;

[0021] The output of the first MP-2 layer is subjected to Cat processing with the input of the first ELAN-H layer, and then the output is input into the third ELAN-H layer, the output of the third ELAN-H layer is input into the second MP-2 layer for processing and also input into the second prediction branch of the prediction network for processing;

[0022] The output of the second MP-2 layer processing is subjected to Cat processing with the output of the SPPCSPC module of the third branch, and then the output is input into the fourth ELAN-H layer, the output of the fourth ELAN-H layer is input into the third prediction branch of the prediction network for processing;

[0023] The processing process of the prediction network is as follows:

[0024] The prediction network includes three parallel prediction branches, each branch including a REP module and a CBM layer; the REP module in the detection stage includes a 1x1 convolution layer and a BN layer; the CBM module includes a convolution layer, a BN layer and a sigmoid layer;

[0025] The prediction network includes three parallel prediction branches for jointly processing and detecting cracks of different sizes.

[0026] Preferably, the second layer to the fourth layer of the feature extraction network correspond to improved ShuffleNetV2 networks respectively including 3 SN_1, 7 SN_1 and 3 SN_1, that is, α1, α2 and α3 are 3, 7 and 3 respectively.

[0027] Further, the ELAN-H layer processing process is as follows:

[0028] First, channel splitting is performed to divide the input into two branches, the short branch is connected to a 1×1 CBS module with a step size of 1, and the long branch is first connected to a 1×1 CBS module with a step size of 1, and then connected to four 3×3 CBS modules with a step size of 1, and then the outputs of the six CBS modules of the short branch and the long branch are subjected to Cat processing, and then a 1×1 CBS module is processed.

[0029] Further, the MP-2 layer processing process is as follows:

[0030] The input is split into two branches by channel splitting, one branch is first subjected to Maxpool operation and then connected to a 1×1 CBS module with a step size of 1, and the other branch is connected to a 1×1 CBS module with a step size of 1 and a 3×3 CBS module with a step size of 1 in turn, and then the outputs of the two branches are subjected to Cat processing.

[0031] Further, the REP module in the detection stage is different in structure from the REP module in the training stage, the REP module in the detection stage is determined based on the REP module in the training stage, the REP module in the training stage is composed of three branches, the first branch is first connected to a 1×1 convolution layer with a step size of 1, and then connected to a BN layer; the second branch is first connected to a 1×1 convolution layer, and then connected to a BN layer; the third branch is only connected to a BN layer, and then the results of the three branches are added; after training, the 1×1 convolution layer and the BN layer of the first branch are retained as the REP module in the test stage.

[0032] A building crack detection system based on deep learning, comprising an image acquisition unit and a crack detection unit;

[0033] The image acquisition unit is used to acquire building crack images.

[0034] The crack detection unit inputs the acquired building crack images into a lightweight crack detection network model to detect building cracks in the building crack images.

[0035] The lightweight crack detection network model comprises a feature extraction network, a feature aggregation network and a prediction network.

[0036] The feature extraction network processing procedure is as follows:

[0037] The input layer is first subjected to a 3*3 convolution block, and then subjected to a maximum value layer to down-sample the input image to obtain a down-sampled feature image; then, three groups of improved ShuffleNetV2 networks are used to extract network features from the down-sampled feature image;

[0038] Each group of improved ShuffleNetV2 networks comprises a basic unit SN_2 and a plurality of basic units SN_1.

[0039] SN_2: the input is connected to two branches, the first branch sequentially passes through a 3*3 depth separable convolution and a 1*1 group convolution; the second branch sequentially passes through a 1*1 group convolution, a 3*3 depth separable convolution and a 1*1 group convolution, and then the two branches are concatenated and then channel integration is performed;

[0040] SN_1: the input is subjected to channel splitting and becomes two branches after the Channel Split operation, the first branch is connected to a 3*3 maximum pooling operation; the second branch sequentially passes through three convolutions with a step of 1, which are a 1*1 group convolution, a 3*3 depth separable convolution and a 1*1 group convolution; then, the two branches are concatenated and then channel integration is performed;

[0041] The three groups of improved ShuffleNetV2 networks are denoted as the second layer to the fourth layer, and the feature extraction process of the second layer to the fourth layer is as follows:

[0042] The down-sampled feature image is input into the second layer, the second layer is first connected to SN_2 and then connected to a1 SN_1, and the output is denoted as the output of the second layer; the features output by the second layer are transmitted to the third layer, and the output is transmitted to the first branch of the feature aggregation network, and the first branch of the feature aggregation network is first processed by the improved Coordinate Attention model;

[0043] The third layer is first connected to SN_2 and then connected to a2 SN_1 to obtain the output of the third layer; the output of the third layer is connected to the fourth layer, and the output is transmitted to the second branch of the feature aggregation network, and the second branch of the feature aggregation network is first processed by the improved Coordinate Attention model;

[0044] The fourth layer first accesses SN_2, then accesses three SN_1s, and then accesses SN_1, and then accesses SN_1, and then the output of the fourth layer is input into the SPPCSPC module of the third branch of the feature aggregation network, and then is sent to the improved Coordinate Attention model for processing;

[0045] The processing process of the feature aggregation network is as follows:

[0046] The first to third branches of the feature aggregation network are respectively connected with a convolution layer; wherein the input features of the improved Coordinate Attention model are respectively subjected to maximum pooling operation in the h direction and the w direction, then the pooled features are subjected to Concat splicing, then 1×1 2d convolution operation is performed on the spliced features in the row direction, then normalization and nonlinear processing are performed, then the processed features are divided into the h direction and the w direction and are respectively subjected to 2d convolution operation, then sigmoid processing is performed in the respective directions to obtain the coordinate weight coefficients, and the coordinate weight coefficients are multiplied with the original input features to obtain the weighted features;

