CN115311531A - Ground penetrating radar underground cavity target automatic detection method of RefineDet network model - Google Patents
Ground penetrating radar underground cavity target automatic detection method of RefineDet network model Download PDFInfo
- Publication number
- CN115311531A CN115311531A CN202210881472.7A CN202210881472A CN115311531A CN 115311531 A CN115311531 A CN 115311531A CN 202210881472 A CN202210881472 A CN 202210881472A CN 115311531 A CN115311531 A CN 115311531A
- Authority
- CN
- China
- Prior art keywords
- image
- ground penetrating
- penetrating radar
- target
- pooling
- 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
Images
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
- G01S13/88—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
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/28—Details of pulse systems
- G01S7/285—Receivers
- G01S7/292—Extracting wanted echo-signals
- G01S7/2923—Extracting wanted echo-signals based on data belonging to a number of consecutive radar periods
-
- 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
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/35—Details of non-pulse systems
- G01S7/352—Receivers
- G01S7/354—Extracting wanted echo-signals
-
- 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
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/273—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion removing elements interfering with the pattern to be recognised
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- 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
- G06V10/774—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
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computer Networks & Wireless Communication (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Electromagnetism (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention provides a method for automatically detecting an underground cavity target of a ground penetrating radar of a RefineDet network model. The method comprises the steps of carrying out background elimination on an acquired echo image of an underground cavity target of the ground penetrating radar to obtain a ground penetrating radar echo image with suppressed transverse ripples; the generated ground penetrating radar echo image is gained, and the position characteristics of the hollow target pixels are highlighted; performing noise reduction processing by using the gained image to inhibit clutter influence; pre-screening the processed ground penetrating radar echo image, and marking the cavity in the image by using labelimg according to the result of manual identification and field confirmation; making the obtained data into an image detection data set; training a RefineDet network model by using the obtained detection data set to obtain a network weight parameter; and carrying out target detection on the underground cavity target ground penetrating radar echo image by using the trained network model. The invention solves the problem that the existing method is difficult to detect the underground cavity target.
Description
Technical Field
The invention belongs to the technical field of target detection of ground penetrating radar echo diagram post-processing, and particularly relates to a ground penetrating radar underground cavity target automatic detection method of a RefineDet network model.
Background
The ground penetrating radar is a detection instrument for nondestructive detection in shallow underground environment. The ground penetrating radar utilizes the difference of electromagnetic dielectric constants in underground media, reflects the difference as the difference of parameters in radar echo data, and can realize high-efficiency detection and intuitively know the distribution of the underground environment by processing the echo data. In order to visually present the echo data for manual analysis, a common approach is to list the multi-channel echo data laterally, on which the data form B-Scan images commonly used in ground penetrating radar analysis are based.
The ground penetrating radar is used as an important geophysical method with high speed, high resolution and no damage, and has important significance and value in underground cavity hidden danger detection research and engineering practice. The ground penetrating radar technology can be suitable for various road conditions, irreversible damage to a detection target cannot be caused, the detection result has the characteristics of good real-time performance and high precision, the requirements of wide application range, no damage, high efficiency and accuracy of road disease detection are met, and the detection task of the underground cavity of the road is fitted. The ground penetrating radar system can be composed of a single or a plurality of pairs of transmitting and receiving antennas, each pair of transmitter and receiver can acquire a single B-Scan image by scanning a candidate area, and the distribution condition of the underground environment can be obtained by analyzing and verifying the B-Scan image. At present, B-Scan images acquired in actual engineering need to be interpreted and interpreted manually, the method is low in efficiency, and the problems of missed detection or false detection are frequent. The detection of the image is the advanced work of image identification, and not only needs to identify whether a cavity target exists in the ground penetrating radar echo image, but also needs to frame information such as the rough position, the size and the like of the target. The detection and identification of underground cavity targets by using the current mainstream partial deep learning method also has problems. The relevant mode information of the underground cavity is difficult to obtain through confirmation, verification and positioning, the mode and the shape of the underground cavity in the B-Scan image are not fixed, and in addition, the acquisition of a large number of ground penetrating radar echo image samples of the underground cavity is a project task with higher requirements.
Disclosure of Invention
The invention aims to solve the problem that the underground cavity target is difficult to detect and identify by the existing method, and provides an automatic detection method for the underground cavity target of the ground penetrating radar of a RefineDet network model.
The invention is realized by the following technical scheme, and provides a ground penetrating radar underground cavity target automatic detection method of a RefineDet network model, which specifically comprises the following steps:
step 1: background elimination is carried out on the obtained ground penetrating radar echo image of the underground cavity target, and a ground penetrating radar echo image with suppressed transverse ripples is obtained;
step 2: the ground penetrating radar echo image generated in the step 1 is gained, the background is suppressed, and the characteristics of the holes submerged in the image are extracted;
and step 3: performing noise reduction on the image data subjected to gain processing in the step 2, and inhibiting the influence of clutter;
and 4, step 4: pre-screening the ground penetrating radar echo image processed in the step 3, and marking the cavity in the image by using labelimg according to the result of manual identification and field confirmation;
and 5: making the image data obtained in the step 4 into a detection network data set in a PASCAL VOC data set format;
step 6: inputting the training set in the data set obtained in the step 5 into a RefineDet network, and training the training set to obtain a weight model;
and 7: and (5) inputting the test set in the data set obtained in the step (5) into the obtained weight model, and carrying out target detection on the underground cavity target ground penetrating radar echo image.
