CN117576487A - Intelligent ground penetrating radar cavity target identification method based on deformable convolution - Google Patents
Intelligent ground penetrating radar cavity target identification method based on deformable convolution Download PDFInfo
- Publication number
- CN117576487A CN117576487A CN202410056333.XA CN202410056333A CN117576487A CN 117576487 A CN117576487 A CN 117576487A CN 202410056333 A CN202410056333 A CN 202410056333A CN 117576487 A CN117576487 A CN 117576487A
- Authority
- CN
- China
- Prior art keywords
- matrix
- ground penetrating
- data
- penetrating radar
- deformable convolution
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 59
- 230000000149 penetrating effect Effects 0.000 title claims abstract description 38
- 238000001514 detection method Methods 0.000 claims abstract description 17
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 239000011159 matrix material Substances 0.000 claims description 51
- 238000012545 processing Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 9
- 238000012549 training Methods 0.000 claims description 9
- 238000012937 correction Methods 0.000 claims description 6
- 238000000354 decomposition reaction Methods 0.000 claims description 6
- 238000010586 diagram Methods 0.000 claims description 6
- 238000002372 labelling Methods 0.000 claims description 6
- 238000012847 principal component analysis method Methods 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 230000002708 enhancing effect Effects 0.000 claims description 3
- 230000036962 time dependent Effects 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 238000005259 measurement Methods 0.000 abstract description 4
- 238000010276 construction Methods 0.000 abstract description 3
- 238000013135 deep learning Methods 0.000 abstract description 2
- 238000000605 extraction Methods 0.000 abstract description 2
- 238000011160 research Methods 0.000 abstract description 2
- 238000000513 principal component analysis Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 239000011800 void material Substances 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000000969 carrier Substances 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
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
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- 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
- G06V10/771—Feature selection, e.g. selecting representative features from a multi-dimensional feature space
-
- 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
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention belongs to the field of ground penetrating radar positioning detection by adopting radio waves, and discloses a ground penetrating radar cavity target intelligent recognition method based on deformable convolution. According to the identification method disclosed by the invention, the actual measurement sample database is utilized to research the construction method of the intelligent identification deep learning network model, the existing model is optimized and improved, and a large amount of actual measurement data is subjected to data preprocessing and data enhancement so as to increase the sample number and diversity of the trainable ground penetrating radar data, so that the adaptability and generalization capability to the target features with different extraction forms are better, and the adjustment capability and the target identification capability are stronger.
Description
Technical Field
The invention belongs to the field of positioning detection of a ground penetrating radar by adopting radio waves, and particularly relates to an intelligent recognition method of road diseases such as a ground penetrating radar cavity target based on deformable convolution.
Background
The urban road and underground engineering with smooth safety are the foundation of urban development, and the urban road and various underground municipal facilities are mutually used as carriers and mutually interact to jointly ensure the safe operation of the city. However, with the rapid development of urban construction, urban road collapse accidents frequently occur, and urban operation safety and normal life of people are seriously affected. Therefore, the existence and development dynamics of various disease risk sources causing ground collapse can be timely, accurately and effectively ascertained, and the method is a great technical problem to be solved in various levels of engineering property units, maintenance departments and the like of managers, roads, underground pipelines and the like. The cavity detection requires to accurately judge whether the ground penetrating radar data contains a cavity target or not and give out cavity coordinate information. Traditional manual identification-based approaches are limited by the experience of the identified person and have slow identification speeds.
The conventional convolution operation is to divide the feature map into parts of the same size as the convolution kernel, and then perform the convolution operation, where the position of each part on the feature map is fixed. The unstructured underground defects have different target forms and different areas, so that the defect identification task is very difficult, the traditional convolutional neural network is used for extracting the whole features by means of a fixed convolutional structure, and the method has the disadvantages of weak adaptation and adjustment capability to the extracted target features with different forms, weak target identification capability and poor generalization capability. In practice, the convolution kernel of conventional neural networks is typically a fixed size, fixed size (3×3,5×5), and it is difficult to adapt to shape changes of the target.
