CN117351321A - Single-stage lightweight subway lining cavity recognition method and related equipment - Google Patents

Single-stage lightweight subway lining cavity recognition method and related equipment Download PDF

Info

Publication number
CN117351321A
CN117351321A CN202311102838.7A CN202311102838A CN117351321A CN 117351321 A CN117351321 A CN 117351321A CN 202311102838 A CN202311102838 A CN 202311102838A CN 117351321 A CN117351321 A CN 117351321A
Authority
CN
China
Prior art keywords
data
tunnel
cavity
stage
detection model
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.)
Pending
Application number
CN202311102838.7A
Other languages
Chinese (zh)
Inventor
周延峰
郑庆周
罗享寰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen University
Original Assignee
Shenzhen University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shenzhen University filed Critical Shenzhen University
Priority to CN202311102838.7A priority Critical patent/CN117351321A/en
Publication of CN117351321A publication Critical patent/CN117351321A/en
Pending legal-status Critical Current

Links

Classifications

    • 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
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/02—Systems using reflection of radio waves, e.g. primary radar systems; Analogous systems
    • G01S13/06—Systems determining position data of a target
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/885—Radar or analogous systems specially adapted for specific applications for ground probing
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0475—Generative networks
    • 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
    • G06N3/094—Adversarial learning
    • 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
    • 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
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Remote Sensing (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Electromagnetism (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Geophysics And Detection Of Objects (AREA)

Abstract

The invention discloses a single-stage lightweight subway lining cavity recognition method and related equipment, wherein the method comprises the following steps: tunnel hole defect data are obtained, and tunnel hole pretreatment is carried out on the tunnel hole defect data to obtain pretreatment data, wherein the tunnel defect data at least comprise tunnel hole position information and tunnel hole size information; inputting the preprocessed data into a trained optimized detection model to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by antagonizing network optimization through real data and ground penetrating radar depth convolution to train. Aiming at the problem that the detection result is inaccurate due to the fact that the single-stage method is still insufficient when the tunnel cavity defect data are detected at present, the accurate cavity detection result can be obtained through the method.