Further, in step 1, performing image background elimination by using a transverse ripple suppression filtering method to obtain a ground penetrating radar echo image with suppressed transverse ripples.
Further, a node type mean linear gain method is used for extracting the characteristics of the cavity target from the redundant background information, the node type mean linear gain can highlight the curve characteristics of the cavity in the background, and the position and shape characteristics of the cavity target can be obtained more clearly.
Further, the node-type mean linear gain method specifically includes:
firstly, dividing a picture into 7 parts according to longitudinal average, and then respectively corresponding a starting line of each part and a final line of an image to a node, namely 8 nodes in total;
and then taking the average value of the maximum value of each row of pixels of each part as the gain size of the corresponding node, obtaining a pre-gain curve through linear interpolation, calibrating the pre-gain curve by using the maximum value to obtain a gain curve, wherein each row of the image corresponds to a point on the gain curve, the point size is the gain size of the row, and the image is gained according to the gain curve.
Further, the noise reduction adopts fast non-local mean denoising.
Further, in step 5, the image data is made into a detection network data set in a PASCAL VOC data set format, wherein a portion of the non-target images and the hole images are allocated to training of the network in a certain proportion, and the remaining portion of the non-target images and the hole images are allocated to testing of the network.
Further, the step 6 specifically includes:
the RefineDet network is divided into 4, 5 or 6 layers; the main network is divided into ResNet50, ghost _ ResNet50 and their different proportion mixing pooling models, wherein the maximum pooling weight and the average pooling weight are increased from 0 to 1, the step length is 0.1, and the connection blocks are divided into original connection blocks and connection blocks adding attention mechanism;
the Ghost module is specifically: in the original process, an input image with the size of w x h is assumed to generate output with the number of channels of n and the size of w '× h' after the input image with the size of w × h and n groups of k × k convolution kernels act; and (3) replacing the traditional convolution operation with simple operation phi in a Ghost module to obtain a Ghost characteristic diagram: firstly, processing an input image by using m groups of k multiplied by k convolution kernels to generate an intrinsic characteristic diagram with m channels and w 'multiplied by h' size, then acting operation phi on the intrinsic characteristic diagram to generate a Ghost characteristic diagram, and obtaining the output of a module after synthesis;
the operation phi adopts a deep convolution DWC, one convolution kernel in the DWC corresponds to one channel, a single convolution kernel only processes the corresponding channel, and the operation phi is equivalent to the operation of a convolution kernel of dxd;
the attention mechanism is specifically as follows:
designing from two dimensions of a channel and a space; the channel attention mechanism is specifically processed as follows: pooling the input characteristic images in two modes respectively to correspondingly obtain description of two angles of background information, processing the results of the two pooling processes by using a multilayer perceptron sharing a weight, then adding the two results, inputting the results into a channel weight judgment function, and selecting a sigmoid function;
output M c Is denoted as M c (F) = σ (MLP (AvgPool (F)) + MLP (MaxPool (F))), where F denotes the input profile and σ denotes the sigmoid function;
the processing of the spatial attention mechanism is specifically as follows: for an input feature image, two abstraction methods are used on a channel of each feature point to obtain multi-dimensional description of the feature image, after an abstraction result is spliced, a convolution is used for checking the splicing result to abstract and further extract features, finally the result is sent to a space weight judgment function, a sigmoid function is used, so that a module obtains the weight of each feature point of an input feature layer, and the weight is multiplied by the original input feature layer based on the space feature weight to obtain a result after space attention processing; output result M s Expressed as:
M s (F)＝σ(f n×n ([AvgPool(F)；MaxPool(F)]) Where F represents an input feature map, F n×n Representing a convolution kernel operation of size n.
Further, the hybrid pooling of the backbone network is specifically: in feature extraction, maximum pooling is used in the shallow layer and average pooling is used in the deep layer.
The invention provides electronic equipment, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor realizes the steps of the method for automatically detecting the underground cavity target of the ground penetrating radar of the RefineDet network model when executing the computer program.
The invention provides a computer-readable storage medium for storing computer instructions, wherein the computer instructions, when executed by a processor, implement the steps of the method for automatically detecting the underground cavity target of the ground penetrating radar of the RefineDet network model.
The invention has the beneficial effects that:
the method comprises the steps of carrying out background elimination, gain and noise reduction on an echo image of an underground cavity target of the existing ground penetrating radar, turning and amplifying the obtained image, manufacturing a data set with a PASCAL VOC data set format, training a RefineDet network model by using data in part of the data set, and testing the obtained network model by using residual data in the data set after training. According to the test result, the invention can be applied to the data which is not seen by the actual network. The invention is used for the detection task of a target image, labels an attention target, namely labels the position of the target on the basis of identifying and classifying the target, and is an advanced task of target identification. The output result is used for realizing the detection of the underground cavity target of the ground penetrating radar echo image, and the detection and identification probability can be effectively improved by adopting the method for detecting the underground cavity target of the ground penetrating radar echo image. The method is used for detecting the underground cavity target, the accuracy rate reaches more than 87%, and the omission factor does not exceed 8.3%.
