CN116977857B - Tunnel crack automatic detection method based on deep learning - Google Patents
Tunnel crack automatic detection method based on deep learning Download PDFInfo
- Publication number
- CN116977857B CN116977857B CN202310986133.XA CN202310986133A CN116977857B CN 116977857 B CN116977857 B CN 116977857B CN 202310986133 A CN202310986133 A CN 202310986133A CN 116977857 B CN116977857 B CN 116977857B
- Authority
- CN
- China
- Prior art keywords
- network
- anchor
- yolov
- clustering
- tunnel
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 71
- 238000013135 deep learning Methods 0.000 title claims abstract description 18
- 238000000034 method Methods 0.000 claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 25
- 238000000605 extraction Methods 0.000 claims abstract description 23
- 230000006872 improvement Effects 0.000 claims description 15
- 230000000694 effects Effects 0.000 description 7
- 238000003064 k means clustering Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 238000004873 anchoring Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/176—Urban or other man-made structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a tunnel crack automatic detection method based on deep learning, which comprises the following steps: selecting images with cracks from the acquired tunnel images, marking detection targets on the images, and forming a data set required by network training by the marked images; generating a priori Anchor frame Anchor in Yolov network by adopting an improved AP clustering algorithm, and replacing a feature extraction module in Yolov network by using a GhostNet feature extraction module; training a model in the improved Yolov network by using the marked dataset, and selecting a training model with highest precision as a final tunnel crack detection model; and performing crack detection by using a tunnel crack detection model. According to the invention, the improved AP clustering method is introduced into automatic detection of the cracks, and the characteristic extraction method of the Yolov network is improved based on GhostNet module, so that the precision of tunnel crack detection can be improved.
Description
Technical Field
The invention relates to the technical field of remote sensing image processing, in particular to a tunnel crack automatic detection method based on deep learning.
Background
Over time, cracks may inevitably occur in the tunnel inner surface due to various external and internal factors. If the crack position can not be detected rapidly and accurately and repaired in time, the normal use of the tunnel can be affected, the tunnel structure can be further damaged seriously, and even the functions of the tunnel structure are damaged, so that huge losses in manpower and financial resources are brought. The traditional tunnel detection is mainly performed by manual operation, an inspector observes through naked eyes, and the inspector uses cameras, crack width measuring instruments, infrared thermometers, ultrasonic flaw detectors, rebound instruments, geological radars and other instruments to assist in identifying diseases. The working time of the subway tunnel in operation for inspection and maintenance is limited, the operation environment is severe, and certain safety risks are brought to detection personnel. Therefore, a convenient and high-precision automatic detection method is needed to replace the traditional detection means. The automatic detection method not only can save a great deal of labor cost, but also can improve the detection precision and quality and ensure the safety of the tunnel structure.
Ren Song et al utilized inception V network as a feature extraction network in SSD (Single Shot multibox Detector) network for tunnel lining crack and water leakage detection. Xue Yadong and the like adopt an improved GoogLeNet model to detect water leakage, cracks, joints and pipelines apparent to the lining of the subway tunnel. The method has the advantages that classification of the subway tunnel lining image crack areas is realized through the improved Alexnet depth convolution network, targets such as cracks and the like are detected by utilizing the SSD full convolution network, and the method has large calculated amount and long time consumption for connecting the connected areas of the images which do not contain diseases with the external rectangle. Long Xuejun, and the like, adopt Yolov model to realize the detection of the crack on the inner wall of the tunnel, and adopt image processing algorithms such as Hessian matrix calculation, iterative self-adaptive segmentation, region growth and the like to realize the extraction of the geometric parameters of the crack. Tunnel cracks have the characteristics of complex background, unobvious cracks, uneven illumination and the like, so that tunnel surface crack extraction is an important research subject, and the existing deep learning related method realizes crack detection of structures such as subways, high-speed tunnels and roads by optimizing structures and parameters of various networks such as CNN, SSD, FCN. The deep learning-based algorithm goes through the development process from a simple convolutional neural network to a complex target detection network, and the crack characteristics are extracted through self-adaptive learning, so that the detection and segmentation of cracks are realized. Under a specific application scene, the detection precision of the existing deep learning method is still to be improved according to the past, and the crack detection effect in a real tunnel environment is still poor. Therefore, it is necessary to study a new and more targeted tunnel crack detection method to effectively overcome the above-mentioned difficulties.