Description

Single-stage lightweight subway lining cavity recognition method and related equipment
Technical Field
The invention relates to the field of tunnel inspection and maintenance, in particular to a single-stage lightweight subway lining cavity recognition method system, an intelligent terminal and a computer readable storage medium.
Background
Periodic inspection and maintenance of subway tunnel liners is critical to ensuring safe and reliable operation of underground subway systems because tunnel liners are constructed of durable materials, but can degrade over time due to various factors such as wear, water intrusion, and environmental conditions, and early discovery of potential problems is critical to preventing service interruption, potential safety hazards, or catastrophic events; however, conventional visual and lidar detection methods can only inspect the lining surface, and a penetrable detection method is required to detect defects such as cracks, voids and delamination inside the lining.
The formation of the tunnel lining cavity is mainly caused by insufficient grouting between lining layers. These voids are typically thin and elongated in the vertical and horizontal dimensions, respectively. In addition, some holes in the lined rock are round, and are larger and deeper due to water and soil erosion. When a cavity appears in the tunnel lining, the dielectric constant difference between the air in the cavity and the surrounding concrete generates strong reflection, and the reflection is shown as high amplitude in the radar chart.
The ground penetrating radar (Ground Penetrating Radar, GPR for short) has the advantages of strong penetrating power, high resolution, high efficiency, non-contact and the like, and is a nondestructive testing method widely applied to tunnel lining detection. However, due to the reliance on expertise and experience, interpretation of tunnel void defect data, i.e., GPR data, is challenging, resulting in differences in interpretation results, thereby affecting the accuracy and consistency of tunnel lining void detection. In addition, the infrastructure, reinforcement materials, and non-uniform lining materials within the tunnel can create complex noise that can interfere with and distort the response of the ground penetrating radar, resulting in human prejudice and misunderstanding. Currently, the deep learning framework includes two types of GPR interpretation: a two-stage method comprising identification and classification, one being a single-stage method. The single-stage method is lightweight, is more suitable for analyzing a large amount of data using a portable device, facilitates interpretation of GPR data, and, among the single-stage methods, you Only Look Once (YOLO detection model, YOLO for short) method is more effective in interpreting GPR data.
However, the single-stage method in the prior art cannot provide accuracy equivalent to that of the two-stage method, and when the tunnel cavity defect data is detected, the disadvantage still exists in using the single-stage method, so that the detection result is inaccurate.
Accordingly, there is a need for improvement and development in the art.
Disclosure of Invention
The invention mainly aims to provide a single-stage lightweight subway lining cavity recognition method system, an intelligent terminal and a computer readable storage medium, and aims to solve the problem that in the prior art, when tunnel cavity disease data are detected, the defect still exists when a single-stage method is used, and the detection result is inaccurate.
In order to achieve the above object, a first aspect of the present invention provides a single-stage lightweight subway lining cavity recognition method, wherein the single-stage lightweight subway lining cavity recognition method includes:
tunnel hole defect data are obtained, and tunnel hole pretreatment is carried out on the tunnel hole defect data to obtain pretreatment data, wherein the tunnel defect data at least comprise tunnel hole position information and tunnel hole size information;
inputting the preprocessed data into a trained optimized detection model to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by antagonizing network optimization through real data and ground penetrating radar depth convolution to train.
Optionally, the obtaining tunnel hole defect data, performing tunnel hole pretreatment on the tunnel hole defect data to obtain pretreated data, specifically includes:
acquiring tunnel cavity defect data through a radar;
and carrying out tunnel cavity preprocessing on the tunnel cavity defect data to obtain preprocessed data, wherein the preprocessing comprises at least one of DC offset elimination, gain adjustment, time zero correction, band-pass filtering and moving average operation.
Optionally, the generating step of the optimized detection model includes:
based on the YOLOv7 detection model, the optimized detection model is obtained by optimizing the feature extraction and fusion process and optimizing the loss function.
Optionally, the optimizing the feature extraction and fusion process specifically includes:
and optimizing the feature extraction and the fusion process through a pure convolution module, a self-attention and convolution fusion module and an optimized high-efficiency aggregation network.
Optionally, the optimizing the loss function specifically includes:
and optimizing the loss function through a weighted cross comparison function.
Optionally, the optimizing detection model generates the synthetic data obtained by optimizing the countermeasure network through the real data and the depth convolution of the ground penetrating radar to train, which specifically includes:
Generating preliminary synthesized data according to a time domain finite difference method;
and generating the preliminary synthetic data for optimizing the countermeasure network according to the depth convolution of the ground penetrating radar after training, and obtaining the synthetic data.
Optionally, the ground penetrating radar deep convolution generating countermeasure network includes a generator network and a discriminator network, and training is performed on the ground penetrating radar deep convolution generating countermeasure network through the generator network and the discriminator network to obtain the trained ground penetrating radar deep convolution generating countermeasure network.
The second aspect of the present invention provides a single-stage lightweight subway lining cavity recognition system, wherein the single-stage lightweight subway lining cavity recognition system includes:
the data preprocessing module is used for acquiring tunnel hole defect data and carrying out tunnel hole preprocessing on the tunnel hole defect data to obtain preprocessed data, wherein the tunnel defect data at least comprises tunnel hole position information and tunnel hole size information;
the data processing module is used for inputting the preprocessed data into the optimized detection model after training to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by optimizing an countermeasure network through the deep convolution of real data and the ground penetrating radar to train.
A third aspect of the present invention provides an intelligent terminal, where the intelligent terminal includes a memory, a processor, and a single-stage lightweight subway lining cavity recognition program stored in the memory and operable on the processor, and the single-stage lightweight subway lining cavity recognition program, when executed by the processor, implements the steps of any one of the single-stage lightweight subway lining cavity recognition methods.
A fourth aspect of the present invention provides a computer-readable storage medium, on which a single-stage lightweight subway lining cavity recognition program is stored, the single-stage lightweight subway lining cavity recognition program implementing the steps of any one of the single-stage lightweight subway lining cavity recognition methods when executed by a processor.
From the above, the tunnel void defect data is obtained in the scheme of the invention, and the tunnel void defect data is subjected to tunnel void pretreatment to obtain pretreatment data, wherein the tunnel defect data at least comprises tunnel void position information and tunnel void size information; inputting the preprocessed data into a trained optimized detection model to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by antagonizing network optimization through real data and ground penetrating radar depth convolution to train.
Compared with the prior art, the method for identifying the single-stage lightweight subway lining cavity, which is provided by the application, aims at solving the problem that the detection result is inaccurate due to the fact that the single-stage method still has the defects when the tunnel cavity disease data are detected, and processes the tunnel cavity disease data by using an optimized detection model obtained by optimizing the feature extraction and fusion process and the loss function of a Yolov7 (one of the Yolov detection models, abbreviated as Yolov 7) detection model, so that the process of extracting local features, global features and fusion features of the optimized detection model can be enhanced, the calculation cost can be reduced without affecting the performance of the model, the accuracy of the model is improved, and meanwhile, the robustness of the model is improved by optimizing the loss function; in the method, the training data adopted in the process of training the optimized detection model of the model are synthesized data obtained by optimizing an antagonism network through real data and depth convolution of the ground penetrating radar, so that the generated synthesized data are more formal, and the number of training data sets is greatly increased due to the fact that the optimized detection model is trained through combination of the synthesized data and the real data, so that the training effect of the optimized detection model is improved, the more accurate optimized detection model can be obtained, tunnel cavity disease data obtained by the ground penetrating radar can be processed in a short time through the obtained optimized detection model, an effective and accurate detection result can be obtained, and a situation that a cavity appears in a tunnel lining can be found through the detection result, so that the corresponding processing can be performed, accidents caused by the cavity in the tunnel can be prevented, and the inside of the tunnel is safer.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments or the description of the prior art will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic flow chart of a single-stage lightweight subway lining cavity recognition method provided by an embodiment of the invention;
fig. 2 is a schematic diagram of a network architecture of YOLOv7 according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a specific structure of a pure convolution module according to an embodiment of the present invention;
FIG. 4 is a schematic structural diagram of an optimized detection model according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of the structure of a self-attention and convolution fusion module provided by an embodiment of the present invention;
FIG. 6 is a schematic diagram of a parallel rolling and optimized efficient aggregation network according to an embodiment of the present invention;
FIG. 7 is a schematic diagram of a simulated void using three different shapes provided by an embodiment of the present invention;
FIG. 8 is a schematic diagram of preliminary synthetic data provided by an embodiment of the present invention;
FIG. 9 is a schematic diagram of a process for generating a countermeasure network training by deep convolution of a ground penetrating radar according to an embodiment of the present invention;
FIG. 10 is a schematic diagram of a data acquisition process according to an embodiment of the present invention;
FIG. 11 is a schematic view of a tunnel structure according to an embodiment of the present invention;
fig. 12 is a schematic diagram of a composition module of a single-stage lightweight subway lining cavity recognition system according to an embodiment of the present invention;
fig. 13 is a schematic block diagram of an internal structure of an intelligent terminal according to an embodiment of the present invention.
Detailed Description
In the following description, for purposes of explanation and not limitation, specific details are set forth such as the particular system architecture, techniques, etc., in order to provide a thorough understanding of the embodiments of the present invention. It will be apparent, however, to one skilled in the art that the present invention may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present invention with unnecessary detail.
It should be understood that the terms "comprises" and/or "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.
It is also to be understood that the terminology used in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
It should be further understood that the term "and/or" as used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes such combinations.
As used in this specification and the appended claims, the term "if" may be interpreted in the context of "when …" or "once" or "in response to a determination" or "in response to a classification. Similarly, the phrase "if determined" or "if classified to [ described condition or event ]" may be interpreted in the context of meaning "upon determination" or "in response to determination" or "upon classification to [ described condition or event ]" or "in response to classification to [ described condition or event ]".