[0047] The output of the convolution layer corresponding to the improved Coordinate Attention model in the third branch of the feature aggregation network is input into the first upsampling layer, and after upsampling, the output is subjected to Cat processing with the output of the convolution layer corresponding to the improved Coordinate Attention model in the second branch of the feature aggregation network, and then is input into the first ELAN-H layer;

[0048] The output of the first ELAN-H layer is input into the second upsampling layer after being processed by the improved Coordinate Attention model and the convolution layer, and after upsampling, the output is subjected to Cat processing with the output of the convolution layer corresponding to the improved Coordinate Attention model in the first branch of the feature aggregation network, and then is input into the second ELAN-H layer, and the output of the second ELAN-H layer is input into the first MP-2 layer for processing and is also input into the first prediction branch of the prediction network for processing;

[0049] The output of the first MP-2 layer is subjected to Cat processing with the input of the first ELAN-H layer, and then is input into the third ELAN-H layer, and the output of the third ELAN-H layer is input into the second MP-2 layer for processing and is also input into the second prediction branch of the prediction network for processing;

[0050] The output of the second MP-2 layer processing is subjected to Cat processing with the output of the SPPCSPC module of the third branch, and then is input into the fourth ELAN-H layer, and the output of the fourth ELAN-H layer is input into the third prediction branch of the prediction network for processing;

[0051] The processing procedure of the prediction network is as follows:

[0052] The prediction network comprises three parallel prediction branches, each branch comprising a REP module and a CBM layer; the REP module in the detection stage comprises a 1*1 convolution layer and a BN layer; the CBM module comprises a convolution layer, a BN layer and a sigmoid layer;

[0053] The prediction network comprises three parallel prediction branches for joint processing, and detects cracks of different sizes.

[0054] Further, the processing procedure of the ELAN-H layer is as follows:

[0055] First, channel splitting is performed, and the input is divided into two branches; the short branch is connected to a 1*1 CBS module with a step size of 1, and the long branch is first connected to a 1*1 CBS module with a step size of 1, and then connected to four 3*3 CBS modules with a step size of 1 in sequence, and then the outputs of the six CBS modules of the short branch and the long branch are subjected to Cat processing, and then subjected to a 1*1 CBS module processing.

[0056] Further, the processing procedure of the MP-2 layer is as follows:

[0057] The input is subjected to channel splitting and divided into two branches; one branch is first subjected to Maxpool operation, and then connected to a 1*1 CBS module with a step size of 1, and the other branch is connected to a 1*1 CBS module with a step size of 1 and a 3*3 CBS module with a step size of 1 in sequence, and then the outputs of the two branches are subjected to Cat processing.

[0058] A computer storage medium, the storage medium stores at least one instruction, the at least one instruction is loaded and run by the processor, and the at least one instruction is loaded and run by the processor.

[0059] A building crack detection device based on deep learning, the device comprises a processor and a memory, the memory stores at least one instruction, the at least one instruction is loaded and run by the processor, and the at least one instruction is loaded and run by the processor.

[0060] The beneficial effects of the present application are:

[0061] 1. The improved ShuffleNetV2 feature extraction network is used to replace the feature extraction network, which reduces the parameters of the network model, reduces the calculation amount of the model, and reduces the complexity of the model, not only improves the detection speed, but also makes the whole model better portable to mobile devices.

[0062] 2. In the down-sampling stage, the double operation of pooling and convolution is used to realize down-sampling, which reduces the loss of features. The improved Coordinate Attention mechanism is also added, which can not only capture cross-channel information, but also obtain direction-aware and position-sensitive information, which makes the model pay more attention to some useful feature information, so that the detection result of the model is more accurate and effective. The overall network structure of the present application can ensure the detection speed and has high crack detection precision. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 It is a building crack detection method flow chart based on deep learning;

[0064] Figure 2 It is a building crack detection method overall network structure diagram based on deep learning;

[0065] Figure 3 It is an improved ShuffleNetV2 network structure diagram when the step is 1;

[0066] Figure 4 It is an improved ShuffleNetV2 network structure diagram when the step is 2;

[0067] Figure 5 It is a network structure diagram of the improved Coordinate Attention model;

[0068] Figure 6 It is a network structure diagram of the ELAN-H module;

[0069] Figure 7 It is a network structure diagram of the MP-2 module;

[0070] Figure 8 It is a network structure diagram of the REP module;

[0071] Figure 9 It is a comparison diagram of the original picture of the crack and the detected picture. DETAILED DESCRIPTION

[0072] The application is a building crack detection method based on a lightweight feature extraction network and an attention model. The application uses a deep learning method to detect cracks in different environments and different light conditions of buildings. In order to reduce the calculation parameters of deep learning and improve the calculation speed of the equipment, a lightweight feature extraction network and an attention model building crack detection method is developed for some surfaces of the building that are difficult to detect and difficult to reach, so as to solve the problem of detecting cracks on the surface of the building. First, the surface crack images of the building are collected, the collected images are subjected to image enhancement operation to obtain more images, all data are randomly divided into a training set, a validation set and a test set, the training set and the test set images are classified and labeled, then the prepared data set is put into the lightweight network model constructed by the application for training, multiple training is performed, the accuracy and model weight parameters of each training are recorded, when the training is completed, the model with the highest accuracy in the training result is selected as the optimal model, and the weight parameter file of the network model is saved. The test set is used for testing, the correctness and the missed detection rate of the crack detection result are observed, the detection result is recorded, and the operation time of each frame of image of the optimal model is recorded.

[0073] The application will be further described in combination with the specific embodiments.

