Ground penetrating radar target detection method and related apparatus

Abstract

Embodiments of the present disclosure provide a target detection method and device of ground penetrating radar, computer equipment, readable storage medium and program, and relate to the technical field of target detection. The detection method comprises: acquiring a two-dimensional profile graph detected by the ground penetrating radar and a plurality of single-channel waveform graphs corresponding to the two-dimensional profile graph; performing multi-modal fusion on the two-dimensional profile graph and the plurality of single-channel waveform graphs to obtain a fused feature graph; and acquiring a detection target by a target detection model according to the fused feature graph, wherein the target detection model is a single-shot multi-box detection neural network model. The target detection method provided by the embodiments of the present disclosure can realize target detection of the ground penetrating radar.

Description

Technical Field

[0001] This disclosure relates to the field of target detection technology, and in particular to a target detection method, apparatus, computer equipment, readable storage medium, and program for ground penetrating radar. Background Technology

[0002] Currently, Ground Penetrating Radar (GPR) is a widely used non-destructive testing (NDT) device for detecting underground targets. It is characterized by its speed, non-destructive nature, safety, and high efficiency, and has gradually become the mainstream method in NDT. Its working principle primarily involves sending high-frequency electromagnetic waves to the object being tested to detect the electromagnetic waves reflected by the target. During propagation, electromagnetic waves encounter different types of media, causing changes in their path, electromagnetic intensity, and waveform. When the receiving antenna receives the reflected electromagnetic waves, multiple continuous electromagnetic wave signals can form a two-dimensional radar waveform diagram displayed on the radar receiving interface. Based on the time, amplitude, and radar waveform data, the type, size, and location of the target inside the detected object can be inferred and identified. Summary of the Invention

[0003] This disclosure provides a target detection method for ground penetrating radar, a method and apparatus for protecting applications, a computer device, a readable storage medium and a program, relating to the field of target detection technology, and establishes a fault detection model that improves the accuracy of fault detection.

[0004] This disclosure provides a target detection method for ground-penetrating radar, comprising: acquiring a two-dimensional profile image detected by the ground-penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image; performing multimodal fusion on the two-dimensional profile image and the multiple single-channel waveform images to obtain a fused feature map; and obtaining a detected target based on the fused feature map through a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model.

[0005] In one embodiment, multimodal fusion of the two-dimensional profile and the plurality of single-channel waveforms includes: encoding and decoding the plurality of single-channel waveforms to obtain a plurality of corresponding feature matrices; converting the two-dimensional profile into a PASCAL VOC dataset; passing the PASCAL VOC dataset through the third convolutional layer of the object detection model to obtain a third convolutional feature map; and fusing the plurality of corresponding feature matrices with corresponding portions of the third convolutional feature map.

[0006] In one embodiment, obtaining the detection target through the target detection model based on the fused feature map includes: inputting the fused feature map into the fourth convolutional layer of the target detection model to obtain the fourth convolutional feature map.

[0007] In one embodiment, obtaining the detected target through the target detection model based on the fused feature map includes: arranging a spatial channel attention mechanism in the ninth convolutional layer of the target detection model to obtain a ninth convolutional spatial channel attention feature map; and arranging a spatial channel attention mechanism in the tenth convolutional layer of the target detection model to obtain a tenth convolutional spatial channel attention feature map.

[0008] In one embodiment, deploying a spatial channel attention mechanism in the ninth convolutional layer of the object detection model to obtain the ninth convolutional spatial channel attention feature map includes: compressing the spatial dimension of the feature map extracted from the ninth convolutional layer of the object detection model using average pooling and max pooling; summing the results of average pooling and max pooling using a shared fully connected layer; obtaining the weight of each channel of the input feature layer using the sigmoid function; obtaining a channel attention feature map based on the weight of each channel; performing global max pooling and global average pooling on the channel attention feature map to obtain two different feature layers; stacking the two different feature layers to obtain a stacked feature map; adjusting the number of channels using a single convolution with 1 channel based on the stacked feature map; obtaining the weight of each feature point in the input feature layer using the sigmoid function; and obtaining the ninth convolutional spatial channel attention feature map based on the weight of each feature point.

[0009] In one embodiment, the method further includes: arranging prior boxes in the fourth, seventh, eighth, ninth, tenth, and eleventh convolutional layers of the target detection model; and optimizing the prior boxes of each convolutional layer.