The following description of the embodiments of the present invention will be made more fully hereinafter with reference to the accompanying drawings, in which embodiments of the invention are shown, it being evident that the embodiments described are only some, but not all embodiments of the 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.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, but the present invention may be practiced in other ways other than those described herein, and persons skilled in the art will readily appreciate that the present invention is not limited to the specific embodiments disclosed below.
Periodic inspection and maintenance of subway tunnel liners is critical to ensuring safe and reliable operation of underground subway systems because tunnel liners are constructed of durable materials, but can degrade over time due to various factors such as wear, water intrusion, and environmental conditions, and early discovery of potential problems is critical to preventing service interruption, potential safety hazards, or catastrophic events; however, conventional visual and lidar detection methods can only inspect the lining surface, and a penetrable detection method is required to detect defects such as cracks, voids and delamination inside the lining.
The ground penetrating radar has the advantages of strong penetrating power, high resolution, high efficiency, non-contact and the like, and is a nondestructive testing method widely applied to tunnel lining detection. Ground penetrating radar exploration relies on the propagation and reflection of electromagnetic waves in tunnel liners. The ground penetrating radar transmitting antenna transmits electromagnetic waves with a central frequency to a target stratum through the antenna, and the electromagnetic waves propagate in a matrix material. When a surface with different electromagnetic properties is encountered, the wave is reflected or refracted back and recorded by the receiver. The reflection amplitude is determined by the contrast of the dielectric constants, whereas the conductivity attenuates the GPR signal as it penetrates downwards, the superimposed signal building a radar map.
The formation of the tunnel lining cavity is mainly caused by insufficient grouting between lining layers. These voids are typically thin and elongated in the vertical and horizontal dimensions, respectively. In addition, some holes in the lined rock are round, and are larger and deeper due to water and soil erosion. When a cavity appears in the tunnel lining, the dielectric constant difference between the air in the cavity and the surrounding concrete generates strong reflection, and the reflection is shown as high amplitude in the radar chart.
In the interpretation process of GPR data, due to dependence on expertise and experience, and under the condition that holes in tunnel lining are distorted and disturbed by steel bars above to generate clutter, interpretation of GPR data is challenging, and interpretation results obtained by different methods are different, so that the accuracy and consistency of tunnel lining hole detection are affected. In addition, the infrastructure, reinforcement materials, and non-uniform lining materials within the tunnel can create complex noise that can interfere with and distort the response of the ground penetrating radar, resulting in human prejudice and misunderstanding. As subway tunnel length increases, manual interpretation of GPR data becomes increasingly time consuming, requiring development of a fast, stable automated GPR interpretation method.
Currently, GPR interpretation based on deep learning has recently become an effective solution to this problem. The deep learning framework has been divided into two types of GPR interpretation: a two-stage method comprising identification and classification, and a single-stage method. There are several methods currently used to facilitate automatic tunnel lining GPR interpretation using a two-stage approach, including tunnel lining defect segmentation using convolutional neural networks, tunnel lining element identification using mask-area based neural networks, and the like. However, these methods are computationally demanding, which presents challenges for operating the server. In contrast, single-stage methods are lightweight and are more suitable for analyzing large amounts of data using portable devices. Among popular single-stage methods such as RetinaNet (retina network), detectanet (detection network), and SqueezeNet (compression network), the You Only Look Once method has proven to be more effective in interpreting GPR data.
However, the single-stage method cannot provide accuracy comparable to that of the two-stage method at present, and thus it is required to improve the accuracy of the single-stage method while maintaining the computational efficiency thereof. Moreover, the target detection of GPR radar maps is more challenging compared to optical images. This is because the superimposed and diffracted subsurface signals make the target object difficult to identify, resulting in unclear boundaries and features. In addition, the low contrast between the different materials in the radar map, coupled with the subtle differences in dielectric properties of the subsurface materials, further complicates the distinction between objects with similar properties, making the radar map noisy and ambiguous. Therefore, it is necessary to improve the feature extraction capability of the single-stage method to improve the accuracy thereof.
In order to solve at least one of the problems, the scheme of the invention acquires tunnel hole defect data, and performs tunnel hole pretreatment on the tunnel hole defect data to obtain pretreated data, wherein the tunnel defect data at least comprises tunnel hole position information and tunnel hole size information; inputting the preprocessed data into a trained optimized detection model to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by antagonizing network optimization through real data and ground penetrating radar depth convolution to train.
Compared with the prior art, the method for identifying the single-stage lightweight subway lining cavity, which is provided by the application, aims at solving the problem that the detection result is inaccurate due to the fact that the single-stage method still has the defects when the tunnel cavity disease data are detected, and processes the tunnel cavity disease data by using an optimized detection model obtained by optimizing the feature extraction and fusion process and the loss function of a Yolov7 (one of the Yolov detection models, abbreviated as Yolov 7) detection model, so that the process of extracting local features, global features and fusion features of the optimized detection model can be enhanced, the calculation cost can be reduced without affecting the performance of the model, the accuracy of the model is improved, and meanwhile, the robustness of the model is improved by optimizing the loss function; in the method, the training data adopted in the process of training the optimized detection model of the model are synthesized data obtained by optimizing an antagonism network through real data and depth convolution of the ground penetrating radar, so that the generated synthesized data are more formal, and the number of training data sets is greatly increased due to the fact that the optimized detection model is trained through combination of the synthesized data and the real data, so that the training effect of the optimized detection model is improved, the more accurate optimized detection model can be obtained, tunnel cavity disease data obtained by the ground penetrating radar can be processed in a short time through the obtained optimized detection model, an effective and accurate detection result can be obtained, and a situation that a cavity appears in a tunnel lining can be found through the detection result, so that the corresponding processing can be performed, accidents caused by the cavity in the tunnel can be prevented, and the inside of the tunnel is safer.
Exemplary method
As shown in fig. 1, the embodiment of the invention provides a single-stage lightweight subway lining cavity recognition method, specifically, the single-stage lightweight subway lining cavity recognition method includes the following steps:
and S100, acquiring tunnel hole defect data, and performing tunnel hole pretreatment on the tunnel hole defect data to obtain pretreated data, wherein the tunnel defect data at least comprises tunnel hole position information and tunnel hole size information.
Specifically, the tunnel cavity defect data is a radar map obtained by a ground penetrating radar, wherein the vertical axis of the radar map represents the detection depth, the position and depth of the underground cavity can be determined by comparing images with different depths, and the shape and size of the cavity can be estimated by observing the signal intensity change in the radar map, so that the tunnel defect data comprises tunnel cavity position information and tunnel cavity size information and tunnel cavity shape information. When the model is adopted to process data, corresponding pre-tunnel hole processing is firstly carried out on the data acquired by the model.
Further, the method for obtaining tunnel void defect data includes performing tunnel void pretreatment on the tunnel void defect data to obtain pretreated data, and specifically includes:
Acquiring tunnel cavity defect data through a radar;
and carrying out tunnel cavity preprocessing on the tunnel cavity defect data to obtain preprocessed data, wherein the preprocessing comprises at least one of DC offset elimination, gain adjustment, time zero correction, band-pass filtering and moving average operation.
Specifically, the tunnel cavity defect data is obtained through a ground penetrating radar, an electromagnetic wave with a central frequency is emitted to a target subway tunnel through an antenna by a ground penetrating radar transmitting antenna, the electromagnetic wave propagates in a matrix material, when the electromagnetic wave encounters surfaces with different electromagnetic characteristics, the electromagnetic wave is reflected or refracted back and recorded by a receiver, the reflection amplitude is determined by the contrast of dielectric constants, the electric conductivity attenuates GPR signals when penetrating downwards, and a radar map is constructed by the superimposed signals, so that the corresponding tunnel cavity defect data is obtained. In one embodiment of the present application, the frequencies of the above ground penetrating radar transmitting antennas are preferably selected from 500MHz and 800 MHz.
After the tunnel cavity defect data is obtained, correspondingly performing tunnel cavity pretreatment on the tunnel cavity defect data to obtain the pretreated data, wherein the tunnel cavity pretreatment comprises DC offset elimination, gain adjustment, time zero correction, band-pass filtering and moving average operation so as to obtain the pretreated data for tunnel inspection.
The D-C drift elimination specifically comprises the following steps: D-C drift refers to the dc offset in the signal, i.e., the offset of the signal at zero frequency; D-C drift cancellation is a method of removing dc offset in a signal, making the value of the signal zero at zero frequency, thus making the signal more stable and accurate.
The gain adjustment includes: the gain adjustment is to adjust the amplitude of the signal to adapt to the specific requirement or range; by gain adjustment, the amplitude of the signal can be amplified or reduced for better processing of the signal in subsequent processing steps.
The time zero correction includes: time zero correction is to eliminate time delays or time offsets in the signal; by correcting the time zero, the starting point of the signal can be aligned or corrected to a desired point in time to ensure accuracy and consistency of the signal over time.
The band-pass filtering includes: bandpass filtering is a signal processing technique that filters out signals in other frequency ranges by selecting signals in a specific frequency range (bandpass); bandpass filtering may be used to remove noise or unwanted frequency components, preserving the frequency components of interest.
The moving average includes: moving average is a method of smoothing a signal to reduce noise and irregular fluctuations in the signal by performing an averaging process on the signal; the moving average typically uses a sliding window, and the average of the signal is calculated within each window and taken as a smoothed signal value.
During preprocessing, a user selects one or more of the specific application scenes, the characteristics of data and the required data quality in advance to perform tunnel cavity preprocessing. After tunnel cavity preprocessing is carried out on tunnel cavity defect data, corresponding preprocessed data are obtained, and the obtained preprocessed data can obtain more accurate detection results when the preprocessed data are processed by the optimized detection model after DC offset elimination, gain adjustment, time zero correction, band-pass filtering and moving average operation.
Step S200, inputting the preprocessed data into a trained optimized detection model to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by antagonizing network optimization through real data and ground penetrating radar depth convolution to train.