[0074] Specific embodiment one, in combination Figure 1 and Figure 2 The present embodiment is described,

[0075] The present embodiment is a building crack detection method based on deep learning, comprising the following steps:

[0076] Step one: use a drone to take pictures, collect building crack images, and perform image enhancement on the collected images to expand the number of images, divide the expanded images into a training set, a validation set and a test set, and label the images of the training set and the validation set to construct the required data set for the experiment.

[0077] Data enhancement processing is performed on the data set: the processing modes of flipping, rotating, adjusting brightness, adjusting contrast, adding Gaussian noise and salt and pepper noise are used for enhancement. Flipping refers to the operation of reversing up, down, left and right with the image midpoint as the center of symmetry, rotating refers to rotating the image around the center by a certain angle, scaling refers to enlarging and reducing the image, adjusting brightness refers to enlarging or reducing the brightness in the image according to a certain value, adjusting contrast refers to stretching or compressing the brightness value range in the image to the specified brightness display range of the display system, and adding Gaussian noise and salt and pepper noise refers to adding Gaussian noise and salt and pepper noise to the image through image processing operation.

[0078] All images in the training set and the validation set are labeled, and the labeling is performed using labelme. The crack is framed with a continuous small frame to better extract the feature information. The json file generated by labeling is converted into a txt format file for network training.

[0079] Step two: build a lightweight crack detection network model

[0080] As shown in Figure 2 , the lightweight crack detection network model includes a feature extraction network, a feature aggregation network, and a prediction network; wherein,

[0081] The feature extraction network first uses a 3x3 convolution block with a step of 2, and then inputs a 3x3 max pooling operation with a step of 2 to downsample the input image to obtain a downsampled feature image. Then, for the downsampled feature image, three groups of improved ShuffleNetV2 network are used to extract network features, and the three groups of improved ShuffleNetV2 network are respectively referred to as the second layer to the fourth layer.

[0082] The improved ShuffleNetV2 network includes two basic units SN_1 and SN_2; wherein,

[0083] SN_1: as shown in Figure 3 , first split the input channels, and after the Channel Split operation, it becomes two branches. In order to better extract feature information, the left branch is connected to a 3x3 max pooling operation with a step of 1. The right branch is sequentially connected to three convolutions with a step of 1, which are 1x1 group convolution, 3x3 depth separable convolution, and 1x1 group convolution. Then, the two branches are concatenated and then integrated.

[0084] SN_2: as shown in Figure 4 , the input is still connected to the left and right two branches, and the depth separable convolution of the left and right two branches has a step of 2. The left branch sequentially passes through 3x3 depth separable convolution and 1x1 group convolution. The right branch sequentially passes through 1x1 group convolution, 3x3 depth separable convolution, and 1x1 group convolution. Then, the two branches are concatenated and then integrated.

[0085] It is found through research that the edges of the cracks and the cracks themselves have obvious differences compared with the surrounding buildings. A 3x3 max pooling operation with a step of 1 is added to the left branch of the ShuffleNetV2 network basic unit SN_1, which will help to extract the main features in the image. After adding the max pooling operation, the network can better capture the edge information and texture information of the cracks to achieve more accurate crack recognition.

[0086] The 1x1 group convolution can be regarded as a convolution operation of each channel with a 1x1 convolution kernel, which can be regarded as a linear combination kernel feature transformation in each channel. This greatly reduces the complexity of the model, thereby reducing the number of parameters and the complexity of calculation.

[0087] After changing the left branch of SN_1 to a 3x3 max pooling operation with a step of 1 and changing all 1x1 ordinary convolutions in the ShuffleNetV2 network to 1x1 group convolutions, the overall calculation parameter amount is reduced while the crack recognition accuracy is improved.

[0088] The improved ShuffleNetV2 network is used to extract features:

[0089] The down-sampled feature image is input into the second layer SN_2, and then input into SN_1. The output is then input into SN_1, and the output is input into SN_1 to obtain the output of the second layer. The features of the second layer output are transmitted to the third layer, and the output is transmitted to the first branch of the feature aggregation network. The first branch of the feature aggregation network first uses the improved CoordinateAttention model for processing. For this layer, we can obtain a larger receptive field, so we can obtain a wider image area to better understand the background information around the target object.

[0090] The third layer first inputs SN_2, then inputs SN_1, and the output is input into SN_1. This SN_1 operation is repeated 7 times to obtain the output of the third layer. The output of the third layer is input into the fourth layer, and the output is transmitted to the second branch of the feature aggregation network. The second branch of the feature aggregation network first uses the improved CoordinateAttention model for processing. For this layer, we repeat the use of SN_1 seven times, and we can obtain more crack feature information, which greatly improves the detection performance.

[0091] The fourth layer first accesses SN_2, then accesses SN_1, and then outputs and accesses SN_1, and then outputs and accesses SN_1. The output result of the fourth layer accesses the SPPCSPC module of the third branch of the feature aggregation network, and is then sent to the improved CoordinateAttention model for processing. For this layer, we continue to perform feature extraction operations on the basis of the above two layers, which enables us to extract feature information of some fine cracks.

[0092] Such a network structure improves the multi-scale detection capability of our model, so that our model can better capture and locate cracks of different sizes in the image structure, thereby improving the detection performance of the target.