[0010] This disclosure provides a target detection device for ground-penetrating radar, comprising: an acquisition module for acquiring a two-dimensional profile image detected by the ground-penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image; a fusion module for performing multimodal fusion of the two-dimensional profile image and the multiple single-channel waveform images to obtain a fused feature map; and a detection module for obtaining a target to be detected based on the fused feature map using a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model.

[0011] This disclosure provides a computer device including a processor, a memory, and an input / output interface; the processor is connected to the memory and the input / output interface respectively, wherein the input / output interface is used to receive data and output data, the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device performs the method as described in any of the above embodiments.

[0012] This disclosure provides a computer-readable storage medium storing a computer program adapted to be loaded and executed by a processor, such that a computer device having the processor performs the method described in any of the above embodiments.

[0013] This disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the above embodiments.

[0014] The ground-penetrating radar target detection method of this application acquires a two-dimensional profile image detected by the ground-penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image; performs multimodal fusion of the two-dimensional profile image and the multiple single-channel waveform images to obtain a fused feature map; and obtains the detected target based on the fused feature map through a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model, which can realize target detection by ground-penetrating radar. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram of an exemplary system architecture for a ground-penetrating radar target detection method to which embodiments of the present disclosure can be applied is shown;

[0017] Figure 2 This is a flowchart of a ground-penetrating radar target detection method provided in an embodiment of this disclosure;

[0018] Figure 3 This is a schematic diagram of the structure and process of the single-shot multi-frame detection neural network model (SSD) for ground-penetrating radar target detection provided in this embodiment of the present disclosure;

[0019] Figure 4 This is a flowchart of a method for multimodal fusion of the two-dimensional cross-sectional view and the plurality of single-channel waveforms provided in this disclosure embodiment;

[0020] Figure 5 This is a flowchart of a method for obtaining a detection target based on the fused feature map using a target detection model, provided in an embodiment of this disclosure.

[0021] Figure 6This is a flowchart of a method for obtaining a ninth convolutional spatial channel attention feature map by arranging a spatial channel attention mechanism in the ninth convolutional layer of the target detection model, as provided in this embodiment of the disclosure.

[0022] Figure 7 This is a schematic diagram of the spatial channel attention mechanism module provided in an embodiment of this disclosure;

[0023] Figure 8 This is a schematic diagram of the channel attention mechanism module provided in an embodiment of this disclosure;

[0024] Figure 9 This is a schematic diagram of the spatial attention mechanism module provided in an embodiment of this disclosure;

[0025] Figure 10 This is a schematic diagram of the structure of a ground-penetrating radar target detection device provided in an embodiment of this disclosure;

[0026] Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. Detailed Implementation

[0027] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0028] In this embodiment of the disclosure, a two-dimensional profile image detected by ground penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image can be obtained based on technologies such as single-shot multi-frame detection (SSD). The two-dimensional profile image and the multiple single-channel waveform images are fused in a multimodal manner to obtain a fused feature map. Based on the fused feature map, a target detection model is used to obtain the target, wherein the target detection model is a single-shot multi-frame detection neural network model, thereby enabling target detection of the two-dimensional profile image of ground penetrating radar.

[0029] The following is a brief explanation of some of the terms used in this disclosure:

[0030] Single-Shot MultiBox Detector (SSD) is a deep learning model for object detection. "Single-shot" refers to a single neural network predicting the locations and classes of multiple objects simultaneously in a single forward pass, while "multi-box" refers to applying multiple convolutional layers across the entire image to predict bounding boxes of different scales and aspect ratios for each predicted object location, and using non-maximum suppression to obtain the final detection result. The advantages of SSD are that it can achieve real-time object detection with relatively low computational cost and can detect objects of different scales and aspect ratios. VGG16 has 16 layers, including 13 convolutional layers and 3 fully connected layers. The SSD neural network used in this application is based on VGG16, with some modifications to the network's fundamental components.

[0031] In neural networks, the concatenation operation combines feature maps of the same size. Its purpose is to fuse features extracted by multiple convolutional feature extraction frameworks or to fuse information from the output layer, specifically by merging the number of channels.

[0032] Prior bounding boxes are pre-defined boxes of different sizes and aspect ratios on an image. These can be manually set within the network model, and their positions can be adjusted during model training to ultimately locate the object being detected.

[0033] The solutions provided in this disclosure involve technologies such as single-shot multi-frame detection.

[0034] Figure 1 A schematic diagram of an exemplary system architecture 100 for which the ground-penetrating radar target detection method of the present disclosure embodiments can be applied is shown.

