CN115423995A - Lightweight curtain wall crack target detection method and system and safety early warning system - Google Patents

Abstract

The invention discloses a lightweight curtain wall crack target detection method and system and a safety early warning system, wherein the target detection method comprises the steps of image collection, image preprocessing, model construction, model training and testing and crack detection, a large number of wall images are collected through the image collection step and preprocessed, then a lightweight curtain wall crack target detection model is constructed, the Ghost convolution is used for replacing the standard convolution in the original network, and the convolution in the C3 structure is replaced by the Ghost convolution, so that the calculated amount can be reduced, and the calculation efficiency is improved; a CA attention mechanism is added at the position where the number of main features extraction network channels is the largest, so that the network pays more attention to valuable features, and the target detection precision of the whole network is improved. The trained model is obtained through the training and testing of the model, and finally the trained model is used for crack detection.

Description

A light-weight curtain wall crack target detection method and system, and safety early warning system

technical field

The invention relates to the technical field of computer vision, in particular to a light-weight curtain wall crack target detection method and system, and a safety early warning system.

Background technique

Object detection is a challenging field in computer vision. Its results have been widely used in many fields, and there have been many practical applications in wall crack detection or wall damage detection. The detection of cracks on the wall is an ever-existing problem, which is more obvious for the taller and taller buildings. The cracks of this kind of high-rise buildings have great hidden dangers to the building itself, and are no longer easy to observe with the naked eye. cracks. But for this kind of crack, it is widespread, and there is no great limitation and cost in using traditional methods for detection.

The YOLO series of algorithms have been updated and iterated for many years. The target detection algorithm of the first-level target type omits the generation of candidate frames. Compared with the second-level target type algorithm itself, it has a certain advantage in detection speed. But also due to the update of the algorithm, the YOLO series target detection algorithm pays more and more attention to the detection accuracy, which makes the network model more and more complex, making the original YOLO model no longer suitable for carrying on common computing units.

It can be seen that the method in the prior art has the technical problem of low detection accuracy.

Contents of the invention

The invention provides a lightweight curtain wall crack target detection method and system, and a safety early warning system, which are used to solve or at least partly solve the technical problem of low detection accuracy existing in the prior art.

In order to solve the above technical problems, the first aspect of the present invention provides a lightweight curtain wall crack target detection method, including:

Collect a large number of wall images;

Preprocessing the collected wall images;

Build a lightweight curtain wall crack target detection model. The model uses an improved target detection network. The improved target detection network includes a backbone feature extraction network, an enhanced feature extraction network, and a detection head. Among them, the backbone feature extraction network uses Ghost convolution to replace the original YOLOv5 The standard convolution in the network, the backbone feature extraction network and the enhanced feature network replace the convolution in the C3 structure of the original YOLOv5 network with Ghost convolution, and add a CA attention module at the position with the largest number of channels in the backbone feature extraction network , to focus on important features related to wall cracks;

Train and test the constructed lightweight curtain wall crack target detection model to obtain a trained lightweight curtain wall crack target detection model;

Use the trained lightweight curtain wall crack target detection model to detect cracks in the image to be detected.

In one embodiment, the collected wall images are preprocessed, including:

Perform data cleaning and sorting on the collected wall images, and uniformly convert the images into PNG or JPG format;

The images are numbered and sorted according to their physical location on the wall.

In one embodiment, the processing of the CA attention module includes:

Using the pooling kernel along the two vertical and horizontal directions for average pooling, a pair of one-dimensional feature encoding along the horizontal and vertical coordinates is obtained, which is encoded as a global receptive field and direction-aware feature representation;

After performing feature fusion on a pair of one-dimensional feature codes, convolution transformation and nonlinear activation function processing are performed, and then the fused features are transformed into two separate tensor sums along the spatial direction, and the two tensor sums Convolve again;

The tensor obtained by re-convolution is normalized using the sigmoid activation function, and then output to the next module in the backbone feature extraction network as a weight to guide the network to learn features in more important channels.