[0093] The processing process of the feature aggregation network is as follows:

[0094] The first to third branches of the feature aggregation network are respectively connected with a convolution layer; wherein the input features of the improved Coordinate Attention model are respectively subjected to maximum pooling operation in the h direction and the w direction, then the pooled features are subjected to Concat splicing, then the spliced features are subjected to 1x1 2d convolution operation in the row direction, then normalization and nonlinear processing are performed, then the processed features are divided into the h direction and the w direction and are respectively subjected to 2d convolution operation, then sigmoid is performed in the respective directions to obtain the weight coefficients of the coordinates, and then the weight coefficients of the coordinates are multiplied with the original input features, so as to obtain the weighted features, as shown in Figure 5

[0095] The Coordinate Attention model introduces position encoding and coordinate information, which enables our model to better understand the relationship between space and position, which makes us less likely to miss detection when processing some fine cracks, because it can finely allocate attention according to the position.

[0096] ​On the basis of the original Coordinate Attention model, we change the average value pooling operation in the h direction and w direction to maximum value pooling operation to help the model better understand the position information of the crack features in the graph. The maximum value pooling operation in the coordinate dimension is to interact the coordinate information with the channel dimension of the feature map to strengthen the model's perception of the position. The maximum value pooling operation will emphasize the maximum value in each coordinate dimension, which will emphasize some positions in the graph. We connect the maximum value pooling operation in the left branch of SN_1 and the Coordinate Attention model, which will emphasize some of the same positions and features. They will enhance each other, so that these features will occupy a larger weight in the Re-weight operation of the Coordinate Attention model. This will make the entire network better understand these important features and greatly improve the performance of the network, enabling the network to quickly and accurately identify cracks. At the same time, this will also improve the robustness of the entire network, which is helpful for processing crack images in different environments and to a certain extent, against noise in the input data.

[0097] The output of the improved Coordinate Attention model in the third branch of the feature aggregation network is sent to the first upsampling layer. After upsampling, it is Cat processed with the output of the improved Coordinate Attention model corresponding to the convolutional layer in the second branch of the feature aggregation network, and then sent to the first ELAN-H layer.

[0098] The output of the first ELAN-H layer is processed by the improved Coordinate Attention model and the convolutional layer, and then sent to the second upsampling layer. After upsampling, it is Cat processed with the output of the improved Coordinate Attention model corresponding to the convolutional layer in the first branch of the feature aggregation network, and then sent to the second ELAN-H layer. The output of the second ELAN-H layer is sent to the first MP-2 layer for processing, and also sent to the first prediction branch of the prediction network for processing.

[0099] The output of the first MP-2 layer is Cat processed with the input of the first ELAN-H layer, and then sent to the third ELAN-H layer. The output of the third ELAN-H layer is sent to the second MP-2 layer for processing, and also sent to the second prediction branch of the prediction network for processing.

[0100] The output of the second MP-2 layer is Cat processed with the output of the SPPCSPC module in the third branch, and then sent to the fourth ELAN-H layer. The output of the fourth ELAN-H layer is sent to the third prediction branch of the prediction network for processing.

[0101] The ELAN-H layer is as followsFigure 6 As shown in FIG. 2, first, channel splitting is performed to divide the input into two branches, the short branch is connected to a 1x1 CBS module with a step size of 1, and the long branch is first connected to a 1x1 CBS module with a step size of 1, and then sequentially connected to four 3x3 CBS modules with a step size of 1, and then the outputs of the six CBS modules in the short branch and the long branch are subjected to Cat processing, and then a 1x1 CBS module is used for processing to obtain more feature information.

[0102] As shown in FIG. 2, first, channel splitting is performed to divide the input into two branches, the short branch is connected to a 1x1 CBS module with a step size of 1, and the long branch is first connected to a 1x1 CBS module with a step size of 1, and then sequentially connected to four 3x3 CBS modules with a step size of 1, and then the outputs of the six CBS modules in the short branch and the long branch are subjected to Cat processing, and then a 1x1 CBS module is used for processing to obtain more feature information. Figure 7 As shown in FIG. 2, first, channel splitting is performed to divide the input into two branches, the short branch is connected to a 1x1 CBS module with a step size of 1, and the long branch is first connected to a 1x1 CBS module with a step size of 1, and then sequentially connected to four 3x3 CBS modules with a step size of 1, and then the outputs of the six CBS modules in the short branch and the long branch are subjected to Cat processing, and then a 1x1 CBS module is used for processing to obtain more feature information.

[0103] The processing process of the prediction network is as follows:

[0104] The prediction network includes three parallel prediction branches, and each branch includes a REP module and a CBM layer.

[0105] As shown in FIG. 2, first, channel splitting is performed to divide the input into two branches, the short branch is connected to a 1x1 CBS module with a step size of 1, and the long branch is first connected to a 1x1 CBS module with a step size of 1, and then sequentially connected to four 3x3 CBS modules with a step size of 1, and then the outputs of the six CBS modules in the short branch and the long branch are subjected to Cat processing, and then a 1x1 CBS module is used for processing to obtain more feature information. Figure 8

[0106] When in the training stage, the REP module is composed of three branches, the first branch is first connected to a 1x1 convolution with a step size of 1, and then connected to a BN layer, and the function of this branch is to extract features; the second branch is first connected to a 1x1 convolution with a step size of 1, and then connected to a BN layer; the third branch is only connected to a BN layer, and then the results of the three branches are added together,

[0107] When in the detection stage, the REP module is composed of a 1x1 convolution with a step size of 1 and a BN layer.

[0108] As shown in FIG. 2, first, channel splitting is performed to divide the input into two branches, the short branch is connected to a 1x1 CBS module with a step size of 1, and the long branch is first connected to a 1x1 CBS module with a step size of 1, and then sequentially connected to four 3x3 CBS modules with a step size of 1, and then the outputs of the six CBS modules in the short branch and the long branch are subjected to Cat processing, and then a 1x1 CBS module is used for processing to obtain more feature information.