Disclosure of Invention
The invention provides a ground penetrating radar cavity target intelligent recognition method based on deformable convolution, aiming at solving the technical problem that the conventional convolutional neural network is poor in recognition capability, and the trainable offset is increased through the deformable convolutional network, so that the method is suitable for the change of the shape of a target and the robustness of target detection is improved.
The invention adopts the following technical scheme:
in the intelligent identification method of the ground penetrating radar cavity target based on deformable convolution, the improvement is that the method comprises the following steps:
step 1, preprocessing data collected by a multi-frequency array ground penetrating radar and accumulated data;
step 2, intercepting images with cavity targets in the data images of each channel;
marking each intercepted image, and enhancing all sample data to obtain a final data set and a label;
step 4, replacing the standard convolution with the deformable convolution in the YOLOv5 to obtain a deformable convolution YOLOv5 model;
and 5, training a YOLOv5 model to obtain model weights for target detection.
Further, the data accumulated in the step 1 comprise urban main road, branch road and sidewalk data detected in the last ten years.
Further, the data preprocessing in step 1 includes zero offset correction, band-pass filtering, moving average, principal component analysis method and automatic gain.
Further, the zero offset correction is to sum up the channel data first, divide the channel data by the number of sampling points to obtain an average value, and then subtract the average value from the channel data to obtain a processing result.
Further, the principal component analysis method comprises the following steps:
the B-scan signal received by the ground penetrating radar is expressed as an orthogonal matrixWherein->For the number of measuring points of a certain line, +.>For the number of sampling points in time, for +.>Singular value decomposition is performed:，
in the above-mentioned method, the step of,，/>orthogonal matrix->Is a symmetric matrix +.>And->Feature vector of>For a diagonal matrix, the diagonal elements are their singular values, arranged from large to small, expressed as: />；
Order the，/>Is 'principal component', corresponding to time sampling information, orthogonal matrix +.>Is the principal component->And corresponding feature image->Weighted sum of->Due to->Is an orthogonal matrix, thus->，/>，Let->The mutually orthogonal rank 1 partial equidistant matrix sequences are formed by:
wherein->The ground penetrating radar receiving signal matrix is decomposed into matrix +.>Is a non-negative linear combination of (a);
derived from the nature of singular value decompositionThis is the matrix +.>"nearest" rank>Matrix of>All ranks are +.>The matrix of (a) satisfies:
，
the ground penetrating radar receiving signal matrix is expressed asWherein->And->Respectively a target echo signal matrix and a direct wave signal matrix, because of the direct wave matrix +.>Rank 1, and the reflected echo signal for the subsurface target will change, in double on the B-scan diagramCurve shape, thus target reflection echo matrix +.>Rank is greater than 1, when->In the time-course of which the first and second contact surfaces,for matrix->Is a matrix +.1, since the direct wave signal energy is far greater than the target echo signal>Middle->Dominant, rank is 1.
Further, the automatic gain is achieved by multiplying the radar record by a time-dependent gain weight function,
，
in the above-mentioned method, the step of,representing radar record after automatic gain, +.>Representing an automatic gain weight function,/->Representing a radar record before automatic gain;
determining the number of time windows by the parameter control points, overlapping half of the time windows between every two adjacent time windows when calculating the average amplitude, and calculating the following parameters according to the control points: the time window is long, the starting time of the time window and the ending time of the time window;
average vibration in the s-th time windowWeb of paperThe calculation formula is as follows:
，
in the above-mentioned method, the step of,a single-pass signal representing a radar acquisition to be processed; />Indicate->The start time of the time window; />Indicate->The expiration time of the time window; />Indicate->The number of samples of the time window;
finally, storing gain parameters of which the average amplitude of each time window corresponds to the center of the respective time window as a control point;
the weighting function is:，
in the above-mentioned method, the step of,represents the weighting factor corresponding to the center of the s-th time window, ">A balance coefficient for adjusting the magnitude of the effective amplitude after the processing; for non-time window centersThe weighting factors of each point are obtained by linear interpolation of the weighting factors of the centers of two adjacent time windows.
Further, in step 2, the number of images taken is 700, and the size of each image is 512×512.
Further, the step 3 specifically includes:
step 31, labeling a cavity target in the image;
step 32, combining the image and the labeling information to establish an image database;
step 33, expanding the data volume by adopting a data enhancement method, wherein the data enhancement method comprises horizontally turning over an image and zooming the image;
step 34, according to 7:2:1 to obtain a final training set, a verification set and a test set.
