A concrete member crack detection method based on a deep learning algorithm

Abstract

The application discloses a concrete member crack detection method based on a deep learning algorithm. In the application, based on the ncnn framework, the developer can encapsulate the deep learning algorithm to output the mobile phone software of the artificial intelligence application without relying on the third-party library, so that the portability of the algorithm is improved, the trained YOLOX algorithm is encapsulated through the ncnn framework and an OpenCV-mobile compiling library, transplanted to the smart phone with an Android system, and cracks generated by the actual stress of the reinforced concrete column are detected, the recognition effect is good, the recognition precisions before and after the transplantation are similar, the transplantation method effectively improves the portability and real-time performance of the algorithm while guaranteeing the recognition precision, so that people can detect more efficiently and quickly in the use process, and the accuracy in the detection process is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of concrete component crack detection, and specifically relates to a concrete component crack detection method based on a deep learning algorithm. BACKGROUND

[0002] A concrete component refers to a component made of concrete (a material mixed by cement, sand, aggregate, etc. according to a certain proportion) as a main structural material. The concrete component is usually used in building and civil engineering and bears functions such as support, force transmission and protection. The concrete component can be classified according to its role in the structure and shape characteristics, and common concrete components include but are not limited to: a column (a vertical column) which is perpendicular to the ground and is used to bear the gravity and other loads of the upper structure of a building. A beam (a horizontal beam) which is a horizontal support structure component and is used to bear the transverse load and transmit the load to a column or a wall. A slab (a floor slab) which is a horizontal component used to cover between layers of a building and support the structure on the upper floor and transmit the load to a wall or a column. A wall which is a vertical component used to separate space, support structure or bear external load. A foundation which is located at the bottom of a building and is used to bear the gravity of the building and transmit the load to the soil of a foundation. The concrete component has advantages such as high strength, good durability and convenient construction and is widely used in building and infrastructure engineering and is one of important component types in modern building structures. The concrete component may crack during transportation and preservation, and if the cracking is not detected in time, the safety and stability of the whole building may be affected.

[0003] However, the common detection method usually adopts manual detection, and in the detection process, there are errors and the detection efficiency is low, thereby affecting the efficiency and accuracy of the whole detection. SUMMARY

[0004] The purpose of the present application is to solve the above-mentioned problems and provide a concrete component crack detection method based on a deep learning algorithm.

[0005] The technical solution adopted by the present application is as follows: a concrete component crack detection method based on a deep learning algorithm, characterized by comprising the following steps:

[0006] S1: first, an image acquisition device is used to acquire images of the concrete component;

[0007] S2: a wireless network module is used to remotely transmit the acquired images of the concrete component to the inside of an image preprocessing module;

[0008] S3: the image preprocessing module first performs image noise reduction processing on the images and then performs image filtering processing on the images;

[0009] S4: Start to establish the concrete component crack dataset, and label the initial road disease dataset through the Labelimg plug-in, which contains four types of labels, namely longitudinal cracks, repairs, network cracks, and transverse cracks

[0010] S5: Use the dataset built in step S4 to train the deep learning model, use GSConv light-weight convolution, input the picture in GSConv first, then perform a normal convolution operation and once down-sampling, then perform DWConv deep convolution, and perform Concat operation on the two convolution results;

[0011] S6: Continue to introduce GSBottleNeck and VoVGSCSP based on GSConv, first replace SC (normal convolution) in the neck with GSConv, the calculation cost is 60% to 70% of the standard convolution. But the learning ability of the model is not inferior to that before the change. In the neck structure, flexible use of GSConv, GSbottleneck and VoV-GSCSP can build a slimneck structure

[0012] S7: When the YOLOX algorithm trained in step S6 is used for crack recognition, it usually needs to use a computer combined with a shooting device in a cooperative manner, and the implementation process is improved compared with traditional manual detection and computer image processing methods, but it is still too cumbersome, and the professional shooting device is not portable. In order to facilitate the actual crack detection operation, the trained YOLOX algorithm is transplanted to the mobile phone based on the ncnn framework and OpenCV-mobile compilation library;

[0013] S8: Use the trained deep learning model to detect cracks in new concrete component images, and determine the crack position and shape in the concrete component through the output results of the model;

[0014] S9: Record and save the output detection results, and the entire concrete component crack detection process based on the deep learning algorithm can be ended.

[0015] In a preferred embodiment, the image acquisition device in step S1 includes a support, an intelligent camera, a lens, a light source, and an IO communication module, wherein the intelligent camera, the lens, and the light source are installed at the wrist end of the industrial robot, the IO communication module is connected with the intelligent camera through a cable, the image acquisition device controls the USB high-definition camera to work by using a JetsonNano development board, starts the image shooting function, and acquires the concrete component image.

[0016] In a preferred embodiment, the wireless network module in step S2 is designed based on the protocol stack of Zigbee 2007, which provides application program interface functions, and the calling API can realize the following functions: ① the coordinator node can realize automatic network establishment; ② the terminal node can automatically discover the coordinator node, and if it joins the network, it can realize automatic binding with the coordinator node; ③ the terminal node can periodically send the collected data to the coordinator node in an end-to-end confirmation sending mode; ④ if the terminal node does not receive the confirmation message from the coordinator node, it will unbind with the coordinator node, and then rediscover the coordinator node and establish binding therewith.

[0017] In a preferred embodiment, in step S3, the image denoising processing step is:

[0018] Firstly, the clear image of the concrete member and the noise image corresponding to the clear image of the concrete member are taken as a training sample, and a training set is constructed therefrom;

[0019] Secondly, a key frame concrete member image denoising model is constructed; the key frame image denoising model comprises an image feature encoding unit for encoding the features of a to-be-denoised key frame image, and an image feature decoding unit for decoding the encoded feature map output by the image feature encoding unit; the image feature encoding unit comprises a feature extraction module and 10 encoding modules, and the image decoding unit comprises 10 decoding modules and an image restoration module.

[0020] Thirdly, the constructed key frame concrete member image denoising model is trained by using the training set to obtain a trained key frame image denoising model.

[0021] Fourthly, the to-be-denoised key frame image is input into the trained image denoising model, and a denoised image is output after calculation.

[0022] In a preferred embodiment, in step S3, the image filtering processing is performed by using a bilateral filtering method to denoise the concrete member image; the bilateral filtering is nonlinear to pixel information, and the coordinates of the pixel points in the concrete member image are considered in the calculation range, and the difference of the surrounding region pixel points is also considered in the calculation range; in this way, the processing of the gray level flat region has a similar effect to Gaussian filtering, and the abnormal region is not processed, so that the edge information of the concrete member image is maintained.

[0023] In a preferred embodiment, in step S4, the concrete member crack data set is labeled, which contains four types of labels, namely longitudinal cracks, repair marks, network cracks and transverse cracks. When establishing the concrete member crack data set, a yolo format txt file is produced by using a python version of a data set labeling tool lambelimg. Each picture corresponds to a label file. The cracks are framed according to the minimum circumscribed rectangle, each crack corresponds to a single frame, the frames do not overlap directly, the visible cracks are labeled, and the small cracks and edge cracks are also labeled to verify the detection ability of the model for small targets. The cracks are named “crack”, and a part of the crack-free pictures are added to detect whether the model will misjudge the cracks.

[0024] In a preferred embodiment, in step S5, the crack image library in the labeled data set is selected as the training sample. The initial weight file is yolox_s.pth; single GPU acceleration training is used; the maximum and minimum learning rates of the algorithm are set to 0.01 and 0.0001 respectively, and the learning rate reduction method is selected as cosine annealing; the stochastic gradient descent (SGD) optimizer is selected, and in order to prevent overfitting, the weight decay is set to 0.0005.