[0109] ​​The three prediction branches have different receptive field sizes due to the convolution feature maps corresponding to different depths, when a large crack appears on the picture, if the receptive field is small, only part of the information of the crack is included, at this time, we are difficult to identify the crack; similarly, when the receptive field is large and the crack is small, the large receptive field often ignores these small cracks. Therefore, three prediction branches are set, which correspond to large, medium and small receptive fields respectively, and they work together to enable the network model to detect cracks of different sizes without changing the resolution of the input image. Among them, the prediction branch corresponding to the shallower convolution feature has a wider receptive field, which can better capture the global features of large size targets, while the prediction branch corresponding to the middle layer of convolution feature has moderate resolution and receptive field, which can balance fine feature extraction and wide coverage features, and the prediction branch corresponding to the deeper convolution feature has a relatively small receptive field but a higher resolution, which enables more accurate detection of small targets.

[0110] The lightweight crack detection network model of the application not only has the characteristics of lightweight, but also has very high recognition accuracy. In the down-sampling stage, the double operation of pooling and convolution is used to realize down-sampling, which reduces the loss of features. At the same time, the improved Coordinate Attention mechanism is added before the convolution layer of the feature aggregation network. The improved Coordinate Attention embeds position information into channel attention, so that the mobile network obtains more regional information and avoids introducing large calculation amount. In order to avoid the loss of position information introduced by 2D global pooling, the channel attention is decomposed into two parallel 1D feature encodings, so as to efficiently integrate spatial coordinate information into the generated attention maps. At the same time, using max pool for pooling operation is helpful to extract important features such as edges and textures, and is not affected by other irrelevant features.

[0111] By reducing the dense connection of the backbone feature extraction network, the overall calculation amount is reduced, by increasing the network depth, more rich features can be obtained, and by adding Coordinate Attention mechanism in the feature aggregation network, the network can pay more attention to feature information. Compared with network models such as YOLO, the calculation amount of the network model of the application is obviously reduced.

[0112] Step three: use the prepared data set to train and verify the constructed lightweight crack detection deep learning network model to obtain the optimal model in the training process.

[0113] Divide the data set of step one, configure the path file, create a yaml file corresponding to your own data set, which includes: training set crack picture and label path, validation set crack picture and label path, crack class number is 1, and crack class name is crack.

[0114] Create a yaml file corresponding to the crack detection network model, name it v7_c, and record the path information of the yaml file. In train.py, set the path parameter to the path parameter of v7_c.

[0115] Before training, the corresponding hyperparameter settings need to be adjusted, including training weight file, data set yaml file, training round number, batch-size, input size, etc. The corresponding parameter settings are weight file v7.pt, data.yaml, 300, 32, 640.

[0116] After 300 rounds of training, the model corresponding to the optimal accuracy is obtained as the optimal model.

[0117] Step four: use the optimal model to detect the test set crack picture, the specific detection steps are as follows:

[0118] Put the optimal model pt file into detect.py file, and configure the remaining hyperparameters, including data set file, input size, etc. Use this model to detect the test set crack, and record the detection accuracy and detection speed.

[0119] We use the detect.py file to detect the test set picture, Figure 9 is a picture randomly selected from the test set detection results. Figure 9 The left graph in shows the original crack picture, Figure 9 The right graph in shows the crack picture detected by the algorithm of the present application. As can be seen from the graph, the original crack graph contains many shadows and noise, and Figure 9 The graph in can be seen, the algorithm of the present application can accurately and precisely frame the position of the crack. Specific implementation method two:

[0121] The embodiment is a building crack detection system based on deep learning, which is a software system corresponding to a building crack detection method based on deep learning. The system comprises an image acquisition unit and a crack detection unit.

[0122] Image acquisition unit: used for acquiring building crack images.

[0123] The crack detection unit inputs the obtained building crack image into a lightweight crack detection network model to perform building crack detection on the building crack image.

[0124] The lightweight crack detection network model comprises a feature extraction network, a feature aggregation network and a prediction network.

[0125] The feature extraction network has the following processing procedure:

[0126] The input layer first passes through a 3x3 convolution block, and then is connected to a maximum value layer to down-sample the input image to obtain a down-sampled feature image; then, three groups of improved ShuffleNetV2 networks are used to extract network features from the down-sampled feature image.

[0127] Each group of improved ShuffleNetV2 networks comprises a basic unit SN_2 and a plurality of basic units SN_1.

[0128] SN_2: the input is connected to two branches, the first branch sequentially passes through a 3x3 depth separable convolution and a 1x1 group convolution; the second branch sequentially passes through a 1x1 group convolution, a 3x3 depth separable convolution and a 1x1 group convolution, then the two branches are concatenated and then channel integration is performed.

[0129] SN_1: the input is split into two branches after the Channel Split operation, the first branch is connected to a 3x3 maximum pooling operation; the second branch is sequentially connected to three convolutions with a step of 1, which are a 1x1 group convolution, a 3x3 depth separable convolution and a 1x1 group convolution; then the two branches are concatenated and then channel integration is performed.

[0130] The three groups of improved ShuffleNetV2 networks are denoted as the second layer to the fourth layer, and the feature extraction process of the second layer to the fourth layer is as follows:

[0131] The down-sampled feature image is input into the second layer, the second layer is first connected to SN_2 and then connected to a1 SN_1, and the output is denoted as the output of the second layer; the features of the output of the second layer are transmitted to the third layer, and the output is transmitted to the first branch of the feature aggregation network, and the first branch of the feature aggregation network is first processed by the improved Coordinate Attention model.