Disclosure of Invention
Aiming at the problems of complex background, insignificant cracks, uneven illumination, low detection precision and the like in tunnel crack detection, the invention provides a tunnel crack automatic detection method based on deep learning, so as to improve the precision of tunnel crack detection.
A tunnel crack automatic detection method based on deep learning comprises the following steps:
Step one, data marking: selecting an image with a crack from the acquired tunnel images, marking detection targets on the image, wherein the marked targets comprise the crack and other targets which are easy to be misjudged as the crack, and the marked images form a data set required by network training;
step two, improving Yolov network: generating a priori Anchor frame Anchor in Yolov network by adopting an improved AP clustering algorithm, and replacing a feature extraction module in Yolov network by using a GhostNet feature extraction module;
step three, model training: training a model in the improved Yolov network by using the marked dataset, and selecting a training model with highest precision as a final tunnel crack detection model;
Step four, crack detection: and (3) performing crack detection by using the tunnel crack detection model selected in the step three.
Further, in the first step, other objects that are easily misjudged as cracks include pipelines, patches and joints.
Further, in the second step, an improved AP clustering algorithm is adopted in the Yolov network to generate a priori Anchor frame Anchor, which specifically includes:
1) Preparing data: extracting the width and the height of object frames of all pictures in the data set to form a two-dimensional array;
2) Setting the clustering number: setting the clustering number K according to the types of objects to be detected;
3) The AP algorithm improvement part provides K clustering centers for the algorithm and gives the algorithm a higher reference degree:
Randomly selecting one sample from the data set as an initial clustering point;
firstly, calculating the shortest distance between each sample and the current existing cluster center, wherein the shortest distance is represented by D (x); then calculating the probability that each sample is selected as the next cluster center, and finally selecting the next cluster center according to a wheel disc method;
Repeating the previous step until K cluster points are selected, and endowing the K cluster points with an initial reference degree of 0.5;
4) And (3) running an AP algorithm: except the K points which are already selected, the initial reference degree of the rest points is 0, the operation is carried out according to an AP algorithm, the final clustering is carried out around the K points, the final clustering center is still near the K points, the clustering centers are combined into a one-dimensional array, and the first K prior Anchor frames Anchor are selected as the final prior Anchor frames Anchor according to the sequence from large to small in area;
5) And (3) importing a model: and importing the newly generated priori Anchor frame Anchor into Yolov network to replace the network self-set Anchor.
Further, the GhostNet feature extraction Module is configured to extract feature information from the dataset, and the core Module of the GhostNet feature extraction Module is Ghost BottleNeck, which is obtained by replacing a general convolution in Inverted Residual with a Ghost Module.
The invention introduces the improved AP clustering method into the automatic detection of the cracks, improves the characteristic extraction method of the Yolov network based on GhostNet module, and has the following beneficial effects:
1. The global optimal clustering information of the priori marked frames is comprehensively utilized, the marked priori frame information can be fully utilized, the most representative Anchor of the global optimal can be selected, and the prediction can be performed more quickly and accurately, so that the detection precision is improved;
2. when the network is constructed, the GhostNet feature extraction module is used for core replacement, and the advanced GhostNet module is used for feature extraction, so that feature information can be extracted and utilized more fully, and feature utilization is more fully, thereby improving the accuracy and the confidence of the network.
Drawings
FIG. 1 is a diagram of three other objects of an embodiment of the present invention, except for a crack;
FIG. 2 is a schematic illustration of an embodiment of the present invention;
FIG. 3 is a block diagram of an embodiment GhostNet of the present invention;
FIG. 4 is a network architecture of embodiments GhoatNet-Yolov of the present invention;
FIG. 5 is a core module structure of embodiment GhostNet of the present invention;
FIG. 6 is a Yolov n versus KG_ Yolov5 network accuracy comparison;
FIG. 7 is a graph showing the results of real tunnel crack detection in accordance with an embodiment of the present invention;
FIG. 8 is a graph of confidence level versus embodiment F1 of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The embodiment of the invention provides a tunnel crack automatic detection method based on deep learning, which comprises the following steps:
Step one, data annotation
The training of the network model for deep learning requires a large number of marked data sets, the marking of the data sets is usually carried out manually by using professional software, the most common software is labelImg, the marked object is an acquired subway tunnel image, and the marked data sets are sent into the network model for training. Specifically, images with cracks are selected from the acquired tunnel images, professional software is used for marking detection targets, the marked images form a data set required by network training, and according to project requirements, the marked targets often comprise cracks and other targets which are easy to be misjudged as the cracks, and as shown in fig. 1, three other targets except the cracks are pipelines, patches and joints.