In practice, when the ground penetrating radar collects relevant data of the underground cavity, the shape of the underground cavity is random and difficult to predict, and meanwhile, the depth, the size and the position of the underground cavity are unknown, so that the detection task is difficult, and great obstruction is generated on the data collection of the underground cavity and the subsequent classification and detection based on deep learning. The invention aims to train a RefineDet network model by using an underground cavity target echo image data set and realize the detection of the underground cavity target echo image of the ground penetrating radar by using the obtained model.
Drawings
FIG. 1 is a flowchart of an automatic detection method for an underground cavity target of a ground penetrating radar based on a RefineDet model according to the invention.
Fig. 2 is a diagram of a RefineDet network model, in which (a) is a 4-layer model, (b) is a 5-layer model, and (c) is a 6-layer model.
Fig. 3 is a schematic diagram of the TCB structure in the RefineDet network model, where (a) is the original structure and (b) is the structure after attention mechanism is added.
Fig. 4 is a schematic diagram of the original convolution and the Ghost convolution operation, where (a) is the original convolution and (b) is the Ghost operation.
FIG. 5 is a schematic of maximum pooling and average pooling.
FIG. 6 is a ground penetrating radar echo original image of a single underground cavity target.
FIG. 7 is an image of a ground penetrating radar echo image of a single underground cavity target after background elimination.
FIG. 8 is a ground penetrating radar echo image gained image of a single underground cavity target.
FIG. 9 is a noise-reduced image of a ground penetrating radar echo image of a single underground cavity target.
Fig. 10 is a diagram of a RefineDet training curve.
Fig. 11 is a training curve of the refineedet model after the attention mechanism is added, wherein (a) is a 4-layer model training curve, (b) is a 5-layer model training curve, and (c) is a 6-layer model training curve.
FIG. 12 shows the results of underground cavity detection, wherein (a) is correct detection, (b) is incorrect detection, and (c) is missing detection.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The first embodiment is as follows:
the invention provides a method for automatically detecting an underground cavity target of a ground penetrating radar based on a RefineDet network model, which comprises the following steps:
step 1: background elimination is carried out on the acquired ground penetrating radar echo image of the underground cavity target, and a ground penetrating radar echo image with suppressed transverse ripples is obtained;
step 2: the ground penetrating radar echo image generated in the step 1 is gained, the background is suppressed, and the characteristics of the holes submerged in the image are extracted;
and step 3: performing noise reduction on the image data subjected to gain processing in the step 2, and inhibiting the influence of clutter;
and 4, step 4: and (4) pre-screening the ground penetrating radar echo images processed in the step (3), and marking the holes in the images by using labelimg according to the result of manual identification and real-field confirmation.
And 5: and (4) making the image data obtained in the step (4) into a detection network data set in a classic PASCAL VOC data set format, wherein part of the non-target images and the cavity images are distributed to the training of the network according to a certain proportion, and the rest of the non-target images and the cavity images are distributed to the testing of the network.
And 6: inputting the training set in the data set obtained in the step 5 into a RefineDet network, and training the training set to obtain a weight model;
and 7: and (5) inputting the test set in the data set obtained in the step (5) into the obtained weight model, and carrying out target detection on the underground cavity target ground penetrating radar echo image.
In the step 1, the process is carried out,
the pretreatment comprises the following steps: background elimination is carried out on the obtained ground penetrating radar echo image of the underground cavity target;
and eliminating the image background by a transverse ripple suppression filtering method to obtain the ground penetrating radar echo image with suppressed transverse ripples.
In the step 2, the process is carried out,
and extracting the characteristics of the cavity target from redundant background information by using a node type mean linear gain method. The node type mean linear gain can highlight the curve characteristics of the cavity in the background, and the position and shape characteristics of the cavity target can be obtained more clearly.
In step 2, the node-type mean linear gain method specifically includes:
firstly, the picture is divided into 7 parts according to the longitudinal average, namely according to the depth of a ground penetrating radar echo image, and then the initial line of each part and the final line of the image respectively correspond to one node, namely 8 nodes in total.
And then taking the average value of the maximum value of each row of pixels of each part as the gain size of the corresponding node, obtaining a pre-gain curve through linear interpolation, calibrating the curve by using the maximum value to obtain a gain curve, wherein each row of the image corresponds to a point on the curve, the size of the point is the gain size of the row, and the image is gained according to the gain curve.
In the step 3, the process is carried out,
the noise reduction treatment specifically comprises the following steps: and fast non-local mean denoising.
The Fast Non-Local mean denoising (Fast Non-Local Means) is an acceleration algorithm based on Non-Local mean denoising (Non-Local Means). NL-means implements filtering based on the similarity between pixels.
For an image, a search frame of size dxd is selected, and a neighborhood frame of size dxd and centered at x and y is selected. The measure of similarity of two neighborhoods is
Where the size of the neighborhood is m n, and x (i, j) and y (i, j) are the pixel values in both neighborhoods. Obtaining the measurement of each pixel point through the similarity
Where h represents a smoothing factor that affects the degree of distortion of the filtering. The final filtering result for point x is: NLmeans (x) = ∑ w (x, y) · y
The fast algorithm aims at the defect that the original method calculates the consumed time point by point, and constructs an integral image of a pixel point, thereby accelerating the filtering. The present invention constructs a 5 x 5 search window and a 3 x 3 neighborhood window.
In the step 4, the process is carried out,
and pre-screening the processed ground penetrating radar echo images, marking holes in the images by using labellimg according to the result of manual identification and field confirmation, and labeling hole targets, wherein more than one hole target may be in one image.
In the step 5, the process is carried out,
the PASCAL VOC formatted data set contains 3 major components, JPEGImages, imageSets, and antibiotics. The JPEGImages are used for storing images of the working objects, a Main folder is contained under the ImageSets folder and used for storing related txt files of the working objects, train txt files for training, test txt files for testing and the like, and the txt files store file name lists of picture files for executing respective tasks. Xml files are used for storing relevant information of corresponding pictures, such as image addresses, coordinates of labeling frames, types of targets in the labeling frames, whether the targets are difficult samples and the like.
In a step 6, the process is carried out,
the RefineDet network is developed based on SSD and belongs to a single-stage detection algorithm. The RefineDet consists of three parts: an Anchor Reference Module (ARM), an Object Detection Module (ODM), and a Connection Block (TCB). The ARM distinguishes foreground and background of the image, reduces the number of samples for the ODM, provides a refined anchor frame and optimizes a frame regression starting point. The OMD performs classification and target location based on the results. The TCB converts the ARM characteristic diagram into a form required by ODM, and avoids the region-by-region RoIploling operation.
The RefineDet network model used by the invention is divided into 4, 5 or 6 layers. The main network of the RefineDet model is divided into ResNet50, ghost _ ResNet50 and their different proportion hybrid pooling models, in which the maximum pooling weight and the average pooling weight are increased from 0 to 1 by a step size of 0.1. As shown in table 1:
TABLE 1 ResNet50 network model architecture
The Ghost _ ResNet50 obtains a required feature map by using a less calculation amount based on a certain similarity between feature maps with the same depth.
The Ghost module is specifically: assume that in the original flow, an input image with a size of w × h is subjected to the convolution kernel of n groups of k × k to generate an output with a channel number of n and a size of w '× h'. And (3) replacing the traditional convolution operation with simple operation phi in a Ghost module to obtain a Ghost characteristic diagram: firstly, processing an input image by using m groups of k multiplied by k convolution kernels to generate an intrinsic feature map with m channels and w '× h' size, then acting operation phi on the intrinsic feature map to generate a Ghost feature map, and obtaining the output of a module after synthesis. As shown in table 2:
table 2 Ghost module utilization results
For operation Φ, the present invention uses deep Convolution (DWC). Unlike ordinary convolution, in DWC one convolution kernel corresponds to one channel, and a single convolution kernel only processes the corresponding channel. The operation Φ may be equivalent to a convolution kernel operation of d × d.
The connection blocks of the RefineDet model are divided into original connection blocks and connection blocks with attention adding mechanisms.
The attention mechanism is specifically as follows:
note that the force mechanism is designed primarily from two dimensions, with different emphasis directions based on different tasks. The CBAM (Convergence Block Attention Module) combines the two dimensions of channel and space information.
The processing core idea of the channel attention mechanism is that meaningful information is mainly focused on. And pooling the input characteristic images in two modes respectively to correspondingly obtain the description of two angles of the background information. And then processing the two pooling results by using a multilayer perceptron sharing the weight, then adding the two results, and inputting the two results into a channel weight judgment function, wherein the sigmoid function is selected by the method.
Output M c Is denoted as M c (F) = σ (MLP (AvgPool (F)) + MLP (MaxPool (F))) where F denotes the input profile and σ denotes the sigmoid function.
In the processing core idea of the space attention mechanism, the main attention is paid toThe location of the meaning information. For the input feature image, two abstract methods are used on the channel of each feature point to obtain the multi-dimensional description of the feature image. After splicing the abstract results, performing abstraction and further feature extraction on the spliced results by using convolution to check, and finally sending the results into a space weight judgment function. Thus, the module obtains the weight of each feature point of the input feature layer. And multiplying the weight by the original input feature layer based on the spatial feature weight to obtain a result after spatial attention processing. Output result M s Can be expressed as: m s (F)＝σ(f n×n ([AvgPool(F)；MaxPool(F)]) Where F represents the input feature map, F n ×n Representing a convolution kernel operation of size n. With CBAM, network computing power can be distributed more efficiently.
The mixing and pooling mode specifically comprises the following steps:
in feature extraction, maximum pooling is used in the shallow layer and average pooling is used in the deep layer. The two pooling modes are used for carrying out region division on the feature map, the maximum pooling takes the maximum value of the action range as an output result, and the average pooling carries out average operation on the action range to obtain a final result. In comparison, the maximum pooling is focused on preserving edge information, while the average pooling is focused on preserving background information, but for the GPR echo image, both the edge and background information contain important information amount, and if the two are combined, higher model accuracy can be obtained. The invention optimizes the model by using mixed pooling, wherein the mixed pooling is based on probability, two pooling modes are selected, and the characteristics of the two pooling modes are combined. As shown in tables 3 and 4:
TABLE 3 adjusting hybrid pooling weight identification result performance in shallow pooling
TABLE 4 adjustment of mixed pooling weight identification result performance in deep pooling
In a step 6, the process is carried out,
in the network parameter setting, where the batch size is 16, the initial learning rate is 0.001, and halved as the number of iterations increases, the optimizer selects SGD with momentum of 0.9 and weight default of 0.0005.
In a step 7 of the method, the step of the method,
the method for carrying out target identification and classification on the ground penetrating radar underground cavity target echo image specifically comprises the steps of using a trained RefineDet model, carrying out feature extraction on the ground penetrating radar underground cavity target echo image which is not input into the system, inputting the extracted feature into the model, and automatically carrying out target detection on the ground penetrating radar underground cavity target echo image. And detecting and outputting the position, confidence degree, length and height of the hole.
Example two:
with reference to fig. 1 to 12, the present invention provides a ground penetrating radar underground cavity target automatic detection method of a RefineDet network model, and the method specifically includes:
step 1: background elimination is carried out on the obtained ground penetrating radar echo image of the underground cavity target, and a ground penetrating radar echo image with suppressed transverse ripples is obtained;
step 2: the ground penetrating radar echo image generated in the step 1 is gained, the background is suppressed, and the characteristics of the holes submerged in the image are extracted;
and step 3: performing noise reduction on the image data subjected to gain processing in the step 2, and inhibiting the influence of clutter;
and 4, step 4: and (3) pre-screening the ground penetrating radar echo image processed in the step (3), marking the cavity in the image by using labelimg according to the result of manual identification and field confirmation, and marking a void label.
And 5: and (4) making the image data obtained in the step (4) into a detection network data set in a PASCAL VOC data set format, wherein 80% of the non-target images and the cavity images are allocated to the training of the network, and 20% of the non-target images and the cavity images are allocated to the testing of the network.
Step 6: inputting the training set in the data set obtained in the step 5 into a RefineDet network, and training the training set to obtain a weight model;
and 7: and (5) inputting the test set in the data set obtained in the step (5) into the obtained weight model, and carrying out target detection on the underground cavity target ground penetrating radar echo image.
The step 1 of preprocessing the acquired ground penetrating radar echo image of the underground cavity target specifically comprises the steps of removing standing waves and filtering the image of the acquired ground penetrating radar echo image of the underground cavity target, removing the standing waves in a direct intercepting mode, and filtering by adopting a transverse mean value filtering method to obtain the ground penetrating radar echo image with the restrained transverse ripples.
In the step 2, the process is carried out,
and extracting the characteristics of the cavity target from redundant background information by using a node type mean linear gain method. The node type mean linear gain can highlight the curve characteristics of the cavity in the background, and the position and shape characteristics of the cavity target can be obtained more clearly.
In step 2, the node-type mean linear gain method specifically includes:
the picture is firstly divided into 7 parts according to the longitudinal average, then the initial line of each part and the final line of the image respectively correspond to one node, namely 8 nodes in total.
And then taking the average value of the maximum value of each row of pixels of each part as the gain size of the corresponding node, obtaining a pre-gain curve through linear interpolation, calibrating the curve by using the maximum value to obtain a gain curve, wherein each row of the image corresponds to a point on the curve, the size of the point is the gain size of the row, and the image is gained according to the gain curve.
In the step 3, the process is carried out,
the noise reduction processing specifically includes: fast non-local mean de-noising, wavelet de-noising, singular value decomposition, discrete cosine transform and Gaussian filtering.
The Fast Non-Local mean denoising (Fast Non-Local Means) is an acceleration algorithm based on Non-Local mean denoising (Non-Local Means). NL-means implements filtering based on the similarity between pixels.
For an image, a search frame of size dxd is selected, and a neighborhood frame of size dxd and centered at x and y is selected. The measure of similarity of two neighborhoods is
Where the size of the neighborhood is m n, and x (i, j) and y (i, j) are the pixel values in both neighborhoods. Obtaining the measurement of each pixel point through the similarity
Where h represents a smoothing factor that affects the degree of distortion of the filtering. The final filtering result for point x is: NLmeans (x) = ∑ w (x, y) · y
The fast algorithm aims at the defect that the original method calculates the consumed time point by point, and constructs an integral image of a pixel point, thereby accelerating the filtering. The invention constructs a 5 x 5 search window and a 3 x 3 neighborhood window.
In the step 4, the process is carried out,
and pre-screening the processed ground penetrating radar echo image, marking the cavity in the image by using labellimg according to the result of manual identification and field confirmation, and marking a void target with a void label.
In the step 5, the process is carried out,
the target detection dataset contains 3 major components, JPEGImages, imageSets, antotations. The JPEGImages are used for storing images of the working objects, a Main folder is contained under the ImageSets folder and used for storing related txt files of the working objects, train txt files for training, test txt files for testing and the like, and the txt files store file name lists of picture files for executing respective tasks. Xml files are used for storing relevant information of corresponding pictures, such as image addresses, coordinates of marking frames, types of targets in the marking frames, whether the targets are difficult samples and the like.
In a step 6, the process is carried out,
the RefineDet network model used by the invention is divided into 4, 5 or 6 layers. The main network of the RefineDet model is divided into ResNet50, ghost _ ResNet50 and their different proportion hybrid pooling models, in which the maximum pooling weight and the average pooling weight are increased from 0 to 1 by a step size of 0.1.
The Ghost _ ResNet50 obtains a required feature map by using a way with smaller calculation amount based on certain similarity between feature maps with the same depth.
The Ghost module is specifically: assume that in the original flow, an input image with a size of w × h is subjected to the convolution kernel of n groups of k × k to generate an output with a channel number of n and a size of w '× h'. And (3) replacing the traditional convolution operation with simple operation phi in a Ghost module to obtain a Ghost characteristic diagram: firstly, processing an input image by using m groups of k multiplied by k convolution kernels to generate an intrinsic feature map with m channels and w '× h' size, then acting operation phi on the intrinsic feature map to generate a Ghost feature map, and obtaining the output of a module after synthesis.
For operation Φ, the present invention uses a deep-Convolution (DWC). Unlike ordinary convolution, in DWC one convolution kernel corresponds to one channel, and a single convolution kernel only processes the corresponding channel. The operation Φ may be equivalent to a convolution kernel operation of d × d.
The connection blocks of the RefineDet model are divided into original connection blocks and connection blocks added with attention mechanisms.
The attention mechanism is specifically as follows:
note that the force mechanism is designed primarily from two dimensions, channel and space.
In the processing core idea of the channel attention mechanism, meaningful information is mainly focused. And pooling the input characteristic images in two modes respectively to correspondingly obtain the description of two angles of the background information. And then processing the two pooling results by using a multilayer perceptron sharing the weight, then adding the two results, inputting the two results into a channel weight judgment function, and selecting a sigmoid function.
Output M c Is denoted as M c (F) = σ (MLP (AvgPool (F)) + MLP (MaxPool (F))) where F represents the input profile,σ denotes a sigmoid function.
In the processing core idea of the spatial attention mechanism, the position of meaningful information is mainly focused. For the input feature image, two abstract methods are used on the channel of each feature point to obtain the multi-dimensional description of the feature image. After splicing the abstract results, performing abstraction and further feature extraction on the spliced results by using convolution check, and finally sending the results into a space weight judgment function and using a sigmoid function. Thus, the module obtains the weight of each feature point of the input feature layer. Based on the space feature weight, multiplying the weight by the original input feature layer to obtain a result after space attention processing. Output result M s Can be expressed as: m s (F)＝σ(f n×n ([AvgPool(F)；MaxPool(F)]) Where F represents the input feature map, F n×n Representing a convolution kernel operation of size n. With CBAM, network computing power can be distributed more efficiently. As shown in tables 5 and 6:
TABLE 5 detection result of attention mechanism RefineDet of layer 4 structure
TABLE 6 results of RefineDet test for multilayer structures
The mixing and pooling mode specifically comprises the following steps:
in feature extraction, maximum pooling is used in the shallow layer and average pooling is used in the deep layer. The method optimizes the model by using mixed pooling, and the mixed pooling is based on probability, selects two pooling modes and combines the characteristics of the two pooling modes.
In a step 6, the process is carried out,
in the network parameter setting, where the batch size is 16, the initial learning rate is 0.001, and halved as the number of iterations increases, the optimizer selects SGD with momentum of 0.9 and weight default of 0.0005.
In a step 7 of the method, the step of the method,
the method for carrying out target identification and classification on the echo images of the underground cavity targets of the ground penetrating radar specifically comprises the steps of using a trained RefineDet model, carrying out feature extraction on the echo images of the underground cavity targets of the ground penetrating radar which are not input into the system, inputting the extracted echo images into the model, and automatically carrying out target detection on the echo images of the underground cavity targets of the ground penetrating radar. And detecting and outputting the position, confidence degree, length and height of the hole.
The invention provides electronic equipment, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor realizes the steps of the method for automatically detecting the underground cavity target of the ground penetrating radar of the RefineDet network model when executing the computer program.
The invention provides a computer-readable storage medium for storing computer instructions, wherein the computer instructions, when executed by a processor, implement the steps of the method for automatically detecting the underground cavity target of the ground penetrating radar of the RefineDet network model.
The memory in the embodiments of the present application may be either volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The non-volatile memory may be a Read Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an electrically Erasable EPROM (EEPROM), or a flash memory. Volatile memory can be Random Access Memory (RAM), which acts as external cache memory. By way of example, and not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), dynamic random access memory (dynamic RAM, DRAM), synchronous Dynamic Random Access Memory (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), SLDRAM (synchronous DRAM), and direct rambus RAM (DR RAM). It should be noted that the memories of the methods described herein are intended to comprise, without being limited to, these and any other suitable types of memories.
In the above embodiments, all or part of the implementation may be realized by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, cause the processes or functions described in accordance with the embodiments of the application to occur, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored on a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, from one website, computer, server, or data center to another website, computer, server, or data center via wire (e.g., coaxial cable, fiber optic, digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that includes one or more available media. The usable medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a Digital Video Disk (DVD)), or a semiconductor medium (e.g., a Solid State Disk (SSD)), among others.
In implementation, the steps of the above method may be performed by integrated logic circuits of hardware in a processor or instructions in the form of software. The steps of a method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware processor, or may be implemented by a combination of hardware and software modules in a processor. The software module may be located in ram, flash memory, rom, prom, or eprom, registers, etc. storage media as is well known in the art. The storage medium is located in a memory, and a processor reads information in the memory and completes the steps of the method in combination with hardware of the processor. To avoid repetition, it is not described in detail here.
It should be noted that the processor in the embodiments of the present application may be an integrated circuit chip having signal processing capability. In implementation, the steps of the above method embodiments may be performed by integrated logic circuits of hardware in a processor or by instructions in the form of software. The processor described above may be a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components. The various methods, steps, and logic blocks disclosed in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software module may be located in ram, flash memory, rom, prom, or eprom, registers, etc. storage media as is well known in the art. The storage medium is located in a memory, and a processor reads information in the memory and completes the steps of the method in combination with hardware of the processor.
The method for automatically detecting the underground cavity target of the ground penetrating radar of the RefineDet network model is described in detail, a specific example is applied to explain the principle and the implementation mode of the method, and the description of the embodiment is only used for helping to understand the method and the core idea of the method; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed, and in summary, the content of the present specification should not be construed as a limitation to the present invention.
Claims (10)
1. A method for automatically detecting an underground cavity target of a ground penetrating radar of a RefineDet network model is characterized by specifically comprising the following steps:
step 1: background elimination is carried out on the acquired ground penetrating radar echo image of the underground cavity target, and a ground penetrating radar echo image with suppressed transverse ripples is obtained;
step 2: the ground penetrating radar echo image generated in the step 1 is gained, the background is suppressed, and the characteristics of the holes submerged in the image are extracted;
and step 3: performing noise reduction on the image data subjected to gain processing in the step 2, and inhibiting the influence of clutter;
and 4, step 4: pre-screening the ground penetrating radar echo image processed in the step 3, and marking the cavity in the image by using labellimg according to the result of artificial identification and real-time confirmation;
and 5: making the image data obtained in the step 4 into a detection network data set in a PASCALVOC data set format;
step 6: inputting the training set in the data set obtained in the step 5 into a RefineDet network, and training the training set to obtain a weight model;
and 7: and (5) inputting the test set in the data set obtained in the step (5) into the obtained weight model, and carrying out target detection on the underground cavity target ground penetrating radar echo image.
2. The method according to claim 1, wherein in step 1, image background elimination is performed by a transverse ripple suppression filtering method, and a ground penetrating radar echo image with suppressed transverse ripples is obtained.
3. The method of claim 1, wherein the hole target feature is extracted from the redundant background information by using a node-based mean linear gain method, wherein the node-based mean linear gain can highlight the curve feature of the hole in the background, so as to obtain the position and shape feature of the hole target more clearly.
4. The method according to claim 3, wherein the nodal-mean linear gain method is specifically:
firstly, dividing a picture into 7 parts according to longitudinal average, and then respectively corresponding a starting line of each part and a final line of an image to a node, namely 8 nodes in total;
and then taking the average value of the maximum value of each row of pixels of each part as the gain size of the corresponding node, obtaining a pre-gain curve through linear interpolation, calibrating the pre-gain curve by using the maximum value to obtain a gain curve, wherein each row of the image corresponds to a point on the gain curve, the point size is the gain size of the row, and the image is gained according to the gain curve.
5. The method of claim 1, wherein the noise reduction employs fast non-local mean denoising.
6. The method according to claim 1, characterized in that in step 5 the image data is made into a detection network dataset in the PASCAL VOC dataset format, wherein part of the non-target images and hole images are allocated to the training of the network in a certain proportion and the remaining part of the non-target images and hole images are allocated to the testing of the network.
7. The method according to claim 1, wherein step 6 is specifically:
the RefineDet network is divided into 4, 5 or 6 layers; the main network is divided into ResNet50, ghost _ ResNet50 and their different proportion mixing pooling models, wherein the maximum pooling weight and the average pooling weight are increased from 0 to 1, the step length is 0.1, and the connection blocks are divided into original connection blocks and connection blocks adding attention mechanism;
the Ghost module specifically comprises: in the original process, an input image with the size of w x h is assumed to generate output with the number of channels of n and the size of w '× h' after the input image with the size of w × h and n groups of k × k convolution kernels act; and (3) replacing the traditional convolution operation with simple operation phi in a Ghost module to obtain a Ghost characteristic diagram: firstly, processing an input image by using m groups of k multiplied by k convolution kernels to generate an intrinsic characteristic diagram with m channels and w 'multiplied by h' size, then acting operation phi on the intrinsic characteristic diagram to generate a Ghost characteristic diagram, and obtaining the output of a module after synthesis;
the operation phi adopts a deep convolution DWC, one convolution kernel in the DWC corresponds to one channel, a single convolution kernel only processes the corresponding channel, and the operation phi is equivalent to the operation of a convolution kernel of dxd;
the attention mechanism is specifically as follows:
designing from two dimensions of a channel and a space; the channel attention mechanism is specifically processed as follows: pooling the input characteristic images in two modes respectively to correspondingly obtain description of two angles of background information, processing the results of the two pooling processes by using a multilayer perceptron sharing a weight, then adding the two results, inputting the results into a channel weight judgment function, and selecting a sigmoid function;
output M c Is denoted as M c (F) = σ (MLP (AvgPool (F)) + MLP (MaxPool (F))), where F denotes the input profile and σ denotes the sigmoid function;
the processing of the spatial attention mechanism is specifically as follows: for an input feature image, two abstraction methods are used on a channel of each feature point to obtain multi-dimensional description of the feature image, after an abstraction result is spliced, a convolution is used for checking the splicing result to abstract and further extract features, finally the result is sent to a space weight judgment function, a sigmoid function is used, so that a module obtains the weight of each feature point of an input feature layer, and the weight is multiplied by the original input feature layer based on the space feature weight to obtain a result after space attention processing; output result M s Expressed as: m s (F)＝σ(f n×n ([AvgPool(F)；MaxPool(F)]) Where F represents an input feature map, F n×n Representing a convolution kernel operation of size n.
8. The method according to claim 7, wherein the hybrid pooling of the backbone network is specifically: in feature extraction, maximum pooling is used in the shallow layer and average pooling is used in the deep layer.
9. An electronic device comprising a memory and a processor, the memory storing a computer program, wherein the processor implements the steps of the method according to any one of claims 1-8 when executing the computer program.
10. A computer-readable storage medium storing computer instructions, which when executed by a processor implement the steps of the method of any one of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210881472.7A CN115311531A (en) | 2022-07-26 | 2022-07-26 | Ground penetrating radar underground cavity target automatic detection method of RefineDet network model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210881472.7A CN115311531A (en) | 2022-07-26 | 2022-07-26 | Ground penetrating radar underground cavity target automatic detection method of RefineDet network model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115311531A true CN115311531A (en) | 2022-11-08 |
Family
ID=83858956
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210881472.7A Pending CN115311531A (en) | 2022-07-26 | 2022-07-26 | Ground penetrating radar underground cavity target automatic detection method of RefineDet network model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115311531A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173618A (en) * | 2023-09-06 | 2023-12-05 | 哈尔滨工业大学 | Ground penetrating radar cavity target identification method based on multi-feature sensing Faster R-CNN |
| CN117409329A (en) * | 2023-12-15 | 2024-01-16 | 深圳安德空间技术有限公司 | Method and system for reducing false alarm rate of underground cavity detection by three-dimensional ground penetrating radar |
| CN117475273A (en) * | 2023-06-28 | 2024-01-30 | 苏州中析生物信息有限公司 | Infrared small target detection method based on improved refine det |
-
2022
- 2022-07-26 CN CN202210881472.7A patent/CN115311531A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117475273A (en) * | 2023-06-28 | 2024-01-30 | 苏州中析生物信息有限公司 | Infrared small target detection method based on improved refine det |
| CN117173618A (en) * | 2023-09-06 | 2023-12-05 | 哈尔滨工业大学 | Ground penetrating radar cavity target identification method based on multi-feature sensing Faster R-CNN |
| CN117173618B (en) * | 2023-09-06 | 2024-04-30 | 哈尔滨工业大学 | Ground penetrating radar cavity target identification method based on multi-feature sensing Faster R-CNN |
| CN117409329A (en) * | 2023-12-15 | 2024-01-16 | 深圳安德空间技术有限公司 | Method and system for reducing false alarm rate of underground cavity detection by three-dimensional ground penetrating radar |
| CN117409329B (en) * | 2023-12-15 | 2024-04-05 | 深圳安德空间技术有限公司 | Method and system for reducing false alarm rate of underground cavity detection by three-dimensional ground penetrating radar |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110472627B (en) | End-to-end SAR image recognition method, device and storage medium | |
| CN115311531A (en) | Ground penetrating radar underground cavity target automatic detection method of RefineDet network model | |
| CN111062321B (en) | SAR detection method and system based on deep convolutional network | |
| CN115291210A (en) | Three-dimensional image pipeline identification method of 3D-CNN ground penetrating radar combined with attention mechanism | |
| CN115661649B (en) | BP neural network-based shipborne microwave radar image oil spill detection method and system | |
| CN111639513A (en) | Ship shielding identification method and device and electronic equipment | |
| CN110543879A (en) | SSD target detection method based on SE module and computer storage medium | |
| CN116012364A (en) | SAR image change detection method and device | |
| CN115019187A (en) | Detection method, device, equipment and medium for SAR image ship target | |
| CN116152651A (en) | Sonar target identification and detection method based on image identification technology and applied to ocean | |
| CN113466839B (en) | Side-scan sonar sea bottom line detection method and device | |
| CN108921017B (en) | Face detection method and system | |
| CN116778176B (en) | SAR image ship trail detection method based on frequency domain attention | |
| CN115311532A (en) | Ground penetrating radar underground cavity target automatic identification method based on ResNet network model | |
| Kou et al. | ISAR image segmentation for space target based on contrastive learning and NL-Unet | |
| CN116047463A (en) | Multi-angle SAR target scattering anisotropy deduction method, device, equipment and medium | |
| US20230401670A1 (en) | Multi-scale autoencoder generation method, electronic device and readable storage medium | |
| CN116756486A (en) | Offshore target identification method and device based on acousto-optic electromagnetic multi-source data fusion | |
| CN116597300A (en) | Unsupervised domain self-adaptive SAR target recognition method integrating and aligning visual features and scattering topological features | |
| CN116047418A (en) | Multi-mode radar active deception jamming identification method based on small sample | |
| CN115527187A (en) | Method and device for classifying obstacles | |
| CN110135280A (en) | A kind of multiple view SAR automatic target recognition method based on sparse representation classification | |
| Shu et al. | On-board wavelet based change detection implementation of SAR flood image | |
| Zhang et al. | Entropy-Based re-sampling method on SAR class imbalance target detection | |
| CN116977746A (en) | Millimeter wave image target classification method, device, equipment and storage medium |
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 |