[0132] The third layer is firstly connected to SN_2 and then connected to a2 SN_1, and the output of the third layer is obtained; the output of the third layer is connected to the fourth layer, and the output is transmitted to the second branch of the feature aggregation network, and the second branch of the feature aggregation network is firstly processed by the improved Coordinate Attention model;

[0133] The fourth layer is firstly connected to SN_2 and then connected to a3 SN_1, and the output is then connected to SN_1 and then connected to SN_1, and the output of the fourth layer is connected to the SPPCSPC module of the third branch of the feature aggregation network and then sent to the improved Coordinate Attention model for processing;

[0134] The processing process of the feature aggregation network is as follows:

[0135] The improved Coordinate Attention model of the first to third branches of the feature aggregation network is connected to a convolution layer respectively; wherein the input features of the improved Coordinate Attention model are respectively subjected to maximum pooling operation in the h direction and the w direction, then the pooled features are concatenated, and then 1×1 2d convolution operation is performed on the row, then normalization and nonlinear processing are performed, the processed features are then divided into the h direction and the w direction and subjected to 2d convolution operation respectively, then sigmoid is performed in the respective directions to obtain the weight coefficients of the coordinates, and the coordinate weight coefficients are multiplied by the original input features to obtain the weighted features;

[0136] The output of the convolution layer corresponding to the improved Coordinate Attention model in the third branch of the feature aggregation network is sent to the first upsampling layer, and after upsampling, the output of the convolution layer corresponding to the improved Coordinate Attention model in the second branch of the feature aggregation network is subjected to Cat processing, and then sent to the first ELAN-H layer;

[0137] The output of the first ELAN-H layer is processed by the improved Coordinate Attention model and the convolution layer, and then sent to the second upsampling layer, and after upsampling, the output of the convolution layer corresponding to the improved Coordinate Attention model in the first branch of the feature aggregation network is subjected to Cat processing, and then sent to the second ELAN-H layer, and the output of the second ELAN-H layer is sent to the first MP-2 layer for processing and also sent to the first prediction branch of the prediction network for processing;

[0138] The output of the first MP-2 layer is subjected to Cat processing with the input of the first ELAN-H layer, and then is sent to the third ELAN-H layer, and the output of the third ELAN-H layer is sent to the second MP-2 layer processing and also to the second prediction branch of the prediction network for processing;

[0139] The output of the second MP-2 layer processing is subjected to Cat processing with the output of the SPPCSPC module of the third branch, and then is sent to the fourth ELAN-H layer, and the output of the fourth ELAN-H layer is sent to the third prediction branch of the prediction network for processing;

[0140] The ELAN-H layer processing process is as follows:

[0141] First, the input is split into two branches, the short branch is connected to a 1x1 CBS module with a step size of 1, and the long branch is first connected to a 1x1 CBS module with a step size of 1, and then connected to four 3x3 CBS modules with a step size of 1 in sequence, and then the outputs of the six CBS modules of the short branch and the long branch are subjected to Cat processing, and then a 1x1 CBS module is processed.

[0142] The MP-2 layer processing process is as follows:

[0143] The input is split into two branches, one branch is first subjected to Maxpool operation, and then connected to a 1x1 CBS module with a step size of 1, and the other branch is connected to a 1x1 CBS module with a step size of 1 and a 3x3 CBS module with a step size of 1 in sequence, and then the outputs of the two branches are subjected to Cat processing.

[0144] The prediction network processing process is as follows:

[0145] The prediction network includes three parallel prediction branches, each branch includes a REP module and a CBM layer; the REP module in the detection stage includes a 1x1 convolution layer and a BN layer; the CBM module includes a convolution layer, a BN layer and a sigmoid layer;

[0146] The prediction network includes three parallel prediction branches for common processing, and detects cracks of different sizes at the same time. Specific implementation three:

[0148] The embodiment is a computer storage medium, and the storage medium stores at least one instruction, the at least one instruction is loaded and run by a processor, and a building crack detection system based on deep learning is implemented to realize the building crack detection method based on deep learning.

[0149] It should be understood that the instructions include a computer program product, software or computerized method corresponding to any method described in the present application; the instructions can be used to program a computer system, or other electronic devices. The computer storage medium can include a readable medium having instructions stored thereon, which can include but not limited to magnetic storage medium, optical storage medium; magneto-optical storage medium including read-only memory ROM, random access memory RAM, erasable programmable memory (such as EPROM and EEPROM) and flash layer, or other types of media suitable for storing electronic instructions. DETAILED DESCRIPTION

[0151] The embodiment is a building crack detection device based on deep learning, which comprises a processor and a memory. It should be understood that any device described in the present application includes a processor and a memory, and the device can also include other units, modules, etc. for display, interaction, processing, control, etc. through signals or instructions, and other functions;

[0152] The memory stores at least one instruction, which is loaded and run by the processor to implement the building crack detection method based on deep learning.