[0035] like Figure 1 As shown, system architecture 100 may include one or more of terminals 101, 102, and 103, a network 104, and a server 105. Network 104 is the medium used to provide a communication link between terminals 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0036] It should be understood that Figure 1 The number of terminals, networks, and servers shown is merely illustrative. Depending on implementation needs, there can be any number of terminals, networks, and servers. For example, server 105 could be a server cluster consisting of multiple servers.

[0037] Terminals 101, 102, and 103 interact with server 105 via network 104, and can receive or send messages, etc. Terminals 101, 102, and 103 can be various electronic devices with displays, including but not limited to smartphones, tablets, laptops, and desktop computers, etc.

[0038] Server 105 can be a server that provides various services. For example, terminal 103 (or terminal 101 or 102) sends a ground-penetrating radar target detection command to server 105. Server 105 can obtain a two-dimensional profile image detected by ground-penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image; perform multimodal fusion of the two-dimensional profile image and the multiple single-channel waveform images to obtain a fused feature map; and obtain the detected target through a target detection model based on the fused feature map, wherein the target detection model is a single-shot multi-frame detection neural network model, thereby realizing target detection by ground-penetrating radar.

[0039] The terminal can be a mobile phone (such as terminal 101), a tablet computer (such as terminal 102), or a desktop computer (such as terminal 101), etc., without limitation. The terminal can display an application, which could be a ground-penetrating radar target detection application, etc. Figure 1 The terminals mentioned are only a portion of the devices cited in this disclosure; the terminals are not limited to these. Figure 1 The equipment listed in the text.

[0040] It is understood that the terminal mentioned in the embodiments of this disclosure can be a user equipment, and the server in the embodiments of this disclosure includes, but is not limited to, a server or a cluster of servers. The terminal mentioned above can be an electronic device, including but not limited to mobile phones, tablets, smart voice interaction devices, smart home appliances, in-vehicle terminals, desktop computers, laptops, handheld computers, in-vehicle equipment, augmented reality / virtual reality (AR / VR) devices, head-mounted displays, smart TVs, wearable devices, digital cameras, webcams, and other mobile internet devices (MIDs) with network access capabilities.

[0041] The servers mentioned above can be cloud servers that provide basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, vehicle-road cooperation, content delivery networks (CDN), and big data and artificial intelligence platforms. They can also be independent physical servers, or server clusters or distributed systems composed of multiple physical servers.

[0042] Optionally, the data involved in the embodiments of this disclosure may be stored in a cloud platform, or may be stored based on cloud storage technology or blockchain technology, without limitation.

[0043] Improving the target recognition accuracy in B-scan (two-dimensional profile) images is of great significance for the field of ground penetrating radar (GPR) underground detection. Due to the complexity of the underground environment, B-scan images often contain a large amount of interference and noise. Currently, B-scan images have low resolution and it is difficult to obtain target outlines. Traditional methods for B-scan image analysis mainly rely on manual recognition and machine learning. However, with the increase in the amount of detection data, manual recognition becomes inefficient and highly subjective. Traditional machine learning methods (Support Vector Machine, Scale Invariant Feature Transform (SIFT), Hough Transform, etc.) do not consider the original radar data when extracting features from B-scan images, resulting in low recognition effectiveness and generalization ability. Both manual data analysis and traditional machine learning methods suffer from drawbacks such as incomplete information, high cost, and time-consuming processes, failing to flexibly distinguish between different types of target objects in B-scan images, thus resulting in low recognition accuracy.

[0044] Figure 2 This is a flowchart illustrating a target detection method for ground-penetrating radar provided in this disclosure. The method provided in this disclosure can be... Figure 1 The terminal or server in the embodiments executes the commands, or the commands are executed by the interaction between the terminal and the server.

[0045] like Figure 2 As shown, the method provided in this disclosure embodiment may include the following steps.

[0046] In step S210, a two-dimensional profile image detected by ground penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image are obtained.

[0047] In this step, the terminal or server acquires a two-dimensional profile image detected by ground-penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image. The two-dimensional profile image is a B-scan image, and the single-channel waveform images are A-scan images. The multiple single-channel waveform images corresponding to the two-dimensional profile image refer to the B-scan image being obtained by processing multiple single-channel waveform images.

[0048] In step S220, the two-dimensional cross-sectional view and the multiple single-channel waveforms are fused in a multimodal manner to obtain a fused feature map.

[0049] In this step, the terminal or server performs multimodal fusion of the two-dimensional profile and the multiple single-channel waveforms to obtain a fused feature map.

[0050] In step S230, the target is obtained by a target detection model based on the fused feature map, wherein the target detection model is a single-shot multi-frame detection neural network model.