[0025] In a preferred embodiment, in step S6, the GS bottleneck module is built by GS onv; and the cross-stage network (GSCPS) module VoV-GSCSP is designed by using a one-shot aggregation method. The VoV-GSCSP module can effectively reduce the complexity of the network without losing accuracy. This method also proposes a parameter memory access cost MAC (Memory Access Cost), which is the storage space required to store the model.

[0026] In a preferred embodiment, in step S7, OpenCV-mobile is a minimized compiled library developed by improving OpenCV software library based on cross-platform computer vision and machine learning, which contains most commonly used image processing functions. ncnn is a high-performance neural network forward calculation dependent framework optimized for mobile phones. In the transplantation process, Android Studio and Visual Studio 2019 software platforms are used. After configuring the protobuf and ncnn framework, some parameters in the YOLOX algorithm structure are modified to adapt to the development of mobile software. An open neural network exchange (ONNX) format file is generated by the toolkit export_onnx.py provided in the YOLOX algorithm, which stores the trained algorithm and facilitates the transfer of the algorithm between different frameworks, and is named yolox.onnx. Then the parameter file yolox-nano.param and the compiled binary file yolox-nano.bin are output by the onnx2ncnn code, and after modifying some parameters, they are imported into the ncnn framework in the Android Studio software, connected to the Redmi K50 Ultra type smart phone based on the Android system customized MIUI system, and the mobile software integrated with the trained YOLOX algorithm is generated, named yolox_demo.

[0027] In a preferred embodiment, in step S8, a yolo format txt file is made by using the python version of the data set labeling tool lambelimg, and each picture corresponds to a labeling file. The cracks are framed according to the minimum circumscribed rectangle, each crack corresponds to a single frame, the frames do not overlap directly, the visible cracks are labeled, and the small cracks and edge cracks are also labeled, verifying the detection ability of the model to small targets. The cracks are named "crack", and a part of the pictures without cracks are added to detect whether the model will misdetect cracks.

[0028] In summary, due to the adoption of the above technical solutions, the present application has the following advantages:

[0029] 1. In this invention, the constructed dataset contains cracks of different shapes, scales, and depths, which can fully verify the practicality and accuracy of the model, providing assistance for building safety and offering a new method for crack detection. Simultaneously, the ncnn framework uses less memory and has a relatively lower CPU usage when running the YOLOX algorithm, resulting in faster benchmark results. Based on the ncnn framework, developers can encapsulate deep learning algorithms into mobile applications for artificial intelligence without relying on third-party libraries, improving the portability of the algorithm. The trained YOLOX algorithm is encapsulated using the ncnn framework and the OpenCV-mobile compilation library, ported to Android smartphones, and used to detect cracks caused by actual stress in reinforced concrete columns. The recognition effect is good, and the recognition accuracy is similar before and after the porting. The porting method effectively improves the portability and real-time performance of the algorithm while ensuring recognition accuracy, allowing users to perform detection more efficiently and quickly, while also reducing errors and improving accuracy during the detection process.

[0030] 2. In this invention, the upsampling operator is replaced with the lightweight operator CARAFE. After improving the structure, training and validation were performed on a public dataset under the same experimental conditions, demonstrating the superiority of the improved algorithm in the crack detection task. Simultaneously, a lightweight transformation of the YOLOv5s backbone structure was performed, further reducing the model network complexity and parameter count, accelerating detection speed, and reducing detection time. Attached Figure Description

[0031] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Reference Figure 1 ,

[0034] Example:

[0035] A method for detecting cracks in concrete components based on deep learning algorithms includes the following steps:

[0036] S1: First, use image acquisition equipment to acquire images of the concrete components;

[0037] S2: Use the wireless network module to remotely transmit the acquired concrete component images to the image preprocessing module.