Further, step 4 specifically includes:
replacing the layer 2, 3 and 4 common convolution of the backup module in YOLOv5 with deformable convolution;
for any point P on the input feature map 0， The convolution operation is expressed as:
，
in the above-mentioned method, the step of,representing +.>Element value of location->Weights representing the corresponding positions of the convolution kernel, +.>Representing +.>Element values at the locations;
deformable convolution inIntroducing an offset for each point based on convolution，
，
Obtaining the offset pixel value by bilinear interpolation:
，
in the above-mentioned method, the step of,pixel value for x point, +.>For pixel values close to the x-point, +.>Pixel value for y-point, +.>Is the pixel value adjacent to the y-point.
Further, in the YOLOv5 model training process in step 5, epoch is 300 times, batch_size is 16, optimizer is SGD, and learning rate is 0.01.
The beneficial effects of the invention are as follows:
according to the identification method disclosed by the invention, the actual measurement sample database is utilized to research the construction method of the intelligent identification deep learning network model, the existing model is optimized and improved, and a large amount of actual measurement data is subjected to data preprocessing and data enhancement so as to increase the sample number and diversity of the trainable ground penetrating radar data, so that the adaptability and generalization capability to the target features with different extraction forms are better, and the adjustment capability and the target identification capability are stronger.
Drawings
FIG. 1 is a flow chart of the disclosed identification method;
FIG. 2 is a schematic diagram of the structure of a backhaul module;
fig. 3 is a schematic diagram of the void target detection evaluation index mAP@0.5;
fig. 4 is a graph of the target measured result.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Example 1:
the embodiment discloses a ground penetrating radar cavity target intelligent identification method based on deformable convolution, which adopts measured data collected by a multi-frequency array ground penetrating radar, obtains a final detection model by adding the deformable convolution into a YOLOv5 model, makes a data set, trains the model, and completes intelligent identification of the cavity target. And the single-stage target detection YOLO model based on deformable convolution enhances the accuracy and detection speed of the detection of the cavities with different shapes. As shown in fig. 1, the method comprises the following steps:
step 1, preprocessing data collected by a multi-frequency array ground penetrating radar and data accumulated in recent years to obtain data with obvious target characteristics;
the data accumulated in step 1 comprises urban arterial road, branch road and sidewalk data detected in the last ten years. The adopted detection equipment comprises a general LTD-2100+400MHz, a general LTD-2600+400MHz, a comprehensive road detection vehicle+multi-channel 400MHz and a three-dimensional ground penetrating radar detection vehicle+400 MHz array. The data processing means for the original data includes: zero offset correction, filtering, moving average, principal component analysis methods and automatic gain.
Zero offset correction:
sometimes, the data on the ground penetrating radar profile is positive or negative or see half-cycle asymmetry, because the data contains DC drift. The dc component needs to be eliminated or suppressed before other processing of the data can take place. The processing method is simple, firstly, the channel data is summed up, the average value is obtained by dividing the sampling point number, then the average value is subtracted from the channel data, the processing result is obtained, and the processing result is expressed as follows by a formula:
，
and (3) filtering:
the ground penetrating radar system is an ultra-wideband system, the receiver can receive external high-power interference signals, and meanwhile, the radar system can generate internal interference signals with certain bandwidth when being interfered by the external. The external interference signals in the radar signals are filtered out, and a band-pass filtering mode is usually adopted.
Sliding average:
the moving average is to average multiple channels of radar data into one channel of data in a sliding rectangular window mode so as to achieve the purpose of smoothing the radar data. The moving average can balance the energy of the direct coupling wave and the ground direct wave, and is helpful for the background elimination of the next step.
The principal component analysis method comprises the following steps:
principal Component Analysis (PCA), also called principal component analysis, is a linear transformation processing method based on minimum mean square error. As a multivariate statistical method, principal component analysis may combine the original features of a set of variables, extract so-called principal components as new features, and ensure that entropy reaches a minimum. Therefore, the number of feature variables can be reduced greatly while basic information can be maintained.