Specifically, after the preprocessing data are obtained, the preprocessing data are input into the optimized detection model after training is completed, and then the picture with the hole detection result can be correspondingly output, wherein the detection result is correspondingly represented in the picture with the hole detection result through the identification frame and the confidence coefficient, and the picture with the hole detection result can know which part of the tunnel has the hole, so that the corresponding part can be quickly repaired, and the use safety of the tunnel is improved.
Further, the optimizing detection model specifically includes:
based on the YOLOv7 detection model, the optimized detection model is obtained by optimizing the feature extraction and fusion process and optimizing the loss function.
Specifically, the optimized detection model in the embodiment of the present application is obtained by improving a YOLOv7 (one of the YOLOv detection models, abbreviated as YOLOv 7) detection model. Compared with the previous version of the detection model of the YOLO series, the YOLO v7 adopts an anchor point-based method, and uses Silu (sigmoid activation function weighted linear combination) as an activation function in a convolution layer module. In addition, it introduces an Effective ELAN (Effective Long-range Aggregation Network, ELAN) module and a REP (Re-parameterization, REP) module into the network architecture. And the ELAN module achieves efficient learning and convergence in deeper networks by controlling the shortest and longest gradient paths. The network architecture of YOLOv7 comprises four parts, namely an input, a backbone network, an intermediate layer and an output layer, as shown in particular in fig. 2.
As shown in fig. 2, in the network architecture of YOLOv7, the backbone network is responsible for feature extraction, and includes a plurality of CNN (Convolutional Neural Networks, convolutional neural network, abbreviated as CNN) base modules CBS (Compact Bilinear Pooling, compact bilinear Pooling, abbreviated as CBS), ELAN modules, and Max Pooling (MP) modules, where the ratio of the channel numbers of MP1 and MP2 is different, so MP1 and MP2 are used for distinguishing. The CBS module uses conventional convolution, batch generation features, as opposed to the convolution layer in a conventional CNN network. The ELAN module enhances the stability and generalization ability of the quantity normalization and SiLU activation function model, while the MP module is used for the downsampling operation. The middle layer, i.e. the neck, is formed by a plurality of modules, including a spatial pyramid pooling and SPPCSPC (patial Pyramid Pooling based on Convolutional and Spatial Pooling Convolutional, spatial pyramid pooling based on convolution and spatial pooling convolution, abbreviated as SPPCSPC) module, a CBS module, an MP2 module, and an UP (UP) module, as a feature fusion part. The main purpose of the method is to fuse different scale characteristics from a backbone network and better detect targets with different sizes. The SPPCSPC module can extract features of different sensing fields, and the CBS module, the MP2 module and the UP module are used for feature fusion and compression, so that parameters and computational complexity are reduced. By means of feature fusion of the middle layer, the YOLOv7 can detect targets with different sizes more accurately, and therefore detection accuracy and efficiency are improved. The output layer, namely a head, outputs the prediction category probability, the confidence score and the boundary frame coordinates of the target by using the three detection heads. Each detection head corresponds to a different feature map scale for detecting objects of different sizes. Each feature map cell predicts the probability and location of an object and outputs a confidence score for the predicted bounding box, indicating whether the object is present in the bounding box. During the training process, each feature map cell is assigned to detect one or more targets, the feature map cell responsible for a particular target predicting the class and location of the target. In YOLOv7, matching of positive and negative samples is based on calculating the intersection ratio (Intersection over Union, intersection ratio, IOU for short) between the target frame and the anchor frame. Specifically, the model calculates the IOU for each target frame and all anchor frames and assigns it to the anchor frame with the highest IOU. If the IOU value of an anchor block and a plurality of target blocks is less than a certain threshold, then it is considered a negative sample. Conversely, if the IOU value of an anchor block and target block is greater than a certain threshold, then it is considered a positive sample. Finally, if the IOU value of an anchor block and target block is between the two thresholds, it is considered to ignore the samples and is not included in the calculation of the penalty function. The YOLOv7 detection model is trained in this way.
The optimized detection model can be obtained after the YOLOv7 detection model is optimized.
Further, the optimizing the feature extraction and fusion process specifically includes:
and optimizing the feature extraction and the fusion process through a pure convolution module, a self-attention and convolution fusion module and an optimized high-efficiency aggregation network.
The specific structure of the pure convolution module (CNXB) is shown in fig. 3, where the pure convolution module includes a pure convolution Block (ConvNeXt Block). The original ELAN module is improved through ConvNeXt Block. The convolution of 4 3×3 in the ELAN modules is replaced by the convnex Block module and facilitated by the convnex Block module, thereby replacing the first ELAN module in the backbone network with one CNXB module and one CBS module.
ConvNeXt Block first extracts and normalizes features through a 7 x 7 deep convolution layer and normalization layer; then, the channel size is expanded four times and nonlinearity is introduced by a 1×1 convolution and a GELU (Gaussian Error Linear Unit, an activation function, abbreviated as GELU) activation function; then, restoring the channel size to the original size through 1×1 convolution, layer scaling and Drop Path (regularization) layers, and finally adding the input and output together; wherein, in ConvNeXt Block, after only one normalized layer is located in the depth convolution, it uses layer normalization (Layer Normalization, LN for short) as normalization. Furthermore, convNeXt Block reduces the requirements for activation and regularization functions because each module employs only one activation function and one regularization function.
CNXB addition location as shown in fig. 4, a feature map of size 160×160×128 is obtained after passing through 4 CBS modules in the backbone network, and the CXNB modules are placed at the end of the CBS module sequence as a special block, with an additional 3×3 convolutional CBS added, in order to capture more feature information and provide more useful information for subsequent tasks. This allows for better processing of the input features and provides a higher quality representation of the features at the output level. Furthermore, the higher computational efficiency of the CXNB module effectively speeds up model training and inference, making it more viable in practical applications.
A fusion module of self-attention and convolution (Attention and Convolution Mix, fusion of self-attention and convolution, abbreviated ACmix), in this embodiment of the application, ACmix combines a self-attention and convolution paradigm, with the advantages of self-attention and convolution, while having less computational overhead than pure convolution or self-attention. The structure of the ACmix module is shown in fig. 5, and the ACmix structure is divided into two stages: stage I and stage II. In stage I, the input features are convolved projected and reshaped into N blocks, respectively, to obtain a containing texel An intermediate feature set of the feature map. In phase II, the structure is split into two paths: self-attention and convolution. In the self-attention path, the intermediate feature sets are clustered into N groups, each group containing 3 feature maps. These 3 feature maps are used as queries, keys and values, respectively, and are processed according to a conventional multi-headed self-care model. In the convolution path, N groups are generated using a lightweight fully-connected layer, each group having K 2 And (3) a characteristic diagram. For the generated feature maps, they are processed by translation and aggregation, convolutionally processing the input features and collecting information from the local perceptual domain. Finally, the outputs of the self-attention and convolution paths are summed.
In ACmix, the standard convolution process is:
wherein K is p,q Core weights representing core positions (p, q); g ij And f ij Output and input feature tensors corresponding to pixel (i, j), respectively; k is the kernel size;is an output characteristic of the core position (p, q); />Is the output characteristic of the kernel position (p, q) after the conversion operator; />Representation of matrix->Translation is carried out, horizontal movement->Units, move in vertical direction +.>p and q are the original row and column indices, < >> K is divided by 2, where k is a constant and is rounded down by brackets.
The multi-head attention mechanism process is as follows:
wherein,projection matrices of queries, keys and values, respectively; />Representing a local pixel region of spatial extent kc centered on (i, j); />The superscript (l) of (i) represents the number of layers or levels, which means that the Query (Query) vector of position (i, j) at level l, +.>Key (Key) vector representing position (a, b) at the first layer，/>A Value (Value) vector representing the position (a, b) at the first layer. />Is->The corresponding attention weights of the internal features.
The convolution portion and the self-attention portion are fused, i.e. the outputs of the self-attention and convolution paths are summed, using different weight fusion:
F out ＝αF att +βF conv ，(6)；
F out 、F att 、F conv the output features of ACmix, self-attention, and convolution, respectively; alpha and beta are learnable parameters.
In YOLOv7, ELAN modules and SPPCSPC modules located at the backbone and neck of the network enable the network to concentrate on important areas that contribute to object detection, thereby improving detection accuracy. In order to enhance the representation capability of the optimized detection model on the cavity characteristics influenced by the steel bars, three ACmix modules are introduced on the basis of Yolov7, as shown in fig. 4. These ACmix modules are placed after the ELAN and SPPCSPC modules, allowing them to run on the feature maps generated by these modules, in particular in the backbone network, one ACmix module is placed after both the second and third ELAN modules on the YOLOv7 basis, and one ACmix module is also placed after the SPPCSPC module in the neck. This arrangement helps to more effectively extract local and global features, thereby enhancing the ability of the network to identify and locate objects. In addition, since ELAN and SPPCSPC modules have focused attention of the network on a key area of object detection, ACmix module can precisely extract valuable feature information, thereby improving accuracy of object detection.
The optimized high-efficiency aggregation network (in the embodiment of the application, the optimized high-efficiency aggregation network is represented by PELAN) adopts a combination of parallel convolution (Parallel Convolution, parallel convolution, abbreviated as PConv) and 1x1CBS modules to replace the CBS modules in the original high-efficiency aggregation network on the basis of the high-efficiency aggregation network, so that a new PELAN module is created. The space features are effectively extracted by reducing redundant computation and memory access simultaneously through parallel convolution. The specific structure of the PELAN and PConv is shown in fig. 6, and the design of PConv makes use of redundancy in the feature diagram, and only performs conventional convolution on part of the input channels without affecting the rest of the channels. Wherein, the floating point operand (Floating Point Operations, floating point operand, FLPs) of PConv is calculated by the following formula:
PConv FLOPs FLOPs that are PConv; h, w, c p The height, width and channel number of the data, respectively; k is the number of rows or columns of 3D tensors in the filter. As can be seen from the floating point operand, whenWhen PConv FLOPs are only 1/16 of the conventional convolution, computing resources can be better utilized.
As shown in fig. 4, on the basis of YOLOv7, the first ELAN module is replaced with CXNB and CBS modules, and all other ELAN modules are replaced with PELANs; the PELAN can flexibly control the convolution operation range through the mask, so that the PELAN can process various scale inputs and simultaneously reduce the calculation cost. This results in an improvement in model generalization and a reduction in computational resource consumption.
It will be appreciated that in the YOLOv7 framework, the backbone network plays a critical role in feature extraction, which is essential for accurate object recognition in deep learning. In the backbone network, CXNB module and CBS module (convolution kernel is 3 and step length is 1) are added after the fourth CBS module, and ACmix module is added at three connection points between backbone and neck, so that the backbone of YOLOv7 is improved, and the effect of feature extraction is improved. In addition, in order to further optimize the model, in the ELAN module of the backbone and the neck, the CBS module (convolution kernel is 3 and step size is 1) is replaced by the PELAN module, so that the calculation complexity and the parameter number of the model are reduced. The improvement makes the meaningful content and the position in the input image sample of the optimized detection model more concentrated, and can extract the characteristic information more effectively even if the reinforcement bar interference exists, thereby improving the detection precision.
Further, the optimizing the loss function specifically includes:
and optimizing the loss function through a weighted cross comparison function.
In particular, unstable convergence of GPR interpretation based on deep learning occurs due to class imbalance and complex background caused by subsurface heterogeneity. Therefore, in the embodiment of the present application, the original loss function weighted intersection set in the YOLOv7 detection model is replaced by an optimized loss function, namely a weighted overlap-and-ratio function Wise-IoU (Weighted Intersection over Union, the weighted overlap-and-ratio function, abbreviated as WIoU), to solve these problems.
The WIoU adopts a dynamic non-monotonic focusing mechanism to balance the bounding box regression of samples with different qualities, thereby improving overall model performance. In the YOLO series model, cross entropy is typically used as a loss function, but in the presence of unbalanced positive and negative samples, it may be detrimental to model accuracy. The WIoU in the present application considers two key indexes, i.e., the cross-over ratio between the target frame and the real frame and the IoU between the predicted frame and the real frame. WIoU adjusts IoU between target and real frames to solve the sample imbalance problem while IoU between predicted and real frames is improved to enhance model performance. In the embodiment of the application, the problem of unbalance of positive and negative samples in the tunnel lining GPR image can be better solved by adopting the WIoU loss function, the model performance is improved, and a more accurate detection result is generated.
Specifically, WIoU loss is defined by the following formulas (8), (9), (10) and (11):
/>
wherein Wg and Hg represent the width and height of the minimum bounding box, respectively,is a monotonic focusing factor, beta is an outlier of a dynamic non-monotonic FM anchor frame, gamma is a non-monotonic focusing factor, alpha and delta are super-parameters, x gt And y gt The GT frame horizontal and vertical point coordinates are represented, x and y represent the predicted frame center point coordinates, and +. >Representing the losses loss of the three WIoU versions, (-) -respectively>Representing the weight of IoU for weighting the calculation +.>r is a scaling factor calculated from a set of constant parameters (β, δ, α) and in some loss functions or evaluation criteria, α and δ are used as super-parameters for adjusting the training or optimization process of the model. The specific values of these parameters are chosen empirically or experimentally in order to obtain better model performance.
Furthermore, before the tunnel cavity defect data is detected by adopting an optimized detection model, the optimized detection model is trained. The dataset used to train the optimized detection model consists of actual real data and synthetic radar map data, i.e. synthetic data. In one embodiment of the present application, the data set is divided into uniformly sized portions, with a predetermined pixel size of 800×320, prior to the optimized detection model training; the final data set consists of 400 parts, which are divided into two sets, a training set and a validation set, wherein the ratio of the training set to the validation set is 8:2. In addition, when the training set is generated, the traditional data enhancement technology is adopted, real data and synthesized data are processed through cutting, translation, brightness adjustment, rotation and mirroring, the quality and the quantity of data in the data set are further enhanced, a data set containing 2000 cavity disease data is finally obtained, in addition, the real data and the synthesized data are processed through fuzzy and scaling operation, and therefore radar diagrams generated by different antennas are simulated. These data enhancement techniques help to diversify the data set and enable better generalization of the model to unseen data. Wherein segments in the dataset containing various void geometries are annotated with lambelimg (graphical image annotation tool) to enable model learning to identify different void shapes and sizes. By using a plurality of diversified data with notes in the data set, the accuracy of the optimized detection model can be improved, and the optimized detection model with better effect can be obtained.
Further, the optimization detection model generates synthetic data obtained by optimizing an countermeasure network through the real data and the depth convolution of the ground penetrating radar to train, and specifically comprises the following steps:
generating preliminary synthesized data according to a time domain finite difference method;
and generating the preliminary synthetic data for optimizing the countermeasure network according to the depth convolution of the ground penetrating radar after training, and obtaining the synthetic data.
Because ground truth is not available, numerical modeling is a popular method of studying the response of ground penetrating radars in different fault situations. Time-Domain Finite Difference (FDTD) methods are commonly used to numerically model GPR propagation. The FDTD method discretizes the electromagnetic field on a spatial grid and uses a time stepping technique to simulate the propagation of electromagnetic waves in a medium. In each time step, maxwell's equations are discretized to calculate the changes in electric and magnetic fields in space. Thus, the FDTD method is computationally efficient for modeling problems for multiple sources or wide frequency ranges. GPRMAX is an open source software that simulates the propagation of electromagnetic waves in a user-defined model using the FDTD method. Preliminary synthetic data can thus be generated by a time-domain finite difference method as well as GPRMAX.
In one embodiment of the application, the lining is composed of two layers of reinforcing steel meshes according to the design standard of subway tunnel lining, and the thickness of the secondary lining is 0.3-0.4 m. In the FDTD model, therefore, the total height is 1.2m, the air space height is 0.2m, the secondary lining thickness is 0.4m, the primary lining thickness is 0.2m, and the remaining 0.4m is the surrounding rock layer. In order to further enhance the generalization capability of the model, FDTD is adopted to model the tunnel lining cavity, and the modeling is respectively performed on three scenes of concrete without reinforcing steel bars, one layer of reinforcing steel bars and two layers of reinforcing steel bars. The embedded voids vary in size, shape, and location as shown in fig. 7. The FDTD model designed in consideration of the requirements of the ground penetrating radar tunnel detection on the spatial resolution and penetration depth is shown in the following table 1.
TABLE 1 FDTD model parameters
As shown in fig. 7, where in fig. 7 e r Is the relative dielectric constant; σ is conductivity (siemens/meter); mu (mu) r Is relative permeability; sigma (sigma) r Is magnetic loss (ohm/meter); three different shapes are used to simulate voids, namely thin ovals, circles and rectangles. Oval voids are located in shallower subsurface representing layer-to-layer boundaries, while circular and rectangular voids are buried deeper, representing voids in the rock behind the lining, the material properties and structure of the lining model being in line with reality.
Preliminary synthetic data, i.e., a preliminary synthetic radar map, may thus be obtained by the FDTD method, and in one embodiment of the present application, as shown in fig. 8, is the obtained preliminary synthetic radar map. In the primary synthetic radar map, the circular cavity generates hyperbolic reflection, while the rectangular cavity generates cross reflection; wherein the reverberation disturbances of the reinforcement and the lining layers are suitably modeled, which lead to a distortion of the cavity response. However, in practice real voids exhibit irregular shapes, whereas regular shapes used in numerically modeling GPR propagation using the FDTD method are too idealized and therefore need to be optimized before they can be used as training samples.
And after the preliminary synthetic data is obtained, generating the preliminary synthetic data for optimizing an countermeasure network according to the depth convolution of the ground penetrating radar after training is completed, and obtaining the synthetic data.
Further, the ground penetrating radar deep convolution generating countermeasure network comprises a generator network and a discriminator network, and training is carried out on the ground penetrating radar deep convolution generating countermeasure network through the generator network and the discriminator network to obtain the trained ground penetrating radar deep convolution generating countermeasure network.
Specifically, the ground penetrating radar depth convolution generating countermeasure network (Ground penetrating radar-Deep Convolutional Generative Adversarial Networks, GPR-DCGAN for short) is essentially a depth convolution generating countermeasure network (Deep Convolutional Generative Adversarial Networks, depth convolution generating countermeasure network, DCGAN for short) for optimizing the generated preliminary synthesized data in the embodiments of the present application. Wherein the GPR-DCGAN architecture includes two deep convolutional neural networks: a generator and a discriminator. The generator in DCGAN consists of five transposed convolutional layers, mapping the low-dimensional noise vector to the high-dimensional image space. On the other hand, the discriminator consists of five conventional convolutional layers, producing a binary classification result indicating whether the radar pattern is authentic or counterfeit, and the countermeasure between the generator and the discriminator network causes the generator to gradually generate a more realistic radar pattern.
The above-generated preliminary synthetic data is required to eliminate the mismatch between the true and synthetic radar maps because of the presence of unrealistic contents. The untrained ground penetrating radar depth convolution generating countermeasure network cannot process the primary synthesized data well, so that the real radar graph and the primary synthesized data generated by the FDTD method are adopted to form the ground penetrating radar depth convolution generating countermeasure network training data, and the ground penetrating radar depth convolution generating countermeasure network training data is used for generating countermeasure network training for the ground penetrating radar depth convolution. During training, initially, the true radar map, i.e., the true data, is designated as "true", and the random noise, i.e., the preliminary synthetic data, is designated as "false" by the discriminator; the discriminator then extracts and learns the differences between the true and preliminary composite data, including error information including dispersion-induced blurring and dispersion-induced noise; the error information is then propagated back through back propagation to the generator, which incorporates the errors into the synthetic radar map to enhance its authenticity. At the same time, the generator and the discriminator are trained and updated together, the generator attempts to increase its output by introducing noise to meet the discriminator requirements, while the discriminator enhances its ability to correctly distinguish between true data and generated data. The actual and modified composite radar map is then input back to the discriminator for the next round of discrimination. This alternating training process continues until the discriminator designates the modified synthetic radar map as "true", which indicates that the synthetic radar map generated by the generator is sufficiently convincing that the discriminator cannot distinguish it from the true radar map. Wherein the specific ground penetrating radar deep convolution generation countermeasure network training process is shown in fig. 9.
In one embodiment of the present application, the learning rate balances convergence speed and stability by fine tuning in GPR-DCGAN, eventually set to 0.0002. The optimizer chosen is "Adam (Adaptive Moment Estimation, adam optimizer, adam for short)", the model is trained from scratch without loading any pre-trained weights. And the input and output image sizes are set to 800 x 320 pixels, training is stopped when the arbiter loss reaches a preset threshold, and training is stopped when the preset threshold is set to approach 0 in one embodiment, and the weight of the generator is used for optimizing the preliminary synthesized data to obtain synthesized data.
The countermeasure network is generated through the deep convolution of the ground penetrating radar after training, and the initial synthesized data is optimized by adopting the weight of the generator obtained through training, so that more real synthesized data is obtained.
And training the optimized detection model through the obtained synthesized data and the real data, thereby obtaining the trained optimized detection model.
In the application, an embodiment of the subway lining cavity recognition method adopting the single-stage lightweight is provided at the same time, the specific data acquisition is acquired by using the antenna frequency of 800MHz and 500MHz respectively, the acquisition area is a section of subway section, and the section is constructed by adopting a mining method. In this application, fig. 10 and 11 illustrate a data acquisition process and a tunnel structure. The track spacing of the radar antenna is 0.02m, the sampling frequency is 12000MHz, the time window is 63ns, the collecting line is a left-right arched part, the ground penetrating radar antenna is closely attached to the tunnel lining surface when collecting data, the radar antenna is pushed to move, and the radar system detects an underground object by sending a wireless pulse signal and recording a received echo signal. And after the tunnel cavity defect data collected by the ground penetrating radar is obtained, preprocessing the tunnel cavity defect data to obtain preprocessed data for tunnel inspection.
Then, in order to evaluate the influence of four different optimization methods performed on the YOLOv7 network in the single-stage lightweight subway lining cavity recognition method on the YOLOv7 network model and know the better effect achieved by optimizing the detection model, the embodiment of the application performs ablation experiments to process the obtained pretreatment data, wherein the method comprises five groups of experiments, and each group of experiments adds different modules for testing and compares with the original YOLOv7 model. The performance of the model is measured using the F1Score (F1 Score ), the number of parameters, the amount of floating point operations (Floating Point Operations, abbreviated as "flow"), the FPS (Frames Per Second, FPS represents the number of frames of images processed per second, used to measure the speed of reasoning of the model, with higher FPS values representing the model being able to reason faster), and the average accuracy at 0.5 threshold (mean Average Precision, average accuracy mean, mAP), etc. Wherein the yolov7 model with ACmix module is called yolov7+acmix, the model with pure convolution module backbone network is called yolov7+cxnb, the model with PELAN module is called yolov7+pelan, the model with WIoU loss function is called yolov7+wiou, and the optimized detection model is called CAPW-YOLO.
Ablation experiments were performed using a computing system equipped with a NVIDIA GeForce RTX A6000GPU of 48GB VRAM; the operating system used was Windows10, and Python version 3.8 was used as the software development framework. The training parameters for the models in the experiments are recorded in tables 2 and 3, respectively.
Parameters (parameters) Value of Parameters (parameters) Value of
Learning rate 0.001 Batch size 32
Weight decay 0.0005 Image size 800×320
Momentum of 0.937 Iteration cycle 200
TABLE 2 Experimental environment
Table 3 ablation experiments
As can be seen from table 3, after adding CNXB module to YOLOv7, the average accuracy at 0.5 threshold increased by 4.8%, and the F1 score increased by 3%; meanwhile, the floating point operand is obviously reduced by 60%, which indicates that the CNXB module not only can improve the accuracy of the optimized detection model, but also greatly reduces the floating point operand of the optimized detection model, thereby leading the optimized detection model to be lighter. After the ACmix module is added, the average precision under the 0.5 threshold value is improved by 1.4%, and the F1 score is increased by 1%; however, adding an ACmix module also results in an 11% increase in parameters and 18.9% increase in floating point operations, and while adding an ACmix module has a positive impact on the accuracy of the model, its implementation is associated with increased computational requirements and reduced efficiency. The PELAN module is added into the YOLOv7 model, so that the model performance is obviously improved, the average precision under the threshold value of 0.5 is improved by 2.7%, the F1 score is improved by 3%, and meanwhile, the complexity of the model is greatly reduced; specifically, the PELAN module successfully reduces the parameters by 23.9% and the floating point operand by 28.4%. These results indicate that the PELAN module can effectively improve the accuracy of the model while reducing the computational cost. As is also apparent from table 3 above, the optimized inspection model was able to detect small-scale defects with insignificant features that YOLOv7 failed to detect, and also showed significant improvement in inspecting defects with significant void features.
Compared with the prior art, the method for identifying the single-stage lightweight subway lining cavity provided by the application aims at solving the problem that the detection result is inaccurate due to the fact that the defect exists when the tunnel cavity disease data are detected at present, the tunnel cavity disease data are processed by using an optimized detection model obtained by optimizing the feature extraction and fusion process and the loss function of a YOLOv7 (one of the YOLOv detection models, namely YOLOv 7) detection model, so that the process of extracting local features, global features and fusion features of the optimized detection model can be enhanced, the calculation cost can be reduced without affecting the performance of the model, the accuracy of the model is improved, and meanwhile, the robustness of the model is improved by optimizing the loss function; in the method, the training data adopted in the training of the optimized detection model of the model are synthesized data obtained by the counter network optimization through real data and the depth convolution of the ground penetrating radar, so that the generated synthesized data are more formal, and the number of training data sets is greatly increased due to the fact that the optimized detection model is trained by adopting the combination of the synthesized data and the real data, so that the training effect of the optimized detection model is improved, a more accurate trained optimized detection model can be obtained, and the obtained optimized detection model can obtain effective and accurate detection results in a short time.
Exemplary apparatus
As shown in fig. 12, corresponding to the above-described single-stage lightweight subway lining void identification method, an embodiment of the present invention also provides a single-stage lightweight subway lining void identification system including:
the data preprocessing module 51 is configured to obtain tunnel hole defect data, and perform tunnel hole preprocessing on the tunnel hole defect data to obtain preprocessed data, where the tunnel defect data at least includes tunnel hole position information and tunnel hole size information;
the data processing module 52 is configured to input the preprocessed data into a trained optimized detection model, and obtain and output a picture with a hole detection result, where the optimized detection model generates synthetic data obtained by antagonizing network optimization through deep convolution of real data and a ground penetrating radar to train.
It should be noted that, the specific structure and implementation manner of the single-stage lightweight subway lining cavity recognition system and each module or unit thereof may refer to the corresponding description in the method embodiment, and will not be repeated herein.
The division method of each module of the single-stage lightweight subway lining cavity recognition system is not limited to a specific one.
Based on the above embodiment, the present invention also provides an intelligent terminal, and a functional block diagram thereof may be shown in fig. 13. The intelligent terminal comprises a processor, a memory, a network interface and a display screen which are connected through a system bus. The processor of the intelligent terminal is used for providing computing and control capabilities. The memory of the intelligent terminal comprises a nonvolatile storage medium and an internal memory. The nonvolatile storage medium stores an operating system and a single-stage lightweight subway lining cavity recognition program. The internal memory provides an environment for the operation of an operating system and a single-stage lightweight subway lining cavity recognition program in a nonvolatile storage medium. The network interface of the intelligent terminal is used for communicating with an external terminal through network connection. The single-stage light subway lining cavity recognition program is executed by a processor to realize the steps of any one of the single-stage light subway lining cavity recognition methods. The display screen of the intelligent terminal can be a liquid crystal display screen or an electronic ink display screen.
It will be appreciated by those skilled in the art that the schematic block diagram shown in fig. 13 is merely a block diagram of a portion of the structure associated with the present invention and is not limiting of the smart terminal to which the present invention is applied, and that a particular smart terminal may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.
In one embodiment, an intelligent terminal is provided, where the intelligent terminal includes a memory, a processor, and a single-stage lightweight subway lining cavity recognition program stored in the memory and capable of running on the processor, where the single-stage lightweight subway lining cavity recognition program is executed by the processor to implement the steps of any one of the single-stage lightweight subway lining cavity recognition methods provided in the embodiments of the present invention.
The embodiment of the invention also provides a computer readable storage medium, wherein the computer readable storage medium is stored with a single-stage lightweight subway lining cavity recognition program, and the single-stage lightweight subway lining cavity recognition program is executed by a processor to realize the steps of any one of the single-stage lightweight subway lining cavity recognition methods provided by the embodiment of the invention.
It should be understood that the sequence number of each step in the above embodiment does not mean the sequence of execution, and the execution sequence of each process should be determined by its function and internal logic, and should not be construed as limiting the implementation process of the embodiment of the present invention.
It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-described division of the functional units and modules is illustrated, and in practical application, the above-described functional distribution may be performed by different functional units and modules according to needs, i.e. the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-described functions. The functional units and modules in the embodiment may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit, where the integrated units may be implemented in a form of hardware or a form of a software functional unit. In addition, the specific names of the functional units and modules are only for distinguishing from each other, and are not used for limiting the protection scope of the present invention. The specific working process of the units and modules in the above device may refer to the corresponding process in the foregoing method embodiment, which is not described herein again.
In the foregoing embodiments, the descriptions of the embodiments are emphasized, and in part, not described or illustrated in any particular embodiment, reference is made to the related descriptions of other embodiments.
Those of ordinary skill in the art will appreciate that the elements and algorithm steps of the examples described in connection with the embodiments disclosed herein may be implemented as electronic hardware, or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
In the embodiments provided in the present invention, it should be understood that the disclosed system/terminal device and method may be implemented in other manners. For example, the system/terminal device embodiments described above are merely illustrative, e.g., the division of the modules or elements described above is merely a logical functional division, and may be implemented in other manners, e.g., multiple elements or components may be combined or integrated into another system, or some features may be omitted, or not performed.
The integrated modules/units described above, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer-readable storage medium. Based on such understanding, the present invention may implement all or part of the flow of the method of the above embodiment, or may be implemented by a computer program to instruct related hardware, where the computer program may be stored in a computer readable storage medium, and when the computer program is executed by a processor, the steps of each method embodiment may be implemented. The computer program comprises computer program code, and the computer program code can be in a source code form, an object code form, an executable file or some intermediate form and the like. The computer readable medium may include: any entity or device capable of carrying the computer program code described above, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, randomAccess Memory), an electrical carrier wave signal, a telecommunications signal, a software distribution medium, and so forth. The content of the computer readable storage medium can be appropriately increased or decreased according to the requirements of the legislation and the patent practice in the jurisdiction.
The above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art will understand that; the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions are not intended to depart from the spirit and scope of the various embodiments of the invention, which are also within the spirit and scope of the invention.

Claims (10)

1. The single-stage lightweight subway lining cavity recognition method is characterized by comprising the following steps of:
tunnel hole defect data are obtained, and tunnel hole pretreatment is carried out on the tunnel hole defect data to obtain pretreatment data, wherein the tunnel defect data at least comprise tunnel hole position information and tunnel hole size information;
inputting the preprocessed data into a trained optimized detection model to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by antagonizing network optimization through real data and ground penetrating radar depth convolution to train.
2. The single-stage lightweight subway lining cavity recognition method according to claim 1, wherein the obtaining tunnel cavity defect data and performing tunnel cavity preprocessing on the tunnel cavity defect data to obtain preprocessed data specifically comprises:
acquiring tunnel cavity defect data through a radar;
and carrying out tunnel cavity preprocessing on the tunnel cavity defect data to obtain preprocessed data, wherein the preprocessing comprises at least one of DC offset elimination, gain adjustment, time zero correction, band-pass filtering and moving average operation.
3. The single-stage lightweight subway lining cavity recognition method according to claim 1, wherein the generating step of the optimization detection model comprises:
based on the YOLOv7 detection model, the optimized detection model is obtained by optimizing the feature extraction and fusion process and optimizing the loss function.
4. The single-stage lightweight subway lining cavity recognition method according to claim 3, wherein the optimizing the feature extraction and fusion process specifically comprises:
and optimizing the feature extraction and fusion process through a pure convolution module, a self-attention and convolution fusion module and an optimized high-efficiency aggregation network.
5. The method for identifying the single-stage lightweight subway lining cavity according to claim 3, wherein the optimizing the loss function specifically comprises:
and optimizing the loss function through a weighted cross-over function.
6. The single-stage lightweight subway lining cavity recognition method according to claim 1, wherein the optimization detection model trains synthetic data obtained by optimizing an countermeasure network through real data and depth convolution of a ground penetrating radar, and specifically comprises the following steps:
generating preliminary synthesized data according to a time domain finite difference method;
generating the preliminary synthetic data for optimizing the countermeasure network according to the depth convolution of the ground penetrating radar after training, and obtaining the synthetic data.
7. The method for identifying the single-stage lightweight subway lining cavity according to claim 6, wherein the ground penetrating radar depth convolution generating countermeasure network comprises a generator network and a discriminator network, and training is carried out on the ground penetrating radar depth convolution generating countermeasure network through the generator network and the discriminator network to obtain the trained ground penetrating radar depth convolution generating countermeasure network.
8. The utility model provides a single-stage lightweight subway lining cavity identification system which characterized in that, single-stage lightweight subway lining cavity identification system includes:
The data preprocessing module is used for acquiring tunnel hole defect data and carrying out tunnel hole preprocessing on the tunnel hole defect data to obtain preprocessed data, wherein the tunnel defect data at least comprises tunnel hole position information and tunnel hole size information;
the data processing module is used for inputting the preprocessed data into the optimized detection model after training to obtain and output a picture with a cavity detection result, wherein the optimized detection model generates synthetic data obtained by optimizing an countermeasure network through the deep convolution of real data and the ground penetrating radar to train.
9. An intelligent terminal, characterized in that the intelligent terminal comprises a memory, a processor and a single-stage lightweight subway lining cavity recognition program which is stored on the memory and can run on the processor, and the single-stage lightweight subway lining cavity recognition program realizes the steps of the single-stage lightweight subway lining cavity recognition method according to any one of claims 1-7 when being executed by the processor.
10. A computer-readable storage medium, wherein a single-stage lightweight subway lining void identification program is stored on the computer-readable storage medium, and the single-stage lightweight subway lining void identification program, when executed by a processor, implements the steps of the single-stage lightweight subway lining void identification method according to any one of claims 1 to 7.
CN202311102838.7A 2023-08-29 2023-08-29 Single-stage lightweight subway lining cavity recognition method and related equipment Pending CN117351321A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311102838.7A CN117351321A (en) 2023-08-29 2023-08-29 Single-stage lightweight subway lining cavity recognition method and related equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311102838.7A CN117351321A (en) 2023-08-29 2023-08-29 Single-stage lightweight subway lining cavity recognition method and related equipment