In one embodiment, the lightweight curtain wall crack target detection model of construction is trained, including:

Divide the preprocessed wall image into a training set, a verification set and a test set according to a preset ratio;

Normalize the divided training set into images with consistent resolution;

Input the normalized training set images to the lightweight curtain wall crack target detection model, generate additional features by Ghost convolution, and focus on important features by CA attention mechanism, and focus on important features based on additional features and CA attention mechanism Get the output image features;

Perform regression prediction and positioning on the verification set according to the output image features, verify the training and convergence, and obtain the detection results, crack degree, recall rate and target detection accuracy of cracks in the image during the training process;

Test the trained light-weight curtain wall crack target detection model on the test set. For the images with cracks found during the test, the position of the crack is obtained according to the number of the wall image, and the final target detection accuracy is greater than the threshold The time model is used as a trained lightweight curtain wall crack target detection model.

In one embodiment, using the trained lightweight curtain wall crack target detection model to perform crack detection on the image to be detected, including:

After preprocessing the image to be detected, input the trained lightweight curtain wall crack target detection model to obtain the crack detection result, wherein the crack detection result uses the anchor box to visually mark the crack;

The crack grade and crack shape are distinguished according to the crack degree, and the crack grade and crack shape are marked next to the anchor frame.

In one embodiment, the method further includes: giving an alarm according to the crack level.

Based on the same inventive concept, the second aspect of the present invention provides a lightweight curtain wall crack target detection system, including:

Image collection module, used to collect a large number of wall images;

A preprocessing module is used to preprocess the collected wall images;

The model building module is used to build a lightweight curtain wall crack target detection model. The model uses an improved target detection network. The improved target detection network includes a backbone feature extraction network, an enhanced feature extraction network and a detection head. Among them, the backbone feature extraction network uses Ghost Convolution replaces the standard convolution in the original YOLOv5 network, the backbone feature extraction network and the enhanced feature network replace the convolution in the C3 structure of the original YOLOv5 network with Ghost convolution, and place the largest number of channels in the backbone feature extraction network Add a CA attention module to focus on important features related to wall cracks;

The training and testing module is used to train and test the constructed lightweight curtain wall crack target detection model to obtain a trained lightweight curtain wall crack target detection model;

The crack detection module is used to use the trained lightweight curtain wall crack target detection model to detect cracks on the image to be detected.

Based on the same inventive concept, the third aspect of the present invention provides a security alarm system, including the light-weight curtain wall crack target detection system described in the second aspect and an alarm module, the alarm module is used for alarming according to the crack level.

Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and operable on the processor, when the processor executes the program, the first A method as described in one aspect.

Compared with prior art, advantage of the present invention and beneficial technical effect are as follows:

A lightweight curtain wall crack target detection method provided by the present invention includes image collection, image preprocessing, model construction, model training and testing, and crack detection steps, and the constructed lightweight curtain wall crack target detection model is implemented on the YOLOv5 network Improvement, use Ghost convolution to replace the standard convolution in the original network, replace the convolution in the C3 structure with Ghost convolution, then replace C3 with GhostC3, which can reduce the amount of calculation and improve calculation efficiency; add a CA attention mechanism In the position where the number of backbone feature extraction network channels is the largest, the network pays more attention to valuable features, which improves the target detection accuracy of the entire network. That is to say, the present invention uses a lightweight target detection model, does not require a large number of computing units, can detect wall cracks in real time and find the location of the cracks, and improves detection accuracy while ensuring efficiency.