[0038] S3: The image preprocessing module first performs image denoising processing on the image, and then performs image filtering processing on the image;

[0039] S4: Start to establish the concrete member crack dataset, and label the initial road disease dataset through the Labelimg plug-in, which contains four types of labeling, namely longitudinal cracks, repairs, network cracks and transverse cracks

[0040] S5: Use the dataset built in step S4 to train the deep learning model, use GSConv light weight convolution, input the picture in GSConv first, then perform a normal convolution operation and then down sampling, then through DWConv deep convolution, and then perform Concat operation on the two convolution results;

[0041] S6: On the basis of GSConv, GSBottleNeck and VoVGSCSP are further introduced, GSConv replaces SC (normal convolution) in the neck, and the calculation cost is 60% to 70% of the standard convolution. But the learning ability of the model is not inferior to that before the change. In the neck structure, GSConv, GSbottleneck and VoV-GSCSP can be flexibly used to build a slimneck structure

[0042] S7: When the YOLOX algorithm trained in step S6 is used for crack recognition, it is usually necessary to use a computer in combination with a shooting device in a cooperative manner, and the implementation process is improved compared with traditional manual detection and computer image processing methods, but it is still too cumbersome, and the professional shooting device is not portable. In order to facilitate the actual crack detection operation, the trained YOLOX algorithm is transplanted to the mobile phone based on the ncnn framework and OpenCV-mobile compilation library;

[0043] S8: Use the trained deep learning model to detect cracks in new concrete member images, and determine the crack position and shape in the concrete member through the output results of the model;

[0044] S9: Record and save the output detection results, and the entire concrete member crack detection process based on the deep learning algorithm can be ended.

[0045] In step S1, the image acquisition device includes a support, an intelligent camera, a lens, a light source and an IO communication module, wherein the intelligent camera, the lens and the light source are installed at the wrist end of the industrial robot, the IO communication module is connected with the intelligent camera through a cable, the image acquisition device controls the USB high-definition camera to work by using the JetsonNano development board, starts the image shooting function, and acquires the concrete member image.

[0046] In step S2, the wireless network module is designed based on the protocol stack of Zigbee 2007, which provides application program interface functions, and the calling API can realize the following functions: ① the coordinator node can realize automatic network establishment; ② the terminal node can automatically discover the coordinator node, and if it joins the network, it can realize automatic binding with the coordinator node; ③ the terminal node can periodically send the collected data to the coordinator node in an end-to-end confirmation sending mode; ④ if the terminal node does not receive the confirmation message of the coordinator node, it will unbind with the coordinator node, and then rediscover the coordinator node and establish binding with it.

[0047] In step S3, the image denoising processing steps are as follows:

[0048] Firstly, the clear image of the concrete member and the noise image corresponding to the clear image of the concrete member are taken as a training sample, and a training set is constructed based on the training sample;

[0049] Secondly, a key frame concrete member image denoising model is constructed; the key frame image denoising model comprises an image feature encoding unit for encoding the features of a to-be-denoised key frame image, and an image feature decoding unit for decoding the encoded feature map output by the image feature encoding unit; the image feature encoding unit comprises a feature extraction module and 10 encoding modules, and the image decoding unit comprises 10 decoding modules and an image restoration module.

[0050] Thirdly, the constructed key frame concrete member image denoising model is trained by using the training set, so as to obtain a trained key frame image denoising model.

[0051] Fourthly, the to-be-denoised key frame image is input into the trained image denoising model, and a denoised image is output after calculation.

[0052] In step S3, the image filtering processing is performed by using the bilateral filtering method to denoise the concrete member image; the bilateral filtering is nonlinear to the pixel information, and the difference of the surrounding pixel points is also considered in the calculation range; thus, the processing of the gray level flat area has a similar effect to the Gaussian filtering, and the abnormal area is not processed, so that the edge information of the concrete member image is maintained.