Publications (1)

Publication Number Publication Date
CN117351321A true CN117351321A (en) 2024-01-05

Family

ID=89356420

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311102838.7A Pending CN117351321A (en) 2023-08-29 2023-08-29 Single-stage lightweight subway lining cavity recognition method and related equipment

Country Status (1)

Country Link
CN (1) CN117351321A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117540178A (en) * 2024-01-09 2024-02-09 武汉大学 Tunnel lining internal cavity defect evaluation method and system

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117540178A (en) * 2024-01-09 2024-02-09 武汉大学 Tunnel lining internal cavity defect evaluation method and system
CN117540178B (en) * 2024-01-09 2024-03-29 武汉大学 Tunnel lining internal cavity defect evaluation method and system

Similar Documents

Publication Publication Date Title
Travassos et al. Artificial neural networks and machine learning techniques applied to ground penetrating radar: A review
Liu et al. GPR-based detection of internal cracks in asphalt pavement: A combination method of DeepAugment data and object detection
CN110287932B (en) Road blocking information extraction method based on deep learning image semantic segmentation
Zhu et al. Intelligent logging lithological interpretation with convolution neural networks
KR101811270B1 (en) Method and system for checking goods
CN108550259A (en) Congestion in road judgment method, terminal device and computer readable storage medium
CN107016677A (en) A kind of cloud atlas dividing method based on FCN and CNN
CN102938071A (en) Fuzzy clustering analysis method for detecting synthetic aperture radar (SAR) image changes based on non-local means
CN114595732B (en) Radar radiation source sorting method based on depth clustering
CN110929621B (en) Road extraction method based on topology information refinement
Su et al. End-to-end deep learning model for underground utilities localization using GPR
CN117351321A (en) Single-stage lightweight subway lining cavity recognition method and related equipment
CN102930519A (en) Method for generating synthetic aperture radar (SAR) image change detection difference images based on non-local means
Ma et al. Automatic detection of steel rebar in bridge decks from ground penetrating radar data
CN104850867A (en) Object identification method based on intuitive fuzzy c-means clustering
CN110706208A (en) Infrared dim target detection method based on tensor mean square minimum error
CN110298410A (en) Weak target detection method and device in soft image based on deep learning
Liang et al. Hybrid support vector machine optimization model for inversion of tunnel transient electromagnetic method
Zhu et al. Evaluating the optimised YOLO-based defect detection method for subsurface diagnosis with ground penetrating radar
Xiong et al. GPR-GAN: A ground-penetrating radar data generative adversarial network
Hu et al. Defects identification and location of underground space for ground penetrating radar based on deep learning
Liu et al. Advances in automatic identification of road subsurface distress using ground penetrating radar: State of the art and future trends
Dai et al. A new B-scan interpretation model in complex environments
CN103824279A (en) Image segmentation method based on organizational evolutionary cluster algorithm
Guo et al. Research on tunnel lining image target recognition method based on YOLOv3

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