In practical application, the B-scan signal received by the ground penetrating radar is expressed as an orthogonal matrixWherein->For the number of measuring points of a certain line, +.>For the number of sampling points in time, for +.>Singular value decomposition is performed:
，
in the above-mentioned method, the step of,，/>orthogonal matrixIs a symmetric matrix +.>And->Feature vector of>For a diagonal matrix, the diagonal elements are their singular values, arranged from large to small, expressed as: />；
Order the，/>Is 'principal component', corresponding to time sampling information, orthogonal matrix +.>Is the principal component->And corresponding feature image->Weighted sum of->Due to->Is an orthogonal matrix, thus->，/>，Let->The mutually orthogonal rank 1 partial equidistant matrix sequences are formed by:
wherein->The ground penetrating radar receiving signal matrix is decomposed into matrix +.>Is a non-negative linear combination of (a);
derived from the nature of singular value decompositionThis is the matrix +.>"nearest" rank>A matrix in whichAll ranks are +.>The matrix of (a) satisfies:
，
the ground penetrating radar receiving signal matrix is expressed asWherein->And->The target echo signal matrix and the direct wave signal matrix are respectively, and the direct wave signal of each measuring point in the same measuring line is not greatly changed, so that the direct wave of each measuring point can be approximately considered to be the same in an ideal condition, namely the direct wave matrix->The rank is 1, but the reflected echo signals of the underground target are changed, and the B-scan diagram is generally in a hyperbolic shape, so the target reflected echo matrix +.>Rank is greater than 1, when->When (I)>For matrix->Is a matrix +.1, since the direct wave signal energy is far greater than the target echo signal>Middle->Dominant, its rank is 1, i.e. +.>Comprises->Is obtained by removing +.>The direct wave can be effectively removed, and the information of the target echo can be well reserved.
Automatic gain:
in the detection process, the amplitudes of different target echo signals are different, and the waveform saturation of the target position is easily caused by adopting conventional linear or curve gain parameter adjustment, so that the position and the quantity information of the target cannot be accurately obtained. Under the condition of full automation, the gain design is needed to be carried out in a weighting mode on the basis of the conventional curve gain, and the gain weighting coefficients of different time windows are adjusted through experimental data and calculation, so that the energy of each effective wave on the radar section is balanced, the processing is convenient for tracking the effective wave, and the weak signal comparison is facilitated.
Automatic gain is achieved by multiplying the radar record by a time-dependent gain weight function,
，
in the above-mentioned method, the step of,representing radar record after automatic gain, +.>Representing an automatic gain weight function,/->Representing a radar record before automatic gain;
for energy-rich reflected signals, the multiplied weight should be small; for reflected signals with small energy, the multiplied weight should be large. In order that the recording of the reflected wave will not be distorted, the weight factor should change slowly over time.
To calculate gain weight functionsFirstly, the whole time window is divided into a plurality of windows with equal time, and the energy size of the windows is used for determining the gain size of the control point.
Determining the number of time windows by the parameter control points, overlapping half of the time windows between every two adjacent time windows when calculating the average amplitude, and calculating the following parameters according to the control points: the time window is long, the starting time of the time window and the ending time of the time window;
first, theAverage amplitude in the respective time window +.>The calculation formula is as follows:
，
in the above-mentioned method, the step of,a single-pass signal representing a radar acquisition to be processed; />Indicate->The start time of the time window; />Indicate->The expiration time of the time window; />Indicate->The number of samples of the time window;
finally, storing gain parameters of which the average amplitude of each time window corresponds to the center of the respective time window as a control point;
the weighting function is:，
in the above-mentioned method, the step of,indicate->Weighting factors corresponding to the centers of the time windows, +.>A balance coefficient for adjusting the magnitude of the effective amplitude after the processing; and (3) for the weighting factors of each point in the center of the non-time window, obtaining the weighting factors by linear interpolation of the weighting factors of the centers of two adjacent time windows.
Step 2, intercepting images with cavity targets in the data images of each channel; the number of images taken is 700, and the size of each image is 512 x 512.