[0153] The above examples of the present application are only to illustrate the calculation model and calculation process of the present application, and are not limited to the embodiments of the present application. For those skilled in the art, other different forms of changes or variations can be made on the basis of the above description, and it is impossible to enumerate all the embodiments here. Any obvious changes or variations derived from the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A deep learning-based method for detecting building cracks, characterized in that, First, images of building cracks are acquired, and then a lightweight crack detection network model is used to detect cracks in the acquired building crack images. The lightweight crack detection network model includes: a feature extraction network, a feature aggregation network, and a prediction network; The feature extraction network processing procedure is as follows: The input layer first passes through a 3×3 convolutional block, and then passes through a maximum value layer to downsample the input image to obtain the downsampled feature image; then, based on the obtained downsampled feature image, three sets of improved ShuffleNetV2 networks are used to extract network features. Each improved ShuffleNetV2 network consists of one basic unit SN_2 and multiple basic units SN_1; SN_2: The input is fed into two branches. The first branch passes through a 3×3 depthwise separable convolution and a 1×1 group convolution in sequence. The second branch passes through a 1×1 group convolution, a 3×3 depthwise separable convolution and a 1×1 group convolution in sequence. Then the two branches are concatenated and channels are integrated. SN_1: The input is split into channels and then split into two branches after the Channel Split operation. The first branch is connected to a 3×3 max pooling operation. The second branch is connected to three convolutions with a stride of 1 in sequence: a 1×1 group convolution, a 3×3 depthwise separable convolution, and a 1×1 group convolution. Then the two branches are concatenated and the channels are integrated. The three improved ShuffleNetV2 networks are referred to as layers two through four, and the feature extraction process of layers two through four is as follows: The downsampled feature image is input into the second layer. The second layer first receives SN_2, then receives α1 SN_1, and the output is recorded as the output of the second layer. The features output from the second layer are transmitted to the third layer, and the output is also transmitted to the first branch of the feature aggregation network. The first branch of the feature aggregation network first uses the improved Coordinate Attention model for processing. The third layer first connects to SN_2, then connects to α2 SN_1, and obtains the output of the third layer; the output of the third layer is connected to the fourth layer, and at the same time the output is transmitted to the second branch of the feature aggregation network. The second branch of the feature aggregation network first uses the improved Coordinate Attention model for processing. The fourth layer first connects to SN_2, then connects to α3 SN_1, then connects the output to SN_1, then connects the output to SN_1 again, and the output of the fourth layer is connected to the SPPCSPC module of the third branch of the feature aggregation network, and then fed into the improved CoordinateAttention model for processing. The processing procedure of the feature aggregation network is as follows: The improved Coordinate Attention model in the first to third branches of the feature aggregation network is connected to a convolutional layer. The input features of the improved Coordinate Attention model are subjected to max pooling in the h and w directions, respectively. Then, the pooled features are concatenated by Concat. After concatenation, a 1×1 2D convolution operation is performed on each row. Then, normalization and non-linear processing are performed. The processed features are then divided into h and w directions and 2D convolution operations are performed on each direction. Then, sigmoid processing is performed in each direction to obtain the coordinate weight coefficients. The coordinate weight coefficients are multiplied with the original input features to obtain the weighted features. The output of the convolutional layer corresponding to the improved Coordinate Attention model in the third branch of the feature aggregation network is fed into the first upsampling layer. After upsampling, it is processed by Cat with the output of the convolutional layer corresponding to the improved Coordinate Attention model in the second branch of the feature aggregation network, and then fed into the first ELAN-H layer. The output of the first ELAN-H layer is processed by the improved Coordinate Attention model and convolutional layer and then fed into the second upsampling layer. After upsampling, the output of the convolutional layer corresponding to the improved Coordinate Attention model of the first branch of the feature aggregation network is processed by Cat and then fed into the second ELAN-H layer. The output of the second ELAN-H layer is fed into the first MP-2 layer for processing and also into the first prediction branch of the prediction network for processing. The output of the first MP-2 layer and the input of the first ELAN-H layer are processed by Cat, and then sent to the third ELAN-H layer. The output of the third ELAN-H layer is sent to the second MP-2 layer for processing, and at the same time, it is also sent to the second prediction branch of the prediction network for processing. The output of the second MP-2 layer and the output of the SPPCSPC module of the third branch are processed by Cat, and then sent to the fourth ELAN-H layer. The output of the fourth ELAN-H layer is sent to the third prediction branch of the prediction network for processing. The processing procedure of the prediction network is as follows: The prediction network consists of three parallel prediction branches, each of which includes a REP module and a CBM layer; the REP module in the detection phase includes a 1×1 convolutional layer and a BN layer; the CBM module includes a convolutional layer, a BN layer and a sigmoid layer. The prediction network consists of three parallel prediction branches that work together to detect cracks of different sizes.

2. The method for detecting building cracks based on deep learning according to claim 1, characterized in that, The improved ShuffleNetV2 network corresponding to the second to fourth layers in the feature extraction network includes 3 SN_1, 7 SN_1, and 3 SN_1, respectively, i.e., α1, α2 and α3 are 3, 7 and 3 respectively.

3. A deep learning-based method for detecting building cracks according to claim 1 or 2, characterized in that, The ELAN-H layer processing procedure is as follows: First, the channel is split, dividing the input into two branches. The short branch is connected to a 1×1 CBS module with a step size of 1, while the long branch is first connected to a 1×1 CBS module with a step size of 1, and then to four 3×3 CBS modules with a step size of 1 in sequence. Then, the output of each of the six CBS modules in the short and long branches is processed by Cat, and then processed by a 1×1 CBS module.

4. The deep learning-based building crack detection method according to claim 3, characterized in that, The MP-2 layer processing procedure is as follows: The input is split into two branches. One branch first performs a Maxpool operation and then connects to a 1×1 CBS module with a step size of 1. The other branch connects to a 1×1 CBS module with a step size of 1 and a 3×3 CBS module with a step size of 1 in sequence. Then, the outputs of the two branches are processed by Cat.

5. The deep learning-based building crack detection method according to claim 4, characterized in that, The REP module in the detection phase has a different structure than the REP module in the training phase. The REP module in the detection phase is determined based on the REP module in the training phase. The REP module in the training phase consists of three branches. The first branch first connects to a 1×1 convolutional layer with a stride of 1, and then connects to a BN layer. The second branch first connects to a 1×1 convolutional layer, and then connects to a BN layer. The third branch only connects to one BN layer, and then the results of the three branches are added together. After training is completed, the 1×1 convolutional layer and BN layer of the first branch are retained as the REP module for the testing phase.

