A concrete crack detection method and system

By acquiring remote sensing images and image sequence data of concrete areas in real time and extracting features using a crack detection model, the problems of low detection accuracy and low efficiency in existing technologies are solved, and efficient and accurate concrete quality detection is achieved.

CN118864452BActive Publication Date: 2024-11-26GUANGDONG OCEAN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411328945.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-24
Publication Date
2024-11-26
Estimated Expiration
2044-09-24

AI Technical Summary

Technical Problem

Existing ultrasonic CT detectors have low accuracy in detecting concrete quality, especially small defects, and are inefficient in measurement. They are also prone to errors in depth of movement and cannot detect the quality of concrete used in buildings in a timely manner.

Method used

By acquiring remote sensing images of the target concrete area in real time, image features are extracted to determine whether there are cracks. When cracks are detected, image sequence data is acquired, and image morphology processing and segmentation are performed. A pre-trained crack detection model is used to extract appearance, parallax, and motion features, and these features are fused to detect concrete quality.

Benefits of technology

It enables timely detection of concrete quality, reduces detection costs, improves detection efficiency and accuracy, and avoids waste of resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118864452B_ABST
    Figure CN118864452B_ABST
Patent Text Reader

Abstract

The invention discloses a concrete crack detection method and system, comprising: acquiring a first remote sensing image of a target concrete area to determine whether cracks occur in the target concrete area; when cracks are detected in the target concrete area, acquiring first image sequence data corresponding to the target concrete area, and performing image morphological processing on the first image in the first image sequence data to obtain the crack area in each first image; segmenting each first image according to the crack area to obtain a second image containing only the crack area and an optical flow image formed by a plurality of second images; inputting the second image and the optical flow image into a pre-trained crack detection model to obtain crack diffusion characteristics of the target concrete area; based on the crack diffusion characteristics, outputting a concrete quality value corresponding to the target concrete area through a preset concrete quality detection model, thereby reducing the cost of quality detection and improving detection efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a concrete crack detection method and system. Background Art

[0002] Concrete is a building material made by mixing cement as a gelling material, sand and gravel as aggregates, and water in a certain proportion. Concrete has the characteristics of abundant raw materials, low price, and simple production process. Therefore, it is widely used in civil engineering, shipbuilding, and machinery industries. At the same time, the quality of concrete affects the safety of all walks of life.

[0003] The methods for testing the quality of concrete in the prior art include experimental testing and on-site testing. Among them, the experimental testing is to analyze the quality of concrete in the laboratory. This method not only takes a long time, but also in actual situations, the quality of concrete will be damaged over time, and it is impossible to test the quality of concrete used in construction in a timely manner.

[0004] On-site inspection uses an ultrasonic CT detector to inspect concrete. The ultrasonic CT detector adopts a single-transmit and single-receive working mode, that is, it transmits ultrasonic waves at a fixed point in one hole, receives ultrasonic waves point by point at a certain interval within a certain depth range in another hole, measures the initial time of the wave, and obtains the velocity reconstruction image of the inter-hole profile (called velocity CT) through special CT software to understand the quality of concrete. At present, the velocity CT detection instrument used in concrete quality inspection has the following main problems:

[0005] The concrete material used for cast-in-place piles is a typical composite material with viscosity. The various defects formed during the pouring process not only have different speeds, but also show different degrees of absorption and attenuation effects on ultrasonic waves. This absorption and attenuation effect is mainly reflected in the amplitude change and spectrum change of the received waveform signal. For some defects, such as segregation, honeycomb pores, and low-intensity areas at the pile head, the sound wave velocity does not change much. Especially when the defect range is small, the detection accuracy is relatively low, and the use of velocity CT sometimes misses detection. This detection instrument uses a single-channel receiving point-by-point moving measurement method, which has low measurement efficiency and is prone to moving depth errors, resulting in false anomalies. Summary of the invention

[0006] In order to solve the above technical problems, the present invention discloses a concrete crack detection method and system, which are used to detect the quality of concrete in time and reduce the cost of detection.

[0007] In order to achieve the above object, the present invention discloses a concrete crack detection method, comprising:

[0008] Acquire a first remote sensing image of a target concrete area in real time, and extract image features of the first remote sensing image to determine whether cracks occur in the target concrete area according to the image features;

[0009] When cracks are detected in the target concrete area, first image sequence data corresponding to the target concrete area is acquired, and image morphological processing is performed on each first image in the first image sequence data to acquire a crack area corresponding to each first image;

[0010] Segment each of the first images according to the crack region to obtain a continuous single second image containing only the crack region and an optical flow image formed by a plurality of the second images;

[0011] Inputting the second image and the optical flow image into a pre-trained crack detection model, and extracting the appearance features and disparity features corresponding to each of the second images and the motion features corresponding to the optical flow image through the crack detection model;

[0012] The crack detection model is used to fuse the appearance feature, the parallax feature and the motion feature to obtain the crack diffusion feature corresponding to the target concrete area;

[0013] Based on the crack diffusion characteristics, the concrete quality of the target concrete area is detected by a preset concrete quality detection model, and a concrete quality value corresponding to the target concrete area is output.

[0014] A concrete crack detection method disclosed in the present invention timely detects cracks by acquiring a first remote sensing image of a target concrete area in real time. During the detection process, the image features of the first remote sensing image are first acquired to determine whether cracks occur in the target concrete area, and then the concrete quality is detected when cracks are detected to avoid waste of resources. When cracks are detected in the area, the diffusion trend of the cracks is related to the quality of the concrete, and thus the trend features are extracted by acquiring image sequence data of the area within a period of time to detect the quality of the concrete.