Marking each intercepted image, and enhancing all sample data to obtain a final data set and a label;
step 31, labeling an underground cavity target in the radar image by using an open source image labeling tool;
step 32, combining the calibrated jpg format image with the generated txt format annotation information to establish an underground cavity radar image database;
step 33, expanding the data volume and increasing the sample diversity by adopting a data enhancement method, wherein the data enhancement method comprises horizontally turning over an image, zooming the image and the like;
step 34, according to 7:2:1 to obtain a final training set, a verification set and a test set.
Step 4, replacing the standard convolution with the deformable convolution in the YOLOv5 to obtain a deformable convolution YOLOv5 model;
replacing the layer 2, 3 and 4 common convolution of the backup module in YOLOv5 with deformable convolution; FIG. 2 is a schematic diagram of the structure of a backhaul module;
for any point P on the input feature map 0 The convolution operation is expressed as:
，
in the above-mentioned method, the step of,representing +.>Element value of location->Weights representing the corresponding positions of the convolution kernel, +.>Representing +.>Element values at the locations;
the deformable convolution introduces an offset for each point based on the convolutionOffset->Is generated by convolving the input signature with another, typically a fraction.
，
Because the position after the offset is added is a non-integer and does not correspond to the pixel points actually existing on the feature map, bilinear interpolation is needed to obtain the offset pixel value:
，
in the above-mentioned method, the step of,pixel value for x point, +.>For pixel values close to the x-point, +.>Pixel value for y-point, +.>Is the pixel value adjacent to the y-point.
And 5, training a YOLOv5 model to obtain model weights for target detection.
In the model training process, epoch is 300 times, the batch_size is 16, the optimizer is SGD, and the learning rate is 0.01. And using the obtained weight file as an inference weight to perform model test. Fig. 3 is a schematic diagram of the void target detection evaluation index mAP@0.5; fig. 4 is a graph of the target measured result.
Claims (10)
1. The intelligent ground penetrating radar cavity target identification method based on deformable convolution is characterized by comprising the following steps of:
step 1, preprocessing data collected by a multi-frequency array ground penetrating radar and accumulated data;
step 2, intercepting images with cavity targets in the data images of each channel;
marking each intercepted image, and enhancing all sample data to obtain a final data set and a label;
step 4, replacing the standard convolution with the deformable convolution in the YOLOv5 to obtain a deformable convolution YOLOv5 model;
and 5, training a YOLOv5 model to obtain model weights for target detection.
2. The intelligent identification method for the ground penetrating radar cavity target based on deformable convolution according to claim 1, wherein the intelligent identification method is characterized by comprising the following steps: the data accumulated in step 1 comprises urban arterial road, branch road and sidewalk data detected in the last ten years.
3. The intelligent identification method for the ground penetrating radar cavity target based on deformable convolution according to claim 1, wherein the intelligent identification method is characterized by comprising the following steps: the data preprocessing in the step 1 comprises zero offset correction, band-pass filtering, moving average, principal component analysis method and automatic gain.
4. The method for intelligently identifying the ground penetrating radar cavity target based on deformable convolution according to claim 3, wherein the method comprises the following steps of: the zero offset correction is performed by summing the trace data, dividing the trace data by the number of sampling points to obtain an average value, and subtracting the average value from the trace data to obtain a processing result.
5. The method for intelligently identifying the ground penetrating radar cavity target based on the deformable convolution according to claim 3, wherein the principal component analysis method comprises the following steps:
the B-scan signal received by the ground penetrating radar is expressed as an orthogonal matrixWherein->For the number of measuring points of a certain line, +.>For the number of sampling points in time, for +.>Singular value decomposition is performed:，
in the above-mentioned method, the step of,，/>orthogonal matrix->Is a symmetric matrix +.>And->Feature vector of>For a diagonal matrix, the diagonal elements are their singular values, arranged from large to small, expressed as: />；
Order the，/>Is 'principal component', corresponding to time sampling information, orthogonal matrix +.>Is the principal component->And corresponding characteristic imageWeighted sum of->Due to->Is an orthogonal matrix, thus->，/>，/>Let->The mutually orthogonal rank 1 partial equidistant matrix sequences are formed by:
wherein->The ground penetrating radar receiving signal matrix is decomposed into matrix +.>Is a non-negative linear combination of (a);
derived from the nature of singular value decompositionThis is the matrix +.>"nearest" rank>A matrix in whichAll ranks are +.>The matrix of (a) satisfies:
，
the ground penetrating radar receiving signal matrix is expressed asWherein->And->Respectively a target echo signal matrix and a direct wave signal matrix, because of the direct wave matrix +.>The rank is 1, and the reflected echo signals of the underground target are changed and are in hyperbolic shape on the B-scan diagram, so that the target reflected echo matrix +.>Rank is greater than 1, when->In the time-course of which the first and second contact surfaces,for matrix->Is a matrix +.1, since the direct wave signal energy is far greater than the target echo signal>Middle->Dominant, rank is 1.