Further, the present invention also provides a safety alarm system, through which the alarm module performs different levels of alarms according to the level of cracks.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are For some embodiments of the present invention, those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a schematic flow chart of a light-weight curtain wall crack target detection method disclosed in an embodiment of the present invention;

Fig. 2 is a schematic diagram of a structural diagram of a lightweight curtain wall crack target detection model (GC-YOLOv5 network) and a basic YOLOv5s network structural diagram in an embodiment of the present invention;

FIG. 3 is a schematic diagram of the working mode of Ghost convolution in the embodiment of the present invention;

Fig. 4 is the structural diagram of CA attention mechanism in the embodiment of the present invention;

Fig. 5 is a flow chart of implementing a safety alarm by using a lightweight curtain wall crack target detection model in an embodiment of the present invention.

detailed description

The invention discloses a light-weight curtain wall crack target detection method, which includes the steps of image collection, image preprocessing, model construction, model training and testing, and crack detection. The image collection step is to take pictures of the wall with drones and other tools. The image preprocessing will obtain the picture number and perform data cleaning and format conversion. The model construction step builds a lightweight curtain wall crack target detection model, which can be conveniently Deployed on devices such as drones, combined with computing units (CPU and GPU, etc.) Test to obtain a model with high detection accuracy. The crack classification detection step is to use the trained model to detect the wall to be inspected, and classify and mark the anchor frame according to the shape and grade of the crack. This method is based on target detection technology to identify wall cracks. The advantage of this method is that it uses a lightweight target detection model, does not require a large number of computing units, and can detect wall cracks and find the location of cracks in real time. The detection accuracy is improved while ensuring the detection efficiency.

Further, the present invention also discloses a safety alarm system, which sends out alarms of different levels according to the grades of anchor frames marked in the crack classification and detection step, so as to ensure that the cracks are dealt with.

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Embodiment one

An embodiment of the present invention provides a lightweight curtain wall crack target detection method, including:

Collect a large number of wall images;

Preprocessing the collected wall images;

Build a lightweight curtain wall crack target detection model. The model uses an improved target detection network. The improved target detection network includes a backbone feature extraction network, an enhanced feature extraction network, and a detection head. Among them, the backbone feature extraction network uses Ghost convolution to replace the original YOLOv5 The standard convolution in the network, the backbone feature extraction network and the enhanced feature network replace the convolution in the C3 structure of the original YOLOv5 network with Ghost convolution, and add a CA attention module at the position with the largest number of channels in the backbone feature extraction network , to focus on important features related to wall cracks;

Train and test the constructed lightweight curtain wall crack target detection model to obtain a trained lightweight curtain wall crack target detection model;

Use the trained lightweight curtain wall crack target detection model to detect cracks in the image to be detected.

Please refer to FIG. 1 , which is a schematic flowchart of a method for detecting cracks in lightweight curtain walls disclosed in an embodiment of the present invention.

In the specific implementation process, drones and other equipment can be used to obtain high-definition images of the wall. In the subsequent training process, they are divided into training images and detection images. The purpose of training images is to train the network model to obtain better detection accuracy. The location of the crack should be marked with the image to help the object detection network model to learn. The detection image is used for actual detection and is directly processed by the target detection network.

The preprocessing of the wall image is mainly to clean the image, convert the format, and number the image for subsequent processing, and determine the location of the crack according to the image number.

The improved target detection network improves the original target detection network YOLOv5, please refer to Figure 2, which is a schematic diagram of the comparison between the structure diagram of the lightweight curtain wall crack target detection model (GC-YOLOv5 network) adopted in the present invention and the basic YOLOv5s network structure diagram .

Improvements mainly include the following parts:

One: Use Ghost convolution to replace the standard convolution in the original network. The Conv structure in YOLOv5 is the basic convolution module, which is composed of ordinary convolution layer, batch normalization layer and activation function, that is, replace the Conv structure with the GhostConv structure;

Two: In the network model, the C3 structure includes some residual structures and three convolutions, which can reduce some parameters. The present invention replaces the convolution in the C3 structure with Ghost convolution, that is, replaces C3 with GhostC3;

Three: Add a CA attention mechanism to the position where the number of backbone feature extraction network channels is the largest, making the network pay more attention to valuable features and improving the target detection accuracy of the entire network.