[0015] After acquiring the image sequence data, firstly, the crack area in each of the first images is acquired, so as to delete the image corresponding to the area without cracks in the first image according to the crack area, thereby reducing the amount of data to be processed later and improving the efficiency of detection. After acquiring the second image sequence for image segmentation, the appearance features and parallax features of each image are extracted through a preset crack detection model to characterize the volume features of the cracks in each second image, and then the diffusion features of the cracks can be obtained by combining the volume features and the motion features corresponding to the time series. Then, the concrete quality is detected based on the diffusion features, thereby reducing the cost of detection and improving the efficiency of detection.

[0016] As a preferred example, extracting the image features of the first remote sensing image to determine whether cracks occur in the target concrete area according to the image features includes:

[0017] Performing minimum noise separation change processing on the first remote sensing image, extracting pixel values ​​of each band corresponding to the first remote sensing image, and calculating the square sum of the pixel values ​​of each band to generate first spectral images corresponding to different bands;

[0018] Integrate the first spectral images corresponding to different bands to obtain a spectral image corresponding to the first remote sensing image;

[0019] A similarity calculation is performed between the spectral image and a historical spectral image corresponding to each type of concrete crack in a pre-constructed concrete crack data set, so as to determine whether cracks occur in the target concrete area according to the result of the similarity calculation.

[0020] The present invention is based on the fact that different types of concrete have different band pixels and color features in the process of forming cracks due to the molecular composition of concrete. Based on this, the present invention obtains the pixel values ​​of each band of the first remote sensing image to extract the corresponding spectral image and generate the corresponding spectral features. Furthermore, by calculating the similarity between the spectral image corresponding to the first remote sensing image and the spectral image corresponding to each crack image in the constructed crack data set, it is determined whether cracks occur in the current target concrete area according to the similarity, thereby improving the efficiency and accuracy of detection.

[0021] As a preferred example, performing image morphological processing on each first image in the first image sequence data to obtain a crack region corresponding to each first image includes:

[0022] Performing binarization processing on the first image to obtain a grayscale image corresponding to the first image;

[0023] Calculating the grayscale value of each pixel in the grayscale image, and comparing the grayscale value with a preset grayscale threshold, so as to perform pixel segmentation on the grayscale image according to the comparison result to obtain an initial crack area;

[0024] The texture features and color distribution features corresponding to the initial crack region are extracted, the image of the initial crack region is segmented by a clustering algorithm, and the results of the image segmentation are merged to generate the crack region corresponding to the first image.

[0025] The present invention recognizes crack areas by extracting grayscale features, texture features and color distribution features in an image, thereby improving the accuracy of the recognition.

[0026] As a preferred example, segmenting each of the first images according to the crack region to obtain a continuous single second image containing only the crack region and an optical flow image formed by a plurality of the second images includes:

[0027] Performing image segmentation on each of the first images according to the crack area to form second image sequence data corresponding to the first image sequence data;

[0028] Decomposing the second image sequence data by frames to obtain continuous single second images corresponding to the second image sequence data;

[0029] According to the second image sequence data, optical flow features of the second image sequence data are extracted by an optical flow method to obtain continuous optical flow images corresponding to the second image sequence data.

[0030] The present invention segments each of the second images according to the crack region to obtain image data containing only the crack image, thereby reducing the amount of data to be processed later and improving the efficiency of detection. Furthermore, the optical flow image is generated to extract the time series features corresponding to the image, thereby constructing the diffusion features of the crack in time and improving the accuracy of the features.

[0031] As a preferred example, the extracting the appearance features and disparity features corresponding to each of the second images and the extracting the motion features corresponding to the optical flow image by the crack detection model includes:

[0032] Performing channel grayscale calculation and structural grayscale calculation on the second image respectively according to the channel grayscale function and structural grayscale function preset in the first feature extraction network in the crack detection model to obtain the channel grayscale and structural grayscale of the second image;

[0033] According to the channel grayscale and the structural grayscale, an erosion value and an expansion value of the channel grayscale to the structural grayscale are obtained by using an erosion function and an expansion function preset in the first feature extraction network;

[0034] According to the erosion value and the dilation value, multiple opening and closing operations are performed by using an opening operation function and a closing operation function preset in the first feature extraction network to obtain appearance features corresponding to the second image.

[0035] The present invention sets the first feature extraction network to extract the basic image features of the image data. To this end, the present invention sets a channel grayscale function and a structural grayscale function in the first feature extraction network to extract the channel grayscale and structural grayscale of the image. Then, the channel grayscale and structural grayscale are corroded and expanded, and then opening and closing operations are performed on them, thereby obtaining the appearance characteristics of the image.

[0036] As a preferred example, the extracting the appearance features and disparity features corresponding to each of the second images and the extracting the motion features corresponding to the optical flow image by the crack detection model includes:

[0037] Extracting position deviations corresponding to different pixels in the second image respectively through a second feature extraction network preset in the crack detection model, and performing convolution processing on the position deviations to obtain disparity features corresponding to the second image;

[0038] The superposition of displacement optical flows between a plurality of adjacent optical flow images in the continuous optical flow images is calculated by using an optical flow superposition formula preset in a third feature extraction network in the crack detection model to obtain a superimposed optical flow;

[0039] The superimposed optical flow is subjected to convolution processing through the third feature extraction network to extract motion features corresponding to the optical flow image.

[0040] The present invention sets different feature extraction networks in the crack detection model to extract the disparity features of each image, and then characterizes the depth of the crack in combination with the disparity features, thereby comprehensively reflecting the volume of the crack. Furthermore, the displacement optical flow between the optical flow images is calculated to reflect the motion characteristics of each image over time, thereby characterizing the diffusion trend of the crack.

[0041] As a preferred example, the method of fusing the appearance feature, the parallax feature and the motion feature through the crack detection model to obtain the crack diffusion feature corresponding to the target concrete area includes:

[0042] Different weights are assigned to the appearance features and the disparity features by presetting a fully connected layer in the crack detection model, so as to obtain a crack volume feature corresponding to each second image by fusing the appearance features and the disparity features according to the weights;

[0043] The crack volume feature and the motion feature are linearly fused through the fully connected layer to output the crack diffusion feature corresponding to the target concrete area.