6. The method for intelligently identifying the ground penetrating radar cavity target based on deformable convolution according to claim 3, wherein the method comprises the following steps of: automatic gain is achieved by multiplying the radar record by a time-dependent gain weight function,
，
in the above-mentioned method, the step of,representing radar record after automatic gain, +.>Representing an automatic gain weight function,/->Representing a radar record before automatic gain;
determining the number of time windows by the parameter control points, overlapping half of the time windows between every two adjacent time windows when calculating the average amplitude, and calculating the following parameters according to the control points: the time window is long, the starting time of the time window and the ending time of the time window;
average amplitude in the s-th time windowThe calculation formula is as follows:
，
in the above-mentioned method, the step of,a single-pass signal representing a radar acquisition to be processed; />Indicating the start time of the s-th time window;indicating the expiration time of the s-th time window; b represents the number of samples of the s-th time window;
finally, storing gain parameters of which the average amplitude of each time window corresponds to the center of the respective time window as a control point;
the weighting function is:，
in the above-mentioned method, the step of,represents the weighting factor corresponding to the center of the s-th time window, ">A balance coefficient for adjusting the magnitude of the effective amplitude after the processing; and (3) for the weighting factors of each point in the center of the non-time window, obtaining the weighting factors by linear interpolation of the weighting factors of the centers of two adjacent time windows.
7. The intelligent identification method for the ground penetrating radar cavity target based on deformable convolution according to claim 1, wherein the intelligent identification method is characterized by comprising the following steps: in step 2, the number of images taken is 700, and the size of each image is 512×512.
8. The method for intelligently identifying the ground penetrating radar cavity target based on the deformable convolution according to claim 1, wherein the step 3 specifically comprises the following steps:
step 31, labeling a cavity target in the image;
step 32, combining the image and the labeling information to establish an image database;
step 33, expanding the data volume by adopting a data enhancement method, wherein the data enhancement method comprises horizontally turning over an image and zooming the image;
step 34, according to 7:2:1 to obtain a final training set, a verification set and a test set.
9. The method for intelligently identifying the ground penetrating radar cavity target based on the deformable convolution according to claim 1, wherein the step 4 specifically comprises the following steps:
replacing the layer 2, 3 and 4 common convolution of the backup module in YOLOv5 with deformable convolution;
for any point P on the input feature map 0 The convolution operation is expressed as:
，
in the above-mentioned method, the step of,representing +.>Element value of location->Weights representing the corresponding positions of the convolution kernel, +.>Representing +.>Element values at the locations;
the deformable convolution introduces an offset for each point based on the convolution，
，
Obtaining the offset pixel value by bilinear interpolation:
，
in the above-mentioned method, the step of,pixel value for x point, +.>For pixel values close to the x-point, +.>Pixel value for y-point, +.>Is the pixel value adjacent to the y-point.