[0053] In step S4, the concrete member crack data set is labeled, which contains four types of labels, namely longitudinal cracks, repair marks, network cracks and transverse cracks. When establishing the concrete member crack data set, a yolo format txt file is produced by using a python version of a data set labeling tool lambelimg. Each picture corresponds to a label file. The cracks are framed according to the minimum circumscribed rectangle, each crack corresponds to a single frame, the frames do not overlap directly, the visible cracks are labeled, and the small cracks and edge cracks are also labeled, so as to verify the detection ability of the model on small targets. The cracks are named as “crack”, and a part of the images without cracks are added to detect whether the model will misjudge the cracks.

[0054] In step S5, the crack image library in the labeled data set is selected as the training sample. The initial weight file is yolox_s.pth; single GPU acceleration training is used; the maximum and minimum learning rates of the algorithm are set to 0.01 and 0.0001 respectively, and the cosine annealing is selected as the learning rate reduction method; the stochastic gradient descent (SGD) optimizer is selected, and in order to prevent overfitting, the weight decay is set to 0.0005.

[0055] In step S6, the GS bottleneck module is built through GS onv; and the cross-stage network (GSCPS) module VoV-GSCSP is designed by using a one-shot aggregation method. The VoV-GSCSP module can effectively reduce the complexity of the network without losing accuracy. The method also proposes a parameter memory access cost MAC (Memory Access Cost), which is the storage space required to store the model.

[0056] In step S7, OpenCV-mobile is a minimized compiled library developed based on the OpenCV software library of cross-platform computer vision and machine learning, containing most commonly used image processing functions. ncnn is a high-performance neural network forward calculation dependent framework optimized for mobile phones. In the transplantation process, Android Studio and Visual Studio 2019 software platforms are used. After configuring the protobuf and ncnn framework, some parameters in the YOLOX algorithm structure are modified to adapt to the development of mobile software. The open neural network exchange (ONNX) format file is generated by the toolkit export_onnx.py provided in the YOLOX algorithm, which stores the trained algorithm and facilitates the transfer of the algorithm between different frameworks, and is named yolox.onnx. Then the parameter file yolox-nano.param and the compiled binary file yolox-nano.bin are output by the onnx2ncnn code, and after modifying some parameters, they are imported into the ncnn framework in the Android Studio software, connected to the Redmi K50 Ultra type smart phone based on the Android system customized MIUI system, and the mobile software integrated with the trained YOLOX algorithm is generated, named yolox_demo.

[0057] In step S8, the yolo format txt file is made by using the python version of the data set labeling tool lambelimg, and each picture corresponds to a labeling file. The cracks are framed according to the minimum enclosing rectangle, each crack corresponds to a single frame, the frames do not overlap directly, the visible cracks are labeled, and the small cracks and edge cracks are also labeled, verifying the detection ability of the model to small targets. The cracks are named "crack", and a part of the pictures without cracks are added to detect whether the model will misdetect the cracks.

[0058] In the application, the constructed data set contains cracks of different shapes, sizes and depths, which can fully test the practicability and accuracy of the model, provide help for building safety and provide a new method for crack detection task. At the same time, the ncnn framework uses less memory before and after the YOLOX algorithm runs, and the CPU occupancy rate is relatively lower when running the YOLOX algorithm, so the benchmark result is faster. Based on the ncnn framework, developers do not need to rely on third-party libraries to encapsulate deep learning algorithms to produce mobile software of artificial intelligence application, to improve the portability of the algorithm. The trained YOLOX algorithm is encapsulated through the ncnn framework and OpenCV-mobile compilation library, transplanted to the smart phone with Android system, and detects the cracks generated by the actual stress of the reinforced concrete column, and the recognition effect is good, and the recognition accuracy before and after transplantation is similar. The transplantation method improves the portability and real-time performance of the algorithm while ensuring the recognition accuracy, so that people can detect more efficiently and quickly in the use process, and also reduces the error in the detection process and improves the accuracy in the detection process.

[0059] In the application, the upsampling part operator is replaced by the lightweight operator CARAFE. After perfecting the structure, training and verification are carried out under the same experimental conditions on the public data set, and the superiority of the improved algorithm in the crack detection task is obtained. At the same time, the backbone structure of yolov5s is converted to be lightweight, further reducing the model network complexity and parameter quantity, accelerating the detection speed and reducing the detection time.