[0044] The present invention integrates the temporal and spatial features to comprehensively and accurately calculate the quality of the concrete using the integrated features, thereby improving the accuracy of detection.

[0045] As a preferred example, the detecting the concrete quality of the target concrete area by using a preset concrete quality detection model and outputting the concrete quality value corresponding to the target concrete area includes:

[0046] Inputting the crack propagation characteristics into a preset concrete quality detection model, predicting the crack propagation characteristics respectively through a plurality of different decision trees preset in the concrete quality detection model, and obtaining a plurality of different first concrete quality values;

[0047] The plurality of different concrete quality values ​​are processed by a preset regression calculation function or classification function to obtain a concrete quality value corresponding to the target concrete area.

[0048] The present invention uses different decision trees preset in the concrete quality detection model to predict the quality of the crack diffusion characteristics, obtains a number of different first concrete quality values, and then performs regression processing or classification processing according to the first concrete quality values ​​to obtain the concrete quality value corresponding to the target concrete area, thereby improving the accuracy of concrete quality detection.

[0049] On the other hand, the present invention discloses a concrete crack detection system, including a region detection module, a crack detection module, an image segmentation module, a feature extraction module, a feature fusion module and a quality detection module;

[0050] The region detection module is used to acquire a first remote sensing image of a target concrete region in real time, and extract image features of the first remote sensing image, so as to determine whether cracks occur in the target concrete region according to the image features;

[0051] The crack detection module is used for acquiring first image sequence data corresponding to the target concrete area when cracks are detected in the target concrete area, and performing image morphological processing on each first image in the first image sequence data to acquire a crack area corresponding to each first image;

[0052] The image segmentation module is used to segment each of the first images according to the crack area to obtain a continuous single second image containing only the crack area and an optical flow image formed by a plurality of the second images;

[0053] The feature extraction module is used to input the second image and the optical flow image into a pre-trained crack detection model, and extract the appearance features and disparity features corresponding to each of the second images and the motion features corresponding to the optical flow image through the crack detection model;

[0054] The feature fusion module is used to fuse the appearance feature, the parallax feature and the motion feature through the crack detection model to obtain the crack diffusion feature corresponding to the target concrete area;

[0055] The quality detection module is used to detect the concrete quality of the target concrete area based on the crack diffusion characteristics through a preset concrete quality detection model, and output a concrete quality value corresponding to the target concrete area.

[0056] The concrete crack detection system disclosed in the present invention timely detects cracks by acquiring a first remote sensing image of a target concrete area in real time. During the detection process, the image features of the first remote sensing image are first acquired to determine whether cracks occur in the target concrete area, and then the concrete quality is detected when cracks are detected to avoid waste of resources. When cracks are detected in the area, the diffusion trend of the cracks is related to the quality of the concrete, and thus the trend features are extracted by acquiring image sequence data of the area within a period of time to detect the quality of the concrete.

[0057] After acquiring the image sequence data, firstly, the crack area in each of the first images is acquired, so as to delete the image corresponding to the area without cracks in the first image according to the crack area, thereby reducing the amount of data to be processed later and improving the efficiency of detection. After acquiring the second image sequence for image segmentation, the appearance features and parallax features of each image are extracted through a preset crack detection model to characterize the volume features of the cracks in each second image, and then the diffusion features of the cracks can be obtained by combining the volume features and the motion features corresponding to the time series. Then, the concrete quality is detected based on the diffusion features, thereby reducing the cost of detection and improving the efficiency of detection.

[0058] As a preferred example, the area detection module includes a spectrum unit and a crack detection unit;

[0059] The spectral unit is used to perform minimum noise separation change processing on the first remote sensing image, extract pixel values ​​of each band corresponding to the first remote sensing image, and calculate the square sum of the pixel values ​​of each band to generate first spectral images corresponding to different bands; integrate the first spectral images corresponding to different bands to obtain a spectral image corresponding to the first remote sensing image;

[0060] The crack detection unit is used to perform similarity calculation between the spectral image and the historical spectral image corresponding to each type of concrete crack in the pre-constructed concrete crack data set, so as to determine whether cracks occur in the target concrete area according to the result of the similarity calculation.

[0061] The present invention is based on the fact that different types of concrete have different band pixels and color features in the process of forming cracks due to the molecular composition of concrete. Based on this, the present invention obtains the pixel values ​​of each band of the first remote sensing image to extract the corresponding spectral image and generate the corresponding spectral features. Furthermore, by calculating the similarity between the spectral image corresponding to the first remote sensing image and the spectral image corresponding to each crack image in the constructed crack data set, it is determined whether cracks occur in the current target concrete area according to the similarity, thereby improving the efficiency and accuracy of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 : is a flow chart of a concrete crack detection method disclosed in an embodiment of the present invention;

[0063] Figure 2 : is a structural schematic diagram of a concrete crack detection system disclosed in an embodiment of the present invention;

[0064] Figure 3 : A flow chart of a concrete crack detection method disclosed in another embodiment of the present invention. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] Embodiment 1

[0067] This embodiment discloses a concrete crack detection method. Specifically, please refer to the specific implementation process of the detection method. Figure 1 , mainly including steps 101 to 106, the steps mainly include:

[0068] Step 101: acquiring a first remote sensing image of a target concrete area in real time, and extracting image features of the first remote sensing image to determine whether cracks occur in the target concrete area according to the image features.

[0069] In this embodiment, the step mainly includes: performing minimum noise separation change processing on the first remote sensing image, extracting pixel values ​​of each band corresponding to the first remote sensing image, and calculating the square sum of the pixel values ​​of each band to generate first spectral images corresponding to different bands; integrating the first spectral images corresponding to different bands to obtain a spectral image corresponding to the first remote sensing image; performing similarity calculation on the spectral image and the historical spectral image corresponding to each type of concrete crack in the pre-constructed concrete crack data set, so as to judge whether cracks occur in the target concrete area according to the result of the similarity calculation.

[0070] This step is based on the fact that different types of concrete have different band pixels and color features in the process of forming cracks due to the molecular composition of concrete. Based on this, the present invention obtains the pixel values ​​of each band of the first remote sensing image to extract the corresponding spectral image and generate the corresponding spectral features. Furthermore, by calculating the similarity between the spectral image corresponding to the first remote sensing image and the spectral image corresponding to each crack image in the constructed crack data set, it is determined whether cracks occur in the current target concrete area based on the similarity, thereby improving the efficiency and accuracy of detection.

[0071] Step 102: When cracks are detected in the target concrete area, first image sequence data corresponding to the target concrete area is obtained, and image morphological processing is performed on each first image in the first image sequence data to obtain a crack area corresponding to each first image.

[0072] In this embodiment, the step mainly includes: binarizing the first image to obtain a grayscale image corresponding to the first image; calculating the grayscale value of each pixel in the grayscale image, and comparing the grayscale value with a preset grayscale threshold, so as to perform pixel segmentation on the grayscale image according to the comparison result to obtain an initial crack area; extracting texture features and color distribution features corresponding to the initial crack area, performing image segmentation on the initial crack area through a clustering algorithm, and performing regional merging on the results of the image segmentation to generate a crack area corresponding to the first image.

[0073] This step identifies the crack area by extracting grayscale features, texture features and color distribution features in the image, thereby improving the accuracy of the identification.

[0074] Step 103: segment each of the first images according to the crack region to obtain a continuous single second image containing only the crack region and an optical flow image formed by a plurality of the second images.

[0075] In this embodiment, the steps mainly include: performing image segmentation on each of the first images according to the crack area to form second image sequence data corresponding to the first image sequence data; decomposing the second image sequence data by frames to obtain a continuous single second image corresponding to the second image sequence data; and extracting the optical flow features of the second image sequence data by the optical flow method according to the second image sequence data to obtain a continuous optical flow image corresponding to the second image sequence data.

[0076] This step segments each of the second images according to the crack region to obtain image data containing only crack images, thereby reducing the amount of data to be processed later and improving the efficiency of detection. Furthermore, the optical flow image is generated to extract the time series features corresponding to the image, thereby constructing the diffusion features of the cracks in time and improving the accuracy of the features.

[0077] Step 104: Input the second image and the optical flow image into a pre-trained crack detection model, and extract the appearance features and disparity features corresponding to each of the second images and the motion features corresponding to the optical flow image through the crack detection model.

[0078] In this embodiment, this step mainly includes: performing channel grayscale calculation and structural grayscale calculation on the second image according to the channel grayscale function and structural grayscale function preset in the first feature extraction network in the crack detection model, so as to obtain the channel grayscale and structural grayscale of the second image; according to the channel grayscale and the structural grayscale, obtaining the erosion value and expansion value of the channel grayscale to the structural grayscale through the erosion function and the expansion function preset in the first feature extraction network; according to the erosion value and the expansion value, performing multiple opening and closing operations through the opening operation function and the closing operation function preset in the first feature extraction network, so as to obtain the appearance features corresponding to the second image.

[0079] Furthermore, the position deviations corresponding to different pixels in the second image are extracted by a second feature extraction network preset in the crack detection model, and the position deviations are convolved to obtain the disparity features corresponding to the second image; the superposition of displacement optical flows between several adjacent optical flow images in the continuous optical flow image is calculated by an optical flow superposition formula preset in a third feature extraction network in the crack detection model to obtain the superimposed optical flow; the superimposed optical flow is convolved by the third feature extraction network to extract the motion features corresponding to the optical flow image.

[0080] In this step, the first feature extraction network is set to extract the basic image features of the image data. To this end, the present invention sets a channel grayscale function and a structural grayscale function in the first feature extraction network to extract the channel grayscale and structural grayscale of the image. Then, the channel grayscale and structural grayscale are corroded and expanded, and then opening and closing operations are performed on them, thereby obtaining the appearance characteristics of the image.

[0081] Furthermore, by setting different feature extraction networks in the crack detection model, the disparity features of each image are extracted, and then the depth of the crack is characterized by combining the disparity features, and the volume of the crack is fully reflected. Furthermore, the displacement optical flow between the optical flow images is calculated to reflect the motion characteristics of each image over time, and then the diffusion trend of the crack is characterized.

[0082] Step 105: The appearance feature, the parallax feature and the motion feature are integrated through the crack detection model to obtain the crack diffusion feature corresponding to the target concrete area.

[0083] In this embodiment, this step mainly includes: assigning different weights to the appearance features and the disparity features by presetting the fully connected layer of the crack detection model, so as to fuse the appearance features and the disparity features according to the weights to obtain the crack volume features corresponding to each of the second images; linearly fusing the crack volume features and the motion features through the fully connected layer, and outputting the crack diffusion features corresponding to the target concrete area.

[0084] This step integrates the temporal and spatial features to comprehensively and accurately calculate the quality of the concrete using the integrated features, thereby improving the accuracy of detection.

[0085] Step 106: Based on the crack diffusion characteristics, the concrete quality of the target concrete area is detected by a preset concrete quality detection model, and a concrete quality value corresponding to the target concrete area is output.

[0086] In this embodiment, this step mainly includes: inputting the crack diffusion characteristics into a preset concrete quality detection model, predicting the crack diffusion characteristics respectively through several different decision trees preset in the concrete quality detection model, and obtaining several different first concrete quality values; processing the several different concrete quality values ​​through a preset regression calculation function or classification function to obtain the concrete quality value corresponding to the target concrete area.

[0087] This step uses different decision trees preset in the concrete quality detection model to predict the quality of the crack diffusion characteristics, obtains several different first concrete quality values, and then performs regression processing or classification processing based on the first concrete quality values ​​to obtain the concrete quality value corresponding to the target concrete area, thereby improving the accuracy of concrete quality detection.

[0088] On the other hand, this embodiment also discloses a concrete crack detection system. For the specific structure of the system, please refer to Figure 2 , including a region detection module 201, a crack detection module 202, an image segmentation module 203, a feature extraction module 204, a feature fusion module 205 and a quality detection module 206.

[0089] The region detection module 201 is used to acquire a first remote sensing image of a target concrete region in real time, and extract image features of the first remote sensing image, so as to determine whether cracks occur in the target concrete region according to the image features.

[0090] The crack detection module 202 is used to obtain first image sequence data corresponding to the target concrete area when cracks are detected in the target concrete area, and perform image morphological processing on each first image in the first image sequence data to obtain the crack area corresponding to each first image.

[0091] The image segmentation module 203 is used to segment each of the first images according to the crack region to obtain a continuous single second image containing only the crack region and an optical flow image formed by a plurality of the second images.

[0092] The feature extraction module 204 is used to input the second image and the optical flow image into a pre-trained crack detection model, and extract the appearance features and disparity features corresponding to each of the second images and the motion features corresponding to the optical flow image through the crack detection model.

[0093] The feature fusion module 205 is used to fuse the appearance feature, the parallax feature and the motion feature through the crack detection model to obtain the crack diffusion feature corresponding to the target concrete area.

[0094] The quality detection module 206 is used to detect the concrete quality of the target concrete area based on the crack diffusion characteristics through a preset concrete quality detection model, and output a concrete quality value corresponding to the target concrete area.

[0095] In this embodiment, the region detection module 201 includes a spectrum unit and a crack detection unit.

[0096] The spectral unit is used to perform minimum noise separation change processing on the first remote sensing image, extract pixel values ​​of each band corresponding to the first remote sensing image, and calculate the square sum of the pixel values ​​of each band to generate first spectral images corresponding to different bands; integrate the first spectral images corresponding to different bands to obtain a spectral image corresponding to the first remote sensing image.

[0097] The crack detection unit is used to perform similarity calculation between the spectral image and the historical spectral image corresponding to each type of concrete crack in the pre-constructed concrete crack data set, so as to determine whether cracks occur in the target concrete area according to the result of the similarity calculation.

[0098] Embodiment 2

[0099] This embodiment further discloses a concrete crack detection method. Specifically, the implementation process of the detection method is as follows: Figure 3 , mainly including steps 301 to 305, wherein the steps are mainly:

[0100] Step 301: Energy enhancement is performed on a first remote sensing image of a target concrete area by using a preset band separation algorithm to calculate a first spectral image corresponding to the first remote sensing image, and then similarity calculation is performed between the first spectral image and a pre-constructed crack image data set to determine whether cracks occur in the target concrete area.

[0101] In this embodiment, the step mainly includes: performing minimum noise separation change processing on the first remote sensing image, extracting pixel values ​​of each band corresponding to the first remote sensing image, and calculating the square sum of the pixel values ​​of each band to generate first spectral images corresponding to different bands; integrating the first spectral images corresponding to different bands to obtain a spectral image corresponding to the first remote sensing image; performing similarity calculation on the spectral image and the historical spectral image corresponding to each type of concrete crack in the pre-constructed concrete crack data set, so as to judge whether cracks occur in the target concrete area according to the result of the similarity calculation.

[0102] Specifically, a band separation algorithm is used to enhance the spectral features corresponding to the target concrete area, so as to facilitate the extraction of spectral features based on the enhanced spectral features. Specifically, principal component analysis method, minimum noise separation change processing and other methods can be used. Furthermore, after using the band separation algorithm to enhance the spectral features in the first remote sensing image, the square sum of the pixel values ​​of each band is calculated to generate a spectral image corresponding to the pixel values ​​of each band, and the spectral images are normalized to reduce interference in the spectral image. Finally, the spectral images of different bands are added and integrated to generate a spectral image corresponding to the first remote sensing image.

[0103] Furthermore, remote sensing images corresponding to each type of concrete crack are obtained through big data retrieval, and spectral images corresponding to the remote sensing images are extracted according to the same spectral feature extraction method, so as to construct a crack data set corresponding to each type of concrete crack based on the spectral images and the remote sensing images.

[0104] Optionally, the pixel coincidence of the spectral image corresponding to the first remote sensing image and the spectral image and remote sensing image corresponding to each concrete crack in the crack data set is calculated, so as to calculate the similarity between the first remote sensing image and each concrete crack according to the pixel coincidence. The similarity is compared with a preset similarity threshold, so as to determine whether cracks occur in the target concrete area according to the comparison result.

[0105] Specifically, a similarity threshold is set. When the similarities between the first remote sensing image and the first spectral image and the remote sensing image and spectral image of each type of concrete crack are less than the similarity threshold, it is determined that no cracks currently occur in the target concrete area. When the similarity between the first remote sensing image and a type of concrete crack is greater than the similarity threshold, it is determined that cracks occur in the target concrete area.

[0106] Step 302: continuously acquiring first image sequence data corresponding to the target concrete area, and segmenting the crack area of ​​each first image in the first image sequence data to extract a second image corresponding to each first image.

[0107] In this embodiment, the step is specifically as follows: binarizing the first image to obtain a grayscale image corresponding to the first image; calculating the grayscale value of each pixel in the grayscale image, and comparing the grayscale value with a preset grayscale threshold, so as to perform pixel segmentation on the grayscale image according to the comparison result to obtain an initial crack area; extracting texture features and color distribution features corresponding to the initial crack area, performing image segmentation on the initial crack area through a clustering algorithm, and performing regional merging on the results of the image segmentation to generate a crack area corresponding to the first image; performing image segmentation on each of the first images according to the crack area to obtain a second image corresponding to each of the first images.

[0108] Specifically, each of the first images is first subjected to noise reduction processing to obtain a denoised image, and then the denoised image is subjected to grayscale processing to generate a corresponding grayscale image, and the grayscale value corresponding to each pixel in the grayscale image is calculated, and then the grayscale value of the grayscale image corresponding to the pre-saved target concrete area in a crack-free state is used as the grayscale threshold, and the first image is preliminarily segmented according to the pixel points whose grayscale values ​​are greater than the grayscale threshold to generate the corresponding initial crack area.

[0109] Secondly, the initial crack area is segmented into superpixels, and the corresponding texture features and color distribution features are extracted based on the results of the superpixel segmentation. Specifically, the texture of the superpixel segmentation result is removed by a filtering method based on a total variation model to obtain the color distribution features, and the superpixel segmentation result is filtered by a short-time windowed Fourier transformer to obtain the texture features.

[0110] Finally, feature fusion techniques such as linear fusion technology and Bayesian fusion are used to fuse the texture features and the color distribution features to generate fusion features for image segmentation. Then, based on the fusion features, a clustering algorithm is used to perform image segmentation on each of the initial crack areas, and the results of the image segmentation are merged to generate a second image corresponding to each of the initial crack areas.

[0111] Step 303: Use the optical flow method to construct several optical flow images corresponding to the second image, and input the optical flow image and the second image into a pre-trained crack detection model to extract the motion features corresponding to the optical flow image and the appearance features and parallax features corresponding to the second image through the crack detection model.

[0112] In this embodiment, the steps are specifically as follows: first, the channel grayscale and structural grayscale of the second image are calculated according to the first feature extraction network preset in the crack detection model; and the erosion value and expansion value of the channel grayscale to the structural grayscale are obtained through the erosion function and the expansion function preset in the first feature extraction network; according to the erosion value and the expansion value, multiple opening and closing operations are performed through the opening operation function and the closing operation function preset in the first feature extraction network to obtain the appearance features corresponding to the second image.

[0113] Specifically, a first feature extraction network is set in the crack detection model to perform morphological feature extraction and binarization operations on each of the second images, extract morphological features and binary features, and perform feature fusion.

[0114] Furthermore, in the first feature extraction network, and Defined in two-dimensional discrete space and Two discrete functions on For each channel of the second image at point The grayscale function at For structural elements at point The grayscale function at for domain, for domain, then right The extended erosion and extended dilation of are defined as:

[0115] ,

[0116] ,

[0117] in, The operator is the distance operator, and Represents the parameters of extended dilatology, indicating Axis pixel neighborhood, and Axis pixel neighborhood.

[0118] In classical morphology, opening and closing operations are used to remove objects smaller than the structural element. Now we hope to find a method that can directly remove objects smaller than the structural element as a whole, and apply the concept of multi-scale analysis to the extended morphological method. The selection of operators is a key point. In this embodiment, the selection definition is as follows:

[0119] .

[0120] Define the extended open and close operations as:

[0121] ,

[0122] .

[0123] According to the definition of the above opening and closing operations, multiple opening and closing operations are performed to obtain multiple morphological feature images corresponding to the second image, thereby obtaining the appearance features of the second image.

[0124] Secondly, the position deviations corresponding to different pixels in the second image are extracted by a second feature extraction network preset in the crack detection model, and convolution processing is performed on the position deviations to obtain the disparity features corresponding to the second image.

[0125] Specifically, the second feature extraction network preset in the crack detection model traverses each pixel point of the second image, and calculates the position deviation corresponding to each pixel point, and the position deviation is convolved by the convolution kernel preset in the second feature extraction network to obtain a disparity map formed by the position deviation.

[0126] Optionally, the disparity map is converted into a depth map according to a parameter conversion formula preset in the crack detection model, and the disparity feature corresponding to the second image is obtained from the depth map, wherein the expression of the parameter conversion formula is:

[0127]

[0128] Wherein, D is the depth, d is the parallax, B is the baseline length, f is the focal length, and the focal length is in pixels. These are the column coordinates of the principal points of the left and right views, respectively.

[0129] Finally, the superposition of the displacement optical flows between several adjacent optical flow images in the continuous optical flow image is calculated by using the optical flow superposition formula preset in the third feature extraction network in the crack detection model to obtain the superimposed optical flow; the superimposed optical flow is convolved with the third feature extraction network to extract the motion features corresponding to the optical flow image.

[0130] Specifically, the optical flow image is input into the third feature extraction network of the crack detection model, and the third feature extraction network includes a convolution layer, a maximum pooling layer, and a fully connected layer. The optical flow image is input into the third feature extraction network, and appearance information such as texture, contour, and interest points are effectively learned from the optical flow image. With the optical flow image as input, the third feature extraction network consists of 13 convolutional and 3 fully connected layers. The convolution kernel size of the 13 convolutional layers is , with a stride of 1, and are stacked into 5 blocks. The kernel size of the maximum pooling layer is , with a step size of 2, connected after each convolution block, the feature information of the optical flow image is obtained through the convolution layer, and the data of the previous layer is convolved using several convolution kernels. The result is output as a two-dimensional vector. After the convolution operation, several two-dimensional outputs can be obtained, and the motion feature map can be output. The motion feature map can be mapped to the feature vector using the fully connected layer. The corresponding motion features are obtained.

[0131] Step 304: The motion feature, appearance feature and parallax feature are integrated through the crack detection model to obtain crack diffusion features corresponding to the target concrete area.

[0132] In this embodiment, this step mainly includes: assigning different weights to the appearance features and the disparity features by presetting the fully connected layer of the crack detection model, so as to obtain the crack volume features corresponding to each second image by fusing the appearance features and the disparity features according to the weights; linearly fusing the crack volume features and the motion features through the fully connected layer, and outputting the crack diffusion features corresponding to the target concrete area.