10. The intelligent identification method for the ground penetrating radar cavity target based on deformable convolution according to claim 1, wherein the intelligent identification method is characterized by comprising the following steps: in the YOLOv5 model training process of step 5, epoch is 300 times, the batch_size is 16, the optimizer is SGD, and the learning rate is 0.01.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410056333.XA CN117576487A (en) | 2024-01-16 | 2024-01-16 | Intelligent ground penetrating radar cavity target identification method based on deformable convolution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410056333.XA CN117576487A (en) | 2024-01-16 | 2024-01-16 | Intelligent ground penetrating radar cavity target identification method based on deformable convolution |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117576487A true CN117576487A (en) | 2024-02-20 |
Family
ID=89862637
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410056333.XA Pending CN117576487A (en) | 2024-01-16 | 2024-01-16 | Intelligent ground penetrating radar cavity target identification method based on deformable convolution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117576487A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112462346A (en) * | 2020-11-26 | 2021-03-09 | 西安交通大学 | Ground penetrating radar roadbed defect target detection method based on convolutional neural network |
| CN116884036A (en) * | 2023-07-11 | 2023-10-13 | 华中农业大学 | Live pig posture detection method, device, equipment and medium based on YOLOv5DA |
| CN117218545A (en) * | 2023-09-27 | 2023-12-12 | 杭州电子科技大学 | LBP feature and improved Yolov 5-based radar image detection method |
-
2024
- 2024-01-16 CN CN202410056333.XA patent/CN117576487A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112462346A (en) * | 2020-11-26 | 2021-03-09 | 西安交通大学 | Ground penetrating radar roadbed defect target detection method based on convolutional neural network |
| CN116884036A (en) * | 2023-07-11 | 2023-10-13 | 华中农业大学 | Live pig posture detection method, device, equipment and medium based on YOLOv5DA |
| CN117218545A (en) * | 2023-09-27 | 2023-12-12 | 杭州电子科技大学 | LBP feature and improved Yolov 5-based radar image detection method |
Non-Patent Citations (3)
| Title |
|---|
| 冯鹏飞: "基于半监督3D-CNN的探地雷达管线识别研究", 中国学位论文全文数据库, 11 November 2022 (2022-11-11), pages 3 * |
| 苏军雄等: "基于可变形卷积神经网络的手势识别方法", 计算机与现代化, 31 December 2018 (2018-12-31) * |
| 马朝猛: "输水管道泄漏无损雷达探测试验研究", 中国优秀硕士学位论文全文数据库（电子期刊）, 31 December 2018 (2018-12-31), pages 038 - 1087 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105787937B (en) | A kind of high resolution remote sensing image road change detecting method based on OSM | |
| CN109255100B (en) | Urban rainfall inversion method based on microwave attenuation characteristic response fingerprint identification | |
| CN105260709A (en) | Water meter detecting method, apparatus, and system based on image processing | |
| CN103941254A (en) | Soil physical property classification recognition method and device based on geological radar | |
| CN109766936A (en) | Image change detection method based on information transmitting and attention mechanism | |
| CN112308873B (en) | Edge detection method for multi-scale Gabor wavelet PCA fusion image | |
| CN105787950A (en) | Infrared image sea-sky-line detection algorithm based on line gradient accumulation | |
| CN109359604B (en) | Method for identifying instrument under shadow interference facing inspection robot | |
| CN111751392A (en) | Steel bar corrosion detection method based on dual-polarization ground penetrating radar | |
| CN105974120A (en) | Automatic detection device and method for C-reactive protein chromaticity | |
| CN111695473A (en) | Tropical cyclone strength objective monitoring method based on long-time and short-time memory network model | |
| CN111414954A (en) | Rock image retrieval method and system | |
| CN105759274A (en) | Typhoon attention area radar rainfall estimation method | |
| CN109901111A (en) | Near-field sound source localization method based on Partial Least Squares Regression | |
| CN115356732A (en) | InSAR deformation result-oriented potential landslide risk area identification method | |
| CN115203625A (en) | Drought and waterlogging index data missing value interpolation method and device | |
| CN111365624A (en) | Intelligent terminal and method for detecting leakage of brine transportation pipeline | |
| CN113920255B (en) | High-efficient mapping system based on point cloud data | |
| CN117576487A (en) | Intelligent ground penetrating radar cavity target identification method based on deformable convolution | |
| CN115761606A (en) | Box electric energy meter identification method and device based on image processing | |
| CN114563771A (en) | Double-threshold laser radar cloud layer detection algorithm based on cluster analysis | |
| CN113191536A (en) | Near-ground environment element prediction model training and prediction method based on machine learning | |
| CN107590824B (en) | Rock particle identification and displacement tracking method based on three-dimensional image processing technology | |
| CN111123356A (en) | Abnormal track intelligent identification method based on first arrival information | |
| CN117173618B (en) | Ground penetrating radar cavity target identification method based on multi-feature sensing Faster R-CNN |
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 |