[0051] In this step, the terminal or server obtains the target to be detected based on the fused feature map through a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model.

[0052] Figure 2 The ground-penetrating radar (GPR) target detection method involves acquiring a two-dimensional profile image detected by the GPR and multiple single-channel waveform images corresponding to the two-dimensional profile image; performing multimodal fusion of the two-dimensional profile image and the multiple single-channel waveform images to obtain a fused feature map; and obtaining the detected target based on the fused feature map using a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model. This enables target detection by the GPR.

[0053] Figure 3 This is a schematic diagram of the structure and process of the single-shot multi-frame detection neural network model (SSD) for ground-penetrating radar target detection provided in this embodiment of the present disclosure. Figure 3 The SSD is based on VGG16, with some modifications to the network infrastructure. Figure 3 In this context, Conv1 represents the first convolutional layer, and so on for the others. Concatenate represents the concatenation operation.

[0054] Figure 4 This is a flowchart illustrating a method for multimodal fusion of the two-dimensional cross-sectional view and the plurality of single-channel waveforms provided in this disclosure. The method provided in this disclosure can be... Figure 1 The terminal or server in the embodiments executes the commands, or the commands are executed by the interaction between the terminal and the server.

[0055] like Figure 4As shown, the method provided in this disclosure embodiment may include the following steps.

[0056] In step S410, the multiple single-channel waveforms are encoded and decoded to obtain multiple corresponding feature matrices.

[0057] In this step, the terminal or server encodes and decodes the multiple single-channel waveforms to obtain multiple corresponding feature matrices.

[0058] In one embodiment, single-channel waveform A-scan data processing is performed to prepare for multimodal fusion of the model in this application. The original A-scan data contains rich feature information. This application considers encoding and decoding the A-scan data to transform it into a feature matrix, which is then fed into the network model for multimodal fusion, improving the network's generalization and expressive capabilities. The principle of an encoder-decoder is to convert signals into digital values ​​so that computers can process them. Converting A-scan signals into feature matrices can be achieved through an encoder-decoder. The specific processing flowchart includes the following steps:

[0059] Data selection: Select A-scan data suitable for the actual application as needed, and remove redundant information;

[0060] Feature encoding: Converting multiple consecutive A-scan data into numerical format through encoding and decoding for computation and analysis.

[0061] Feature combination: The decoded features are combined into a feature matrix and used as input data for subsequent multimodal fusion in the model.

[0062] Feature normalization dimensionality reduction: Standardize or normalize feature data to make it comparable across different units and avoid errors caused by dimensionality.

[0063] Feature dimensionality reduction: This involves reducing the dimensionality of the feature matrix to decrease the size of the dataset and computational complexity, while ensuring the correlation between data and the preservation of useful information.

[0064] In step S420, the two-dimensional profile is converted into a PASCAL VOC dataset.

[0065] In this step, the terminal or server converts the two-dimensional profile into a PASCAL VOC dataset.

[0066] In one embodiment, after obtaining the corresponding two-dimensional profile B-scan data, preprocessing operations such as background removal and data filtering are performed on the B-scan images to filter out clutter and reduce noise interference. Data annotation is then performed on the B-scan images, and finally, the data is used as model input for training in the PASCAL VOC dataset format. The B-scan images are then converted to the PASCAL VOC dataset format.

[0067] In step S430, the PASCAL VOC dataset is passed through the third convolutional layer of the object detection model to obtain the third convolutional feature map.

[0068] In this step, the terminal or server obtains the third convolutional feature map by passing the PASCAL VOC dataset through the third convolutional layer of the object detection model.

[0069] like Figure 3 As shown, the B-scan image is converted into the PASCAL VOC dataset and then passed through the first, second, and third convolutional layers to obtain the third convolutional feature map.

[0070] In step S440, the plurality of corresponding feature matrices are fused with the corresponding portions of the third convolutional feature map.

[0071] In this step, the terminal or server fuses the multiple corresponding feature matrices with the corresponding portions of the third convolutional feature map.

[0072] like Figure 3 As shown, the multiple corresponding feature matrices are fused with the corresponding portions of the third convolutional feature map, and the fused matrices are then input into the fourth convolutional layer for further feature extraction. Here, "multiple corresponding feature matrices are fused with the corresponding portions of the third convolutional feature map" means that the single-channel waveform corresponding to the feature matrix is ​​fused with the portions of the same single-channel waveform in the two-dimensional cross-sectional image.

[0073] In one embodiment, obtaining the detection target through the target detection model based on the fused feature map includes: inputting the fused feature map into the fourth convolutional