[0133] Specifically, an adaptive weight matching algorithm is set in the fully connected layer of the crack detection model to adaptively assign different weights to the appearance features and the disparity features according to the adaptive weight matching algorithm, so as to generate crack volume features obtained by fusing the appearance features and the disparity features according to the weights.

[0134] Furthermore, a mid-term fusion strategy is used to linearly combine the extracted fracture volume features and motion features. The mid-term fusion process is to integrate the extracted fracture volume features and motion features, and then fuse the motion features and fracture volume features to obtain fracture diffusion features. The fused fracture diffusion features are as follows:

[0135]

[0136]

[0137] in, Represents the fracture volume characteristics, represents the motion feature, and f is the fused crack diffusion feature. The fused crack diffusion feature is normalized.

[0138] Step 305: Based on the crack diffusion characteristics, the concrete quality of the target concrete area is detected by a preset concrete quality detection model, and a concrete quality value corresponding to the target concrete area is output.

[0139] In this embodiment, this step is specifically as follows: inputting the crack diffusion characteristics into a preset concrete quality detection model, predicting the crack diffusion characteristics respectively by using several different decision trees preset in the concrete quality detection model, and obtaining several different first concrete quality values; processing the several different concrete quality values ​​by using a preset regression calculation function or classification function, and obtaining the concrete quality value corresponding to the target concrete area.

[0140] Specifically, in this embodiment, the concrete quality detection model is constructed according to the attention-based random forest model. The attention-based random forest model can adaptively adjust the contribution of each decision tree, so that in the calculation process of the concrete quality value, more attention is paid to the features that have important contributions to the target prediction. In this embodiment, the training process of the concrete quality detection model is:

[0141] The constructed concrete crack data set is randomly sampled to generate n different sampling data sets. For each sampling data set, a random forest algorithm is used to generate k decision trees. For each decision tree, an attention mechanism is used to adjust the weight of each feature in the decision tree.

[0142] Assume that the weight of the i-th feature in the j-th decision tree is , then the attention score of this feature is:

[0143]

[0144] In order to adaptively adjust the importance of each feature, an attention mechanism is introduced to adjust the importance of each feature in the calculation process of concrete quality value by learning the attention score of each feature in each decision tree.

[0145] For the jth decision tree, let the attention score vector be , then the attention weight of the i-th feature in the decision tree for:

[0146]

[0147] Where m is the total number of features in the decision tree.

[0148] Attention score The calculation method is:

[0149]

[0150] in, is the weight matrix, is the weighted average of the embedding vectors of all features in the jth decision tree, is the bias term, is the activation function (such as the sigmoid function).

[0151] After generating several different decision trees in the concrete quality detection model, the extracted crack diffusion characteristics are respectively input into each decision tree to calculate the concrete quality value, and the concrete quality value is obtained. Results , and then the final result is obtained by voting or averaging. In this embodiment, a regression method can be used to predict a specific value, and the final prediction result is for:

[0152]

[0153] The crack diffusion characteristics can also be classified to obtain an interval value. For example: (0, 0.5mm], (0.5mm, 1mm], (1mm, 1.5mm], (1.5mm, 2mm], corresponding to 4 categories respectively, the final result is for:

[0154]

[0155] in Represents an indicator function. When the condition in the brackets is true, , otherwise .

[0156] A concrete crack detection method disclosed in this embodiment timely detects cracks by acquiring a first remote sensing image of a target concrete area in real time. During the detection process, the image features of the first remote sensing image are first acquired to determine whether cracks occur in the target concrete area, and then the concrete quality is detected when cracks are detected to avoid waste of resources. When cracks are detected in the area, the diffusion trend of the cracks is related to the quality of the concrete, and thus the trend features are extracted by acquiring image sequence data of the area within a period of time to detect the quality of the concrete.

[0157] After acquiring the image sequence data, firstly, the crack area in each of the first images is acquired, so as to delete the image corresponding to the area without cracks in the first image according to the crack area, thereby reducing the amount of data to be processed later and improving the efficiency of detection. After acquiring the second image sequence for image segmentation, the appearance features and parallax features of each image are extracted through a preset crack detection model to characterize the volume features of the cracks in each second image, and then the diffusion features of the cracks can be obtained by combining the volume features and the motion features corresponding to the time series. Then, the concrete quality is detected based on the diffusion features, thereby reducing the cost of detection and improving the efficiency of detection.

[0158] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting concrete cracks, It is characterized in that include: Acquire a first remote sensing image of a target concrete area in real time, and extract image features of the first remote sensing image to determine whether cracks occur in the target concrete area according to the image features; When cracks are detected in the target concrete area, first image sequence data corresponding to the target concrete area is acquired, and image morphological processing is performed on each first image in the first image sequence data to acquire a crack area corresponding to each first image; Segment each of the first images according to the crack region to obtain a continuous single second image containing only the crack region and an optical flow image formed by a plurality of the second images; Inputting the second image and the optical flow image into a pre-trained crack detection model, and extracting the appearance features and disparity features corresponding to each of the second images and the motion features corresponding to the optical flow image through the crack detection model; The crack detection model is used to fuse the appearance feature, the parallax feature and the motion feature to obtain the crack diffusion feature corresponding to the target concrete area; Based on the crack diffusion characteristics, the concrete quality of the target concrete area is detected by a preset concrete quality detection model, and a concrete quality value corresponding to the target concrete area is output.

2. A concrete crack detection method according to claim 1, It is characterized in that The step of extracting the image features of the first remote sensing image to determine whether cracks occur in the target concrete area according to the image features includes: Performing minimum noise separation change processing on the first remote sensing image, extracting pixel values ​​of each band corresponding to the first remote sensing image, and calculating the square sum of the pixel values ​​of each band to generate first spectral images corresponding to different bands; Integrate the first spectral images corresponding to different bands to obtain a spectral image corresponding to the first remote sensing image; A similarity calculation is performed between the spectral image and a historical spectral image corresponding to each type of concrete crack in a pre-constructed concrete crack data set, so as to determine whether cracks occur in the target concrete area according to the result of the similarity calculation.

3. A concrete crack detection method according to claim 1, It is characterized in that The performing image morphological processing on each first image in the first image sequence data to obtain a crack region corresponding to each first image includes: Performing binarization processing on the first image to obtain a grayscale image corresponding to the first image; Calculating the grayscale value of each pixel in the grayscale image, and comparing the grayscale value with a preset grayscale threshold, so as to perform pixel segmentation on the grayscale image according to the comparison result to obtain an initial crack area; The texture features and color distribution features corresponding to the initial crack region are extracted, the image of the initial crack region is segmented by a clustering algorithm, and the results of the image segmentation are merged to generate the crack region corresponding to the first image.

4. A concrete crack detection method according to claim 3, It is characterized in that The step of segmenting each of the first images according to the crack region to obtain a continuous single second image containing only the crack region and an optical flow image formed by a plurality of the second images includes: Performing image segmentation on each of the first images according to the crack area to form second image sequence data corresponding to the first image sequence data; Decomposing the second image sequence data by frames to obtain continuous single second images corresponding to the second image sequence data; According to the second image sequence data, optical flow features of the second image sequence data are extracted by an optical flow method to obtain continuous optical flow images corresponding to the second image sequence data.

5. A method for detecting concrete cracks according to claim 1, It is characterized in that The extracting the appearance features and the disparity features corresponding to each of the second images and the extracting the motion features corresponding to the optical flow image by using the crack detection model includes: Performing channel grayscale calculation and structural grayscale calculation on the second image respectively according to the channel grayscale function and structural grayscale function preset in the first feature extraction network in the crack detection model to obtain the channel grayscale and structural grayscale of the second image; According to the channel grayscale and the structural grayscale, an erosion value and an expansion value of the channel grayscale to the structural grayscale are obtained by using an erosion function and an expansion function preset in the first feature extraction network; According to the erosion value and the dilation value, multiple opening and closing operations are performed by using an opening operation function and a closing operation function preset in the first feature extraction network to obtain appearance features corresponding to the second image.

6. A concrete crack detection method according to claim 1, It is characterized in that The extracting the appearance features and the disparity features corresponding to each of the second images and the extracting the motion features corresponding to the optical flow image by using the crack detection model includes: Extracting position deviations corresponding to different pixels in the second image respectively through a second feature extraction network preset in the crack detection model, and performing convolution processing on the position deviations to obtain disparity features corresponding to the second image; The superposition of displacement optical flows between a plurality of adjacent optical flow images in the continuous optical flow images is calculated by using an optical flow superposition formula preset in a third feature extraction network in the crack detection model to obtain a superimposed optical flow; The superimposed optical flow is subjected to convolution processing through the third feature extraction network to extract motion features corresponding to the optical flow image.

7. A method for detecting concrete cracks according to claim 1, It is characterized in that The method of fusing the appearance feature, the parallax feature and the motion feature through the crack detection model to obtain the crack diffusion feature corresponding to the target concrete area includes: Different weights are assigned to the appearance features and the disparity features by presetting a fully connected layer in the crack detection model, so as to obtain a crack volume feature corresponding to each second image by fusing the appearance features and the disparity features according to the weights; The crack volume feature and the motion feature are linearly fused through the fully connected layer to output the crack diffusion feature corresponding to the target concrete area.

8. A method for detecting concrete cracks according to claim 1, It is characterized in that The detecting the concrete quality of the target concrete area by using a preset concrete quality detection model and outputting a concrete quality value corresponding to the target concrete area includes: Inputting the crack propagation characteristics into a preset concrete quality detection model, predicting the crack propagation characteristics respectively through a plurality of different decision trees preset in the concrete quality detection model, and obtaining a plurality of different first concrete quality values; The plurality of different concrete quality values ​​are processed by a preset regression calculation function or classification function to obtain a concrete quality value corresponding to the target concrete area.

9. A concrete crack detection system, It is characterized in that It includes area detection module, crack detection module, image segmentation module, feature extraction module, feature fusion module and quality detection module; The region detection module is used to acquire a first remote sensing image of a target concrete region in real time, and extract image features of the first remote sensing image, so as to determine whether cracks occur in the target concrete region according to the image features; The crack detection module is used for acquiring first image sequence data corresponding to the target concrete area when cracks are detected in the target concrete area, and performing image morphological processing on each first image in the first image sequence data to acquire a crack area corresponding to each first image; The image segmentation module is used to segment each of the first images according to the crack area to obtain a continuous single second image containing only the crack area and an optical flow image formed by a plurality of the second images; The feature extraction module is used to input the second image and the optical flow image into a pre-trained crack detection model, and extract the appearance features and disparity features corresponding to each of the second images and the motion features corresponding to the optical flow image through the crack detection model; The feature fusion module is used to fuse the appearance feature, the parallax feature and the motion feature through the crack detection model to obtain the crack diffusion feature corresponding to the target concrete area; The quality detection module is used to detect the concrete quality of the target concrete area based on the crack diffusion characteristics through a preset concrete quality detection model, and output a concrete quality value corresponding to the target concrete area.

10. A concrete crack detection system according to claim 9, It is characterized in that The regional detection module includes a spectrum unit and a crack detection unit; The spectral unit is used to perform minimum noise separation change processing on the first remote sensing image, extract pixel values ​​of each band corresponding to the first remote sensing image, and calculate the square sum of the pixel values ​​of each band to generate first spectral images corresponding to different bands; integrate the first spectral images corresponding to different bands to obtain a spectral image corresponding to the first remote sensing image; The crack detection unit is used to perform similarity calculation between the spectral image and the historical spectral image corresponding to each type of concrete crack in the pre-constructed concrete crack data set, so as to determine whether cracks occur in the target concrete area according to the result of the similarity calculation.