Please refer to FIG. 3 , which is a schematic diagram of the Ghost convolution working method in the embodiment of the present invention.

是在Ghost卷积中使用常规卷积的比例，而

是使用线性运算的比例，w'和h'是输出特征图的长和宽，c是输出的通道数，k和d分别是常规卷积和先行操作的内核长宽，一般内核长宽相同所以用到的是k*k和d*d，且此处由于设备运算的并行性，一般使得k和d相等。The function of Ghost convolution is to use some relatively cheap linear operations to generate some redundant features, and replace some features generated by standard convolution with redundant features. Although these redundant features consume a small amount of calculation, they can be greatly improved. Detection accuracy, its working method is shown in Figure 3. Among them, n is the number of convolutions used, and s is a coefficient, so that

is the ratio of regular convolution used in Ghost convolution, and

is the ratio of using linear operations, w' and h' are the length and width of the output feature map, c is the number of output channels, k and d are the kernel length and width of the conventional convolution and the preceding operation, respectively, and the general kernel length and width are the same so k*k and d*d are used, and here, due to the parallelism of device operations, k and d are generally equal.

According to the working method of Ghost convolution, compared with the corresponding standard convolution, its speedup ratio is:

In general, the present invention detects wall cracks in real time through the improved YOLO network model GC-YOLOv5, and can detect whether there are wall cracks. At the same time, due to sorting the wall images, the crack position can be obtained, and at the same time, it can be visualized Mark the grade and shape of the crack. This lightweight GC-YOLOv5 target detection network can be run on common tools such as drones, which can ensure real-time detection while ensuring the safety of inspectors and saving manpower.

In one embodiment, the collected wall images are preprocessed, including:

Perform data cleaning and sorting on the collected wall images, and uniformly convert the images into PNG or JPG format;

The images are numbered and sorted according to their physical location on the wall.

In the specific implementation process, the unresolved images are retaken, and the images are also segmented. Number and sort images by column or row.

In one embodiment, the processing of the CA attention module includes:

Using the pooling kernel along the two vertical and horizontal directions for average pooling, a pair of one-dimensional feature encoding along the horizontal and vertical coordinates is obtained, which is encoded as a global receptive field and direction-aware feature representation;

After performing feature fusion on a pair of one-dimensional feature codes, convolution transformation and nonlinear activation function processing are performed, and then the fused features are transformed into two separate tensor sums along the spatial direction, and the two tensor sums Convolve again;

The tensor obtained by re-convolution is normalized using the sigmoid activation function, and then output to the next module in the backbone feature extraction network as a weight to guide the network to learn features in more important channels.

Specifically, the role of the CA attention module is to obtain the importance of each feature channel from the horizontal and vertical spatial directions. In this process, the spatial relationship of the entire image can be obtained. This spatial relationship can help the network Better focus on valuable spatial information. The overall steps of the work of the CA attention module are shown in Figure 4.

The main modules of the convolutional neural network model are convolutions. In addition to the Conv standard convolution and C3 residual convolution modules that have been introduced, there is also an SPPF structure, including three serial convolutions with the same convolution kernel. It is to fuse more features with different resolutions to get more information.

Different from the general improved models in the prior art, the present invention not only uses the GhostConv and GhostC3 modules in the backbone feature extraction network, but also changes the corresponding modules in the subsequent enhanced feature extraction network to GhostConv and GhostC3 modules, through This improvement makes the module itself more lightweight, which can make the model meet the requirements of real-time monitoring. Therefore, the model adopted by the present invention is lighter than the general model that only improves the backbone feature extraction network. In addition, since the model replaces more GhostConv and GhostC3 modules, due to its calculation method, some features will be lost, resulting in a slight decrease in final accuracy. Therefore, the present invention also adds a CA attention module, which uses the attention mechanism to make The network pays more attention to the original high-value features, and reduces the weight of redundant features generated by GhostConv and GhostC3 modules. The combination of the two makes the model highly lightweight while maintaining high detection accuracy.