Step two, improving Yolov network
The random selection of the initial points by the K-means clustering method tends to cause instability of results, namely, the results of each time the K-means clustering algorithm is operated can be different, and the random process causes a significant defect in the K-means clustering algorithm: the convergence situation is strongly dependent on the selection of the initial center. The target object in the tunnel crack detection process has fewer types, and the characteristic extraction method needs to be adaptively improved. According to the embodiment of the invention, an Anchor generation mode which improves an AP clustering algorithm and is used in Yolov networks is adopted, so that the Anchor is more representative, the detection precision is improved, and a GhostNet module is used for fully extracting and utilizing rich characteristic information so as to obtain a network training model which is more robust and higher in precision. The embodiment of the invention is realized based on Yolov networks.
The second step specifically comprises the following steps:
Step 2.1: the AP clustering algorithm is improved to generate an priori Anchor frame Anchor, and the specific flow is as follows:
1) Preparing data: the width and height of the object frame (i.e. the bounding box) of all pictures in the dataset are extracted to form a two-dimensional array.
2) Setting the clustering number: setting the clustering number K according to the types of objects to be detected;
3) The AP algorithm improvement part provides K clustering centers for the algorithm and gives the algorithm a higher reference degree:
Randomly selecting one sample from the data set as an initial clustering point;
Firstly, calculating the shortest distance between each sample and the current existing cluster center (namely the distance between each sample and the nearest cluster center), wherein the shortest distance is represented by D (x); then calculating the probability that each sample is selected as the next cluster center, and finally selecting the next cluster center according to a wheel disc method;
Repeating the previous step until K cluster points are selected, and endowing the K cluster points with an initial reference degree of 0.5;
4) And (3) running an AP algorithm: except the K points which are already selected, the initial reference degree of the rest points is 0, the operation is carried out according to an AP algorithm, the final clustering is carried out around the K points, the final clustering center is still near the K points, the clustering centers are combined into a one-dimensional array, and the first K prior Anchor frames Anchor are selected as the final prior Anchor frames Anchor according to the sequence from large to small in area.
5) And (3) importing a model: and importing the newly generated priori Anchor frame Anchor into Yolov network to replace the network self-set Anchor.
The priori anchoring frames Anchor can provide a certain number of candidate frames in the prediction stage of the network, and the candidate frames are strong in representativeness and can detect cracks more accurately. The improved AP algorithm can be used for optimizing the local optimal problem existing in the original network method, so that the generated prior Anchor frame Anchor tends to be globally optimal, and the method is more representative.
Step 2.2: replacement of feature extraction modules in Yolov5 networks using GhostNet modules
The embodiment of the invention uses GhostNet modules for crack extraction for the first time, uses GhostNet feature extraction modules to replace the feature extraction modules in Yolov5 networks for extracting feature information from a dataset, as shown in fig. 4. The GhostNet module extracts the features deeper, so that the crack detection precision is higher.
In the deep learning network training process, a large number of even redundant feature maps are usually included, so that comprehensive understanding of input data is ensured. Not all feature maps need to be convolved, and can be generated using inexpensive operations such as "Ghost". I.e. a simple linear operation of one of the feature maps for the purpose of generating more similar feature maps, which are considered to be "ghosts" of each other.
The core Module of GhostNet is Ghost BottleNeck, ghost BottleNeck, which is very similar to Inverted Residual in MobileNetv2, and can be considered to be obtained by replacing the general convolution in Inverted Residual with a Ghost Module, as shown in fig. 5, where in the left (a) structure, stride=1, its function is to deepen the depth of the network, and consists of two Ghost modules in series, where the first Ghost Module can expand the number of channels, and the second Ghost Module can reduce the number of channels to be consistent with the input. In the right side (b), to compress the height and width of the feature map to half of the input, a Deepwise convolution of stride=2 is added between the two Ghost modules. And a depth separable convolution with a step length of 2x2 and a common convolution with a step length of 1x1 are added to the residual edge part, so that the alignment of Add operation is ensured.
Step three, model training
Training the model in the Yolov network after improvement by using the marked data set, and selecting the training model with highest precision as a final tunnel crack detection model.
And fourthly, detecting cracks.