[0060] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0061] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting cracks in a concrete structure based on a deep learning algorithm, characterized in that: The method comprises the following steps: S1: first use the image acquisition device to collect the image of the concrete member; S2: use the wireless network module to remotely transmit the collected concrete member image to the inside of the image preprocessing module; S3: the image preprocessing module first performs image noise reduction processing on the image, and then performs image filtering processing on the image; S4: start to establish the concrete member crack data set, and label the initial road disease data set through the Labelimg plug-in, which contains four types of labels, namely longitudinal cracks, repairs, network cracks and transverse cracks S5: use the data set constructed in step S4 to train the deep learning model, use GSConv light weight convolution, input the picture in GSConv first, then perform a normal convolution operation and then perform a down sampling, then perform DWConv deep convolution, and perform Concat operation on the two convolution results; S6: on the basis of GSConv, continue to introduce GSBottleNeck and VoVGSCSP, first, GSConv replaces SC (normal convolution) in the neck, and the calculation cost is 60% to 70% of the standard convolution; but the learning ability of the model is not inferior to that before the change; in the neck structure, flexible use of GSConv, GSbottleneck and VoV-GSCSP can build a slimneck structure; S7: when the YOLOX algorithm trained in step S6 is used for crack identification, it is usually necessary to use a computer in combination with a shooting device in a cooperative manner, and the implementation process is improved compared with traditional manual detection and computer image processing methods, but it is still too cumbersome, and the professional shooting device has poor portability; in order to facilitate the actual crack detection operation, the trained YOLOX algorithm is transplanted to the mobile phone based on the ncnn framework and OpenCV-mobile compilation library; S8: use the trained deep learning model to detect cracks in new concrete member images, and determine the crack position and shape in the concrete member through the output result of the model; S9: record and save the output detection result, and the whole concrete member crack detection process based on the deep learning algorithm is ended. 2.The concrete member crack detection method based on a deep learning algorithm according to claim 1, characterized in that: In step S1, the image acquisition device comprises a support, an intelligent camera, a lens, a light source and an IO communication module, wherein the intelligent camera, the lens and the light source are installed at the wrist end of the industrial robot, the IO communication module is connected with the intelligent camera through a cable, the image acquisition device controls the USB high-definition camera to work by using the JetsonNano development board, starts the image shooting function, and collects the image of the concrete member. 3.The concrete member crack detection method based on deep learning algorithm according to claim 1, characterized in that: In the step S2, the wireless network module is designed based on the protocol stack of Zigbee2007, which provides application program interface functions, and the calling API can realize the following functions: ① the coordinator node can realize automatic network establishment; ② the terminal node can automatically discover the coordinator node, and if it joins the network, the automatic binding with the coordinator node can be realized; ③ the terminal node can periodically send the collected data to the coordinator node in the end-to-end confirmation sending mode; ④ if the terminal node does not receive the confirmation message of the coordinator node, it will unbind with the coordinator node, and then rediscover the coordinator node and establish the binding with it. 4.The concrete member crack detection method based on a deep learning algorithm of claim 1, wherein: In the step S3, the image denoising processing steps are as follows: Firstly, the clear image of the concrete member and the noise image corresponding to the clear image of the concrete member are taken as a training sample, and a training set is constructed based on the training sample; Secondly, a key frame concrete member image denoising model is constructed; the key frame image denoising model comprises an image feature encoding unit for encoding the features of a to-be-denoised key frame image, and an image feature decoding unit for decoding the encoded feature map output by the image feature encoding unit; the image feature encoding unit comprises a feature extraction module and 10 encoding modules, and the image decoding unit comprises 10 decoding modules and an image restoration module. Thirdly, the constructed key frame concrete member image denoising model is trained by using the training set, so as to obtain a trained key frame image denoising model. Fourthly, the to-be-denoised key frame image is input into the trained image denoising model, and a denoised image is output after calculation. 5.The concrete member crack detection method based on deep learning algorithm according to claim 1, wherein: In the step S3, the image filtering processing is performed by using the bilateral filtering method to denoise the concrete member image; the bilateral filtering is nonlinear to the pixel information, and the coordinates of the pixel points in the concrete member image are considered in the calculation range, and the difference of the surrounding area pixel points is also considered in the calculation range; thus, the processing of the gray level flat area has a similar effect to the Gaussian filtering, and the abnormal area is not processed, so that the edge information of the concrete member image is maintained.