Specifically, the present invention uses Ghost convolution to replace the standard convolution in the original network. The function of Ghost convolution is to use some linear operations with relatively cheap calculation costs to generate some redundant features, and replace the standard convolution with redundant features. Some features, although these redundant features consume a small amount of calculation, they can greatly improve the detection accuracy, which makes it unnecessary to use a large number of computing units when using this method to identify wall cracks, and can achieve real-time detection of wall cracks. It detects the cracks of the body and finds the location of the cracks and distinguishes the severity level of the cracks, and then gives different levels of alarms.

Furthermore, the present invention adds a CA attention mechanism to the position where the number of backbone feature extraction network channels is the largest, making the network pay more attention to valuable features and improving the target detection accuracy of the entire network. This attention mechanism effectively improves the object detection accuracy that is inevitably reduced due to the use of Ghost convolution. The combination of the two methods makes the target detection network simplified and lightweight while maintaining the detection accuracy.

In one embodiment, the lightweight curtain wall crack target detection model of construction is trained, including:

Divide the preprocessed wall image into a training set, a verification set and a test set according to a preset ratio;

Normalize the divided training set into images with consistent resolution;

Input the normalized training set images to the lightweight curtain wall crack target detection model, generate additional features by Ghost convolution, and focus on important features by CA attention mechanism, and focus on important features based on additional features and CA attention mechanism Get the output image features;

Perform regression prediction and positioning on the verification set according to the output image features, verify the training and convergence, and obtain the detection results, crack degree, recall rate and target detection accuracy of cracks in the image during the training process;

Test the trained light-weight curtain wall crack target detection model on the test set. For the images with cracks found during the test, the position of the crack is obtained according to the number of the wall image, and the final target detection accuracy is greater than the threshold The time model is used as a trained lightweight curtain wall crack target detection model.

During the specific implementation process, the preset ratio can be set according to the actual situation, such as 8:1:1, 7:1:2 and so on. The threshold can also be set according to the actual situation, such as 70%, 80% and so on.

In addition, when the target detection accuracy is less than 80%, it is necessary to expand the data set and return to the data set division step to retrain to ensure the detection accuracy.

In one embodiment, using the trained lightweight curtain wall crack target detection model to perform crack detection on the image to be detected, including:

After preprocessing the image to be detected, input the trained lightweight curtain wall crack target detection model to obtain the crack detection result, wherein the crack detection result uses the anchor box to visually mark the crack;

The crack grade and crack shape are distinguished according to the crack degree, and the crack grade and crack shape are marked next to the anchor frame.

Specifically, according to the crack program, cracks are divided into three grades: mild, moderate, and severe. This visualization result can be used to visually discover crack information during human inspection.

In one embodiment, the method further includes: giving an alarm according to the crack level.

Specifically, when a crack is found in the crack detection step, different levels of alarms are issued according to the crack level. Mild and moderate cracks can be marked with the crack level and crack shape in the submission log, and according to the wall image processing. , mark the specific location of the wall where the crack appears in the log. For severe cracks, an alarm is issued directly, and the drone equipped with the lightweight target detection network should suspend follow-up work and continue to detect the wall until the alarm is eliminated.

Specifically, please refer to FIG. 5 , which is an implementation flowchart of implementing a safety alarm by using a lightweight curtain wall crack target detection model in an embodiment of the present invention.

Compared with the original YOLOv5s, the present invention uses the lighter CG-YOLOv5 network, and uses Ghost convolution to replace the original convolution, so that the amount of calculation is greatly reduced. Ghost convolution uses cheaper linear operations to generate some redundant features, and uses redundant features Fill in the features generated by the original standard convolution. The use of redundant features will inevitably reduce part of the detection accuracy, so the CA attention mechanism is inserted to make the network pay more attention to important features. The CG-YOLOv5 network using these two methods can greatly reduce the complexity of the network while maintaining the detection accuracy, thereby reducing the requirements for hardware.

Embodiment two

Based on the same inventive concept, this embodiment provides a lightweight curtain wall crack target detection system, including:

Image collection module, used to collect a large number of wall images;

A preprocessing module is used to preprocess the collected wall images;

The model building module is used to build a lightweight curtain wall crack target detection model. The model uses an improved target detection network. The improved target detection network includes a backbone feature extraction network, an enhanced feature extraction network and a detection head. Among them, the backbone feature extraction network uses Ghost Convolution replaces the standard convolution in the original YOLOv5 network, the backbone feature extraction network and the enhanced feature network replace the convolution in the C3 structure of the original YOLOv5 network with Ghost convolution, and place the largest number of channels in the backbone feature extraction network Add a CA attention module to focus on important features related to wall cracks;

The training and testing module is used to train and test the constructed lightweight curtain wall crack target detection model to obtain a trained lightweight curtain wall crack target detection model;

The crack detection module is used to use the trained lightweight curtain wall crack target detection model to detect cracks on the image to be detected.

Since the system introduced in the second embodiment of the present invention is the system used to implement the target detection method for lightweight curtain wall cracks in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the system The specific structure and deformation, so it will not be repeated here. All systems used in the method in the first embodiment of the present invention belong to the scope of protection of the present invention.

Embodiment three

Based on the same inventive concept, the present invention also provides a security alarm system, including the light-weight curtain wall crack target detection system described in Embodiment 2 and an alarm module, the alarm module is used for alarming according to the crack level.

Because the system introduced in Embodiment 3 of the present invention is implemented based on the system described in Embodiment 2, and based on the system described in Embodiment 2 of the present invention, those skilled in the art can understand the security alarm system, its specific structure and deformation, so it will not be described here. Let me repeat. Embodiment four

Based on the same inventive concept, the present application also provides a computer device, including storage, a processor, and a computer program stored in the storage and operable on the processor. The method in Embodiment 1 is implemented when the processor executes the above program.

Since the computer equipment introduced in Embodiment 4 of the present invention is the computer equipment used to implement the target detection method for lightweight curtain wall cracks in Embodiment 1 of the present invention, based on the method described in Embodiment 1 of the present invention, those skilled in the art can understand The specific structure and deformation of the computer device will not be repeated here. All computer equipment used in the method in Embodiment 1 of the present invention belongs to the scope of protection intended by the present invention.

Those skilled in the art should understand that the embodiments of the present invention may be provided as methods, systems, or computer program products. Accordingly, the present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

While preferred embodiments of the present invention have been described, additional changes and modifications can be made to these embodiments by those skilled in the art once the basic inventive concept is appreciated. Therefore, it is intended that the appended claims be construed to cover the preferred embodiment as well as all changes and modifications which fall within the scope of the invention.

Apparently, those skilled in the art can make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. In this way, if the modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and equivalent technologies, the present invention also intends to include these modifications and variations.

Claims (9)

1. A lightweight curtain wall crack target detection method is characterized by comprising the following steps:

collecting a large number of wall images;

preprocessing the collected wall body image;

constructing a lightweight curtain wall crack target detection model, wherein the model adopts an improved target detection network, the improved target detection network comprises a trunk feature extraction network, an enhanced feature extraction network and a detection head, the trunk feature extraction network uses Ghost convolution to replace standard convolution in the original YOLOv5 network, the trunk feature extraction network and the enhanced feature network replace convolution in the C3 structure of the original YOLOv5 network with Ghost convolution, and a CA attention module is added at the position with the maximum trunk feature extraction network channel number to pay attention to important features related to wall cracks;

training and testing the constructed lightweight curtain wall crack target detection model to obtain a trained lightweight curtain wall crack target detection model;

and carrying out crack detection on the image to be detected by using the trained lightweight curtain wall crack target detection model.

2. The method for detecting the crack target of the light-weight curtain wall as claimed in claim 1, wherein the preprocessing of the collected wall body image comprises:

cleaning and sorting the collected wall body images, and uniformly converting the images into PNG or JPG formats;

and numbering and sequencing according to the physical positions of the images on the wall.

3. The method for detecting the crack target of the light-weight curtain wall as claimed in claim 1, wherein the CA attention module comprises:

performing average pooling by using pooling kernels along two vertical and horizontal directions to obtain a pair of one-dimensional feature codes along a horizontal coordinate and a vertical coordinate, wherein the pair of feature codes are feature representations with global receptive field and direction perception;

performing feature fusion on a pair of one-dimensional feature codes, then performing convolution transformation and nonlinear activation function processing, then changing the fused features into two independent tensor sums along the space direction again, and performing convolution again on the two tensor sums;

and carrying out normalization processing on the tensor obtained by the convolution again and the sigmoid activation function, and outputting the tensor to the next module in the trunk feature extraction network to be used as a weight for guiding the network to learn more important features in the channel.

4. The method for detecting the crack target of the lightweight curtain wall as claimed in claim 2, wherein training the constructed lightweight curtain wall crack target detection model comprises:

dividing the preprocessed wall body image into a training set, a verification set and a test set according to a preset proportion;

carrying out normalization processing on the divided training sets to obtain images with consistent resolution;

inputting the training set image after normalization processing into a lightweight curtain wall crack target detection model, generating additional features by Ghost convolution, paying attention to the important features by a CA attention mechanism, and obtaining output image features according to the additional features and the important features paid attention by the CA attention mechanism;

performing regression prediction and positioning on the verification set according to the output image characteristics, verifying training and convergence conditions, and obtaining a detection result, a crack degree, a recall rate and target detection precision of cracks in the image in the training process;

and testing the trained lightweight curtain wall crack target detection model on a test set, obtaining the position of a crack of the image with the crack found in the test process according to the serial number of the wall body image, and taking the model with the final target detection precision larger than the threshold value as the trained lightweight curtain wall crack target detection model.

5. The method for detecting the crack target of the light-weight curtain wall as claimed in claim 1, wherein the step of performing crack detection on the image to be detected by using the trained light-weight curtain wall crack target detection model comprises the following steps:

preprocessing an image to be detected, and inputting the preprocessed image into a trained lightweight curtain wall crack target detection model to obtain a crack detection result, wherein the crack detection result uses anchor frame visual crack marking;

and distinguishing the crack grade and the crack shape according to the crack degree, and marking the crack grade and the crack shape beside the anchor frame.

6. The method for detecting the crack target of the light-weight curtain wall as claimed in claim 5, wherein the method further comprises: and alarming according to the crack grade.

7. The utility model provides a light-weight curtain wall crack target detecting system which characterized in that includes:

the image collection module is used for collecting a large number of wall images;

the preprocessing module is used for preprocessing the collected wall body images;

the model construction module is used for constructing a lightweight curtain wall crack target detection model, the model adopts an improved target detection network, the improved target detection network comprises a trunk feature extraction network, a reinforced feature extraction network and a detection head, wherein the trunk feature extraction network uses a Ghost convolution to replace a standard convolution in an original YOLOv5 network, the trunk feature extraction network and the reinforced feature network replace a convolution in a C3 structure of the original YOLOv5 network with the Ghost convolution, and a CA attention module is added at the position with the largest number of main feature extraction network channels and used for paying attention to important features related to wall cracks;

the training and testing module is used for training and testing the constructed lightweight curtain wall crack target detection model to obtain a trained lightweight curtain wall crack target detection model;

and the crack detection module is used for carrying out crack detection on the image to be detected by utilizing the trained lightweight curtain wall crack target detection model.

8. A safety warning system, characterized by comprising the light-weight curtain wall crack target detection system as claimed in claim 7 and a warning module, wherein the warning module is used for giving a warning according to the crack grade.

9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method according to any of claims 1 to 6 when executing the program.

Images