And (3) performing crack detection by using the tunnel crack detection model selected in the step three. In order to obtain the best detection accuracy, the acquired image is cut (segmented) to a size close to the training data set according to the requirement, and then the image is input into a model for crack detection.
Step five, verifying the precision
And comparing the influence degree of different improvement methods on the network precision, comparing the precision before and after the network improvement, and quantitatively and qualitatively analyzing the result.
Hardware configuration used in this experiment: windows11 operating system, running memory 64GB, processor: the display card is NVIDIA RTX5000 16GB, based on a Python3.6, pytorch1.10, cuda 10.2.2 deep learning framework, random gradient descent (SGD) optimization is adopted, the initial learning rate is 0.01, the momentum transfer coefficient (momentum) is 0.98, the weight attenuation coefficient is 0.001, the batch size is 32, the total training times epoch are set to 1000, if the precision of the continuous 100 epochs is not improved, the training is stopped, and all models are trained on the same equipment.
(1) Improved AP clustering algorithm for generating Anchor
TABLE 1 training results of New Anchor on Yolov n network
For Yolov n network, respectively testing training effects of the network on an Anchor (old_Anchor) generated by an initial method and an Anchor (New_Anchor) generated by an improved AP clustering method, counting change situations of Precision indexes such as Precision, R, mAP_0.5, mAP_0.5:0.95 and the like, stopping iteration at 635 th epoch by the original network, and stopping iteration at 734 th epoch by the network configured with an improved AP clustering algorithm, wherein the change situations of Precision indexes such as Precision, R, mAP_0.5, mAP_0.5:0.95 and the like are calculated. From the figure, it can be seen that the accuracy jump of the network in the initial training is superior to the Anchor generated by the K-means method in four statistical accuracy aspects after the network is stabilized and Yolov n network for generating the Anchor by the improved AP clustering method is configured. Network P is improved from 67.5% to 75.1%, and 17.6% is improved; network R is improved from 66.8% to 68.9%, and the improvement is 2.1% points; mAP_0.5 is improved from 70% to 72.8%, and is improved by 2.8% points; mAP_0.5:0.9 is improved from 43.9% to 45.3%, and is improved by 1.4% points; as can be seen from the above, on the Yolov n model, the improvement on the accuracy P of the network by the Anchor is most obvious.
In order to further embody improving the improvement of the Anchor on the network precision capability, the influence of two kinds of Anchor on the overall precision of four kinds of modules is respectively tested, and the precision situation is counted as shown in the following table 2. As can be seen from table 2, the improved Anchor has different degrees of improvement in the overall accuracy of the four models, with 10.7 percent improvement in P on MobileNetv modules; p of EFFICIENTNET modules raised by 10.1 percentage points; p on GhostNet modules is lifted to a certain degree; the lifting effect on ShuffleNetv modules is least pronounced.
Table 2 improvement of the accuracy influence of Anchor on each model
| Model | P | R | mAP_0.5 | mAP_0.5：0.95 |
| MobileNetv2 | 0.107 | 0.058 | 0.073 | 0.043 |
| EfficientNet | 0.101 | 0.081 | 0.09 | 0.134 |
| ShuffleNetv2 | 0.015 | 0.011 | 0.011 | 0.001 |
| GhostNet | 0.027 | 0.04 | 0.036 | 0.06 |
The result shows that the detection capability of the network to the tunnel cracks can be improved by improving the Anchor generation mode.
(2) GhostNet module extracts features
After the improved Anchor is configured, the feature extraction module is further replaced by GhostNet, the new network is named kg_ Yolov5, and the accuracy pair with the original Yolov network is as shown in fig. 6. The Yolov n is compared with the improvement method in precision to show the rationality and effectiveness of the improvement method in the text, KG_ Yolov is greatly improved compared with the precision of the original network, wherein the accuracy P is improved by 27.2%, and the recall rate R is improved by 20.1%
(3) Detection result generation
Test comparison was performed using real subway tunnel images, and the results shown in fig. 7 were obtained. The confidence map of the network is shown in fig. 8.
The fig. 7 contains labeled class 4 targets, and it can be seen that the Yolov n network on the left has better detection effect on non-crack targets, but has poor detection effect on cracks, and a plurality of cracks cannot be detected; the detection result of the improved network of the invention can detect the missed detection crack and the confidence of other targets is higher. Fig. 8 is a graph of the F1 value and the confidence level of the kg_ Yolov5 network, where the F1 value can comprehensively reflect the detection capability of the network, and it can be seen that when the confidence level reaches 0.513, the highest F1 value is 0.90, and even when the confidence level is 0.8, the F1 value still approaches 0.8, which indicates that the reliability of the detection result of the network is high.