6. The method of claim 1, wherein the method is based on a deep learning algorithm. In the step S4, the concrete member crack data set is labeled, which contains four types of labels, namely longitudinal cracks, repair marks, network cracks and transverse cracks; when the concrete member crack data set is established, a yolo format txt file is produced by using a python version of a data set labeling tool lambelimg, and each picture corresponds to a label file; the cracks are framed according to the minimum circumscribed rectangle, each crack corresponds to a single frame, the frames are not overlapped, the visible cracks are labeled, the small cracks and edge cracks are labeled, and the detection ability of the model for small targets is verified; the cracks are named as "crack", and a part of the pictures without cracks are added to detect whether the model will misdetect the cracks.

7. The method of claim 1, wherein the method is based on a deep learning algorithm. In the step S5, the crack image library in the labeled data set is selected as the training sample; the initial weight file is yolox_s.pth; single GPU acceleration training is used; the maximum and minimum learning rates of the algorithm are set to 0.01 and 0.0001 respectively, and the cosine annealing is selected as the learning rate reduction method; the stochastic gradient descent (SGD) optimizer is selected, and in order to prevent overfitting, the weight decay is set to 0.0005. 8.The concrete member crack detection method based on deep learning algorithm according to claim 1, wherein: In the step S6, the GS bottleneck module is built through GS onv; the cross-stage network (GSCPS) module VoV-GSCSP is designed by using the one-shot aggregation method; the VoV-GSCSP module can effectively reduce the complexity of the network without losing accuracy, and the method simultaneously proposes a parameter memory access cost (MAC), that is, the storage space required to store the model. 9.The concrete member crack detection method based on deep learning algorithm according to claim 1, wherein: In the step S7, OpenCV-mobile is a minimized compiled library improved and developed based on the OpenCV software library of cross-platform computer vision and machine learning, which contains most of the commonly used image processing functions; ncnn is a high-performance neural network forward calculation dependent framework optimized for mobile phones, and in the transplantation process, Android Studio and Visual Studio 2019 software platforms are used for transplantation; after the protobuf and ncnn frameworks are configured, some parameters in the YOLOX algorithm structure are modified to adapt to the development of mobile software, and the open neural network exchange format file is generated by the toolkit export_onnx.py provided in the YOLOX algorithm, which stores the trained algorithm and facilitates the transfer of the algorithm between different frameworks, and is named yolox.onnx; then the parameter file yolox-nano.param and the compiled binary file yolox-nano.bin are output by the onnx2ncnn code, and after modifying some parameters, they are imported into the ncnn framework in the Android Studio software, connected to the Redmi K50 Ultra type intelligent mobile phone based on the MIUI system customized based on the Android system, and the mobile software integrated with the trained YOLOX algorithm is generated, which is named yolox_demo. 10.The concrete member crack detection method based on a deep learning algorithm according to claim 1, wherein: In the step S8, the python version of the data set labeling tool lambelimg is used to produce a yolo format txt file, and each picture corresponds to a labeled file; the smallest circumscribed rectangle is used to frame the cracks, and each crack corresponds to a single frame, the frames do not overlap directly, the visible cracks are labeled, and the small cracks and edge cracks are also labeled, and the detection ability of the model for small targets is verified; the cracks are named "crack", and a part of the pictures without cracks are added to detect whether the model will misdetect the cracks.

Images