6. A building crack detection system based on deep learning, characterized in that, Includes an image acquisition unit and a crack detection unit; Image acquisition unit: used to acquire images of cracks in buildings; Crack Detection Unit: Inputs the acquired building crack images into a lightweight crack detection network model to detect building cracks in the images; The lightweight crack detection network model includes: a feature extraction network, a feature aggregation network, and a prediction network; The feature extraction network processing procedure is as follows: The input layer first passes through a 3×3 convolutional block, and then passes through a maximum value layer to downsample the input image to obtain the downsampled feature image; then, based on the obtained downsampled feature image, three sets of improved ShuffleNetV2 networks are used to extract network features. Each improved ShuffleNetV2 network consists of one basic unit SN_2 and multiple basic units SN_1; SN_2: The input is fed into two branches. The first branch passes through a 3×3 depthwise separable convolution and a 1×1 group convolution in sequence. The second branch passes through a 1×1 group convolution, a 3×3 depthwise separable convolution and a 1×1 group convolution in sequence. Then the two branches are concatenated and channels are integrated. SN_1: The input is split into channels and then split into two branches after the Channel Split operation. The first branch is connected to a 3×3 max pooling operation. The second branch is connected to three convolutions with a stride of 1 in sequence: a 1×1 group convolution, a 3×3 depthwise separable convolution, and a 1×1 group convolution. Then the two branches are concatenated and the channels are integrated. The three improved ShuffleNetV2 networks are referred to as layers two through four, and the feature extraction process of layers two through four is as follows: The downsampled feature image is input into the second layer. The second layer first receives SN_2, then receives α1 SN_1, and the output is recorded as the output of the second layer. The features output from the second layer are transmitted to the third layer, and the output is also transmitted to the first branch of the feature aggregation network. The first branch of the feature aggregation network first uses the improved Coordinate Attention model for processing. The third layer first connects to SN_2, then connects to α2 SN_1, and obtains the output of the third layer; the output of the third layer is connected to the fourth layer, and at the same time the output is transmitted to the second branch of the feature aggregation network. The second branch of the feature aggregation network first uses the improved Coordinate Attention model for processing. The fourth layer first connects to SN_2, then connects to α3 SN_1, then connects the output to SN_1, then connects the output to SN_1 again, and the output of the fourth layer is connected to the SPPCSPC module of the third branch of the feature aggregation network, and then fed into the improved CoordinateAttention model for processing. The processing procedure of the feature aggregation network is as follows: The improved Coordinate Attention model in the first to third branches of the feature aggregation network is connected to a convolutional layer. The input features of the improved Coordinate Attention model are subjected to max pooling in the h and w directions, respectively. Then, the pooled features are concatenated by Concat. After concatenation, a 1×1 2D convolution operation is performed on each row. Then, normalization and non-linear processing are performed. The processed features are then divided into h and w directions and 2D convolution operations are performed on each direction. Then, sigmoid processing is performed in each direction to obtain the coordinate weight coefficients. The coordinate weight coefficients are multiplied with the original input features to obtain the weighted features. The output of the convolutional layer corresponding to the improved Coordinate Attention model in the third branch of the feature aggregation network is fed into the first upsampling layer. After upsampling, it is processed by Cat with the output of the convolutional layer corresponding to the improved Coordinate Attention model in the second branch of the feature aggregation network, and then fed into the first ELAN-H layer. The output of the first ELAN-H layer is processed by the improved Coordinate Attention model and convolutional layer and then fed into the second upsampling layer. After upsampling, the output of the convolutional layer corresponding to the improved Coordinate Attention model of the first branch of the feature aggregation network is processed by Cat and then fed into the second ELAN-H layer. The output of the second ELAN-H layer is fed into the first MP-2 layer for processing and also into the first prediction branch of the prediction network for processing. The output of the first MP-2 layer and the input of the first ELAN-H layer are processed by Cat, and then sent to the third ELAN-H layer. The output of the third ELAN-H layer is sent to the second MP-2 layer for processing, and at the same time, it is also sent to the second prediction branch of the prediction network for processing. The output of the second MP-2 layer and the output of the SPPCSPC module of the third branch are processed by Cat, and then sent to the fourth ELAN-H layer. The output of the fourth ELAN-H layer is sent to the third prediction branch of the prediction network for processing. The processing procedure of the prediction network is as follows: The prediction network consists of three parallel prediction branches, each of which includes a REP module and a CBM layer; the REP module in the detection phase includes a 1×1 convolutional layer and a BN layer; the CBM module includes a convolutional layer, a BN layer and a sigmoid layer. The prediction network consists of three parallel prediction branches that work together to detect cracks of different sizes.

7. A building crack detection system based on deep learning according to claim 6, characterized in that, The ELAN-H layer processing procedure is as follows: First, the channel is split, dividing the input into two branches. The short branch is connected to a 1×1 CBS module with a step size of 1, while the long branch is first connected to a 1×1 CBS module with a step size of 1, and then to four 3×3 CBS modules with a step size of 1 in sequence. Then, the output of each of the six CBS modules in the short and long branches is processed by Cat, and then processed by a 1×1 CBS module.

8. A building crack detection system based on deep learning according to claim 7, characterized in that, The MP-2 layer processing procedure is as follows: The input is split into two branches. One branch first performs a Maxpool operation and then connects to a 1×1 CBS module with a step size of 1. The other branch connects to a 1×1 CBS module with a step size of 1 and a 3×3 CBS module with a step size of 1 in sequence. Then, the outputs of the two branches are processed by Cat.

9. A computer storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor according to any one of claims 6 to 8, a deep learning-based building crack detection system.

10. A deep learning-based building crack detection device, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor as described in any one of claims 6 to 8: a deep learning-based building crack detection system.

Images