On the basis of designing and constructing a tunnel crack automatic detection method based on deep learning, the embodiment of the invention comprehensively compares the improved precision change condition and detection effect of the network, and finally can obtain the following conclusion:
1. compared with the original K-means clustering method, the Anchor frame generated by the method is more representative, accords with the shape and size distribution condition of the crack, can detect the crack target more accurately, and solves the problem that the existing method tends to be locally optimal rather than globally optimal. On the basis of deep learning understanding improvement Yolov network, experiments on various modules prove that the novel Anchor generation mode can effectively improve the crack detection precision.
2. Secondly, the use of GhostNet modules improves the Backbone portion of Yolov for the problem of finer cracks and difficult detection, and the new network structure is shown in fig. 4. The GhostNet module can better capture the details and texture characteristics of the crack, the detection accuracy is improved, the overall accuracy reaches 94.3%, the recall rate reaches 86.9%, the average accuracy mAP_0.5 reaches 93.1%, and the average accuracy mAP_0.5:0.95 reaches 75.1%.
3. Finally, the trained improved network model is used for crack detection, so that a good effect is obtained; the improved network not only can detect the missed crack, but also has higher confidence of the target. The improved method can realize the detection requirement of tunnel crack detection in a specific application scene.
The foregoing is merely illustrative embodiments of the present invention, and the present invention is not limited thereto, and any changes or substitutions that may be easily contemplated by those skilled in the art within the scope of the present invention should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.
Claims (3)
1. The tunnel crack automatic detection method based on deep learning is characterized by comprising the following steps of:
Step one, data marking: selecting an image with a crack from the acquired tunnel images, marking detection targets on the image, wherein the marked targets comprise the crack and other targets which are easy to be misjudged as the crack, and the marked images form a data set required by network training;
step two, improving Yolov network: generating a priori Anchor frame Anchor in Yolov network by adopting an improved AP clustering algorithm, and replacing a feature extraction module in Yolov network by using a GhostNet feature extraction module;
step three, model training: training a model in the improved Yolov network by using the marked dataset, and selecting a training model with highest precision as a final tunnel crack detection model;
Step four, crack detection: performing crack detection by using the tunnel crack detection model selected in the step three;
in the second step, an improved AP clustering algorithm is adopted in Yolov networks to generate a priori Anchor frame Anchor, which specifically includes:
1) Preparing data: extracting the width and the height of object frames of all pictures in the data set to form a two-dimensional array;
2) Setting the clustering number: setting the clustering number K according to the types of objects to be detected;
3) The AP algorithm improvement part provides K clustering centers for the algorithm and gives the algorithm a higher reference degree:
randomly selecting a sample from the data set as an initial clustering point;
firstly, calculating the shortest distance between each sample and the current existing cluster center, wherein the shortest distance is represented by D (x); then calculating the probability that each sample is selected as the next cluster center, and finally selecting the next cluster center according to a wheel disc method;
Repeating the previous step until K cluster points are selected, and endowing the K cluster points with an initial reference degree of 0.5;
4) And (3) running an AP algorithm: except the K points which are already selected, the initial reference degree of the rest points is 0, the operation is carried out according to an AP algorithm, the final clustering is carried out around the K points, the final clustering center is still near the K points, the clustering centers are combined into a one-dimensional array, and the first K prior Anchor frames Anchor are selected as the final prior Anchor frames Anchor according to the sequence from large to small in area;
5) And (3) importing a model: and importing the newly generated priori Anchor frame Anchor into Yolov network to replace the network self-set Anchor.
2. The automatic tunnel crack detection method based on deep learning as claimed in claim 1, wherein the other objects which are easily misjudged as cracks in the first step include pipelines, patches and joints.
3. The automatic tunnel crack detection method based on deep learning as claimed in claim 1, wherein the GhostNet feature extraction Module is used for extracting feature information from the dataset, and the core Module of the GhostNet feature extraction Module is Ghost BottleNeck, which is obtained by replacing a general convolution in Inverted Residual with a Ghost Module.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310986133.XA CN116977857B (en) | 2023-08-07 | 2023-08-07 | Tunnel crack automatic detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310986133.XA CN116977857B (en) | 2023-08-07 | 2023-08-07 | Tunnel crack automatic detection method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116977857A CN116977857A (en) | 2023-10-31 |
| CN116977857B true CN116977857B (en) | 2024-06-11 |
Family
ID=88477937
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310986133.XA Active CN116977857B (en) | 2023-08-07 | 2023-08-07 | Tunnel crack automatic detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116977857B (en) |
Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106603324A (en) * | 2015-10-20 | 2017-04-26 | 富士通株式会社 | Training set acquisition device and training set acquisition method |
| CN106778834A (en) * | 2016-11-29 | 2017-05-31 | 北京信息科技大学 | A kind of AP based on distance measure study clusters image labeling method |
| CN107767293A (en) * | 2017-09-20 | 2018-03-06 | 国网浙江省电力公司电力科学研究院 | A kind of larger power user divided method based on improvement AP and K means clusters |
| CN108376302A (en) * | 2018-05-23 | 2018-08-07 | 贵州大学 | The prediction model of ELM complexity petrochemical industry production capacity based on AP clusters |
| CN108460486A (en) * | 2018-03-05 | 2018-08-28 | 国网上海市电力公司 | A kind of voltage deviation prediction technique based on improvement clustering algorithm and neural network |
| CN110543913A (en) * | 2019-09-04 | 2019-12-06 | 西南科技大学 | Genetic algorithm-based neighbor propagation clustering method |
| CN111586567A (en) * | 2020-05-22 | 2020-08-25 | 中国电子科技集团公司第五十四研究所 | Network cooperative positioning method based on anchor node |
| CN111811818A (en) * | 2020-06-02 | 2020-10-23 | 桂林电子科技大学 | Rolling bearing fault diagnosis method based on AP clustering algorithm of specified clustering number |
| CN112598152A (en) * | 2020-11-18 | 2021-04-02 | 天津大学 | High-voltage circuit breaker fault diagnosis method based on improved affine propagation clustering |
| CN112926635A (en) * | 2021-02-01 | 2021-06-08 | 中国人民解放军空军工程大学 | Target clustering method based on iterative adaptive neighbor propagation algorithm |
| CN114913423A (en) * | 2022-05-25 | 2022-08-16 | 中国电建集团成都勘测设计研究院有限公司 | Model training method and extraction method for surrounding rock fracture information |
| CN115131711A (en) * | 2022-07-07 | 2022-09-30 | 重庆科技学院 | Video image flame detection method based on YoloV4-Lite network |
| CN115984234A (en) * | 2023-01-12 | 2023-04-18 | 淮阴工学院 | Improved Yolov 5-based bridge crack detection and classification method |
| CN116030266A (en) * | 2021-10-25 | 2023-04-28 | 南京理工大学 | Pavement crack detection and classification method based on improved YOLOv3 under natural scene |
| CN116310272A (en) * | 2022-12-30 | 2023-06-23 | 西安交通大学 | Fan blade defect detection method based on improved YOLOv4 |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8798349B2 (en) * | 2011-07-05 | 2014-08-05 | The Chinese University Of Hong Kong | Systems and methods for detecting arterial input function (AIF) |
-
2023
- 2023-08-07 CN CN202310986133.XA patent/CN116977857B/en active Active
Patent Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106603324A (en) * | 2015-10-20 | 2017-04-26 | 富士通株式会社 | Training set acquisition device and training set acquisition method |
| CN106778834A (en) * | 2016-11-29 | 2017-05-31 | 北京信息科技大学 | A kind of AP based on distance measure study clusters image labeling method |
| CN107767293A (en) * | 2017-09-20 | 2018-03-06 | 国网浙江省电力公司电力科学研究院 | A kind of larger power user divided method based on improvement AP and K means clusters |
| CN108460486A (en) * | 2018-03-05 | 2018-08-28 | 国网上海市电力公司 | A kind of voltage deviation prediction technique based on improvement clustering algorithm and neural network |
| CN108376302A (en) * | 2018-05-23 | 2018-08-07 | 贵州大学 | The prediction model of ELM complexity petrochemical industry production capacity based on AP clusters |
| CN110543913A (en) * | 2019-09-04 | 2019-12-06 | 西南科技大学 | Genetic algorithm-based neighbor propagation clustering method |
| CN111586567A (en) * | 2020-05-22 | 2020-08-25 | 中国电子科技集团公司第五十四研究所 | Network cooperative positioning method based on anchor node |
| CN111811818A (en) * | 2020-06-02 | 2020-10-23 | 桂林电子科技大学 | Rolling bearing fault diagnosis method based on AP clustering algorithm of specified clustering number |
| CN112598152A (en) * | 2020-11-18 | 2021-04-02 | 天津大学 | High-voltage circuit breaker fault diagnosis method based on improved affine propagation clustering |
| CN112926635A (en) * | 2021-02-01 | 2021-06-08 | 中国人民解放军空军工程大学 | Target clustering method based on iterative adaptive neighbor propagation algorithm |
| CN116030266A (en) * | 2021-10-25 | 2023-04-28 | 南京理工大学 | Pavement crack detection and classification method based on improved YOLOv3 under natural scene |
| CN114913423A (en) * | 2022-05-25 | 2022-08-16 | 中国电建集团成都勘测设计研究院有限公司 | Model training method and extraction method for surrounding rock fracture information |
| CN115131711A (en) * | 2022-07-07 | 2022-09-30 | 重庆科技学院 | Video image flame detection method based on YoloV4-Lite network |
| CN116310272A (en) * | 2022-12-30 | 2023-06-23 | 西安交通大学 | Fan blade defect detection method based on improved YOLOv4 |
| CN115984234A (en) * | 2023-01-12 | 2023-04-18 | 淮阴工学院 | Improved Yolov 5-based bridge crack detection and classification method |
Non-Patent Citations (3)
| Title |
|---|
| 《Gravity theory-based affinity propagation clustering algorithm and its applications》;limin Wang et al.;《Tehnicki Vjesnik》;20180831;全文 * |
| 《基于指定聚类数目AP 算法的滚动轴承故障诊断》;杨庆勇 等;《组合机床与自动化加工技术》;20210430;全文 * |
| 《基于改进的半监督AP 聚类雷达信号分选》;金银波 等;《现代电子技术》;20141001;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116977857A (en) | 2023-10-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20210319561A1 (en) | Image segmentation method and system for pavement disease based on deep learning | |
| CN110660052B (en) | Hot-rolled strip steel surface defect detection method based on deep learning | |
| CN108428229A (en) | It is a kind of that apparent and geometric properties lung's Texture Recognitions are extracted based on deep neural network | |
| CN111985552B (en) | Method for detecting diseases of thin strip-shaped structure of airport pavement under complex background | |
| CN108711148A (en) | A kind of wheel tyre defect intelligent detecting method based on deep learning | |
| CN111178392B (en) | Aero-engine hole detection image damage segmentation method based on deep neural network | |
| CN113962960A (en) | Pavement disease detection method based on deep learning | |
| CN116206112A (en) | Remote sensing image semantic segmentation method based on multi-scale feature fusion and SAM | |
| CN117173568A (en) | Target detection model training method and target detection method | |
| CN114972759A (en) | Remote sensing image semantic segmentation method based on hierarchical contour cost function | |
| CN112489026A (en) | Asphalt pavement disease detection method based on multi-branch parallel convolution neural network | |
| CN116229354A (en) | Face image surrounding rock grade identification method based on characteristic parameter automatic extraction | |
| CN115601544A (en) | High-resolution image landslide detection and segmentation method | |
| CN116152674A (en) | Dam unmanned aerial vehicle image crack intelligent recognition method based on improved U-Net model | |
| Wang et al. | Instance segmentation of soft‐story buildings from street‐view images with semiautomatic annotation | |
| CN116977857B (en) | Tunnel crack automatic detection method based on deep learning | |
| CN116468690B (en) | Subtype analysis system of invasive non-mucous lung adenocarcinoma based on deep learning | |
| Bush et al. | Deep Neural Networks for visual bridge inspections and defect visualisation in Civil Engineering | |
| CN117197085A (en) | Road rapid-inspection image pavement disease detection method based on improved YOLOv8 network | |
| CN116883831A (en) | Method for identifying occluded tunnel surface cracks | |
| CN111815625A (en) | Early lung cancer prediction method based on deep learning | |
| CN110059544B (en) | Pedestrian detection method and system based on road scene | |
| Fakhri et al. | Road crack detection using gaussian/prewitt filter | |
| CN115456973A (en) | Method, device and equipment for establishing leakage water disease detection and identification model | |
| Wijaya et al. | Building crack due to lombok earthquake classification based on glcm features and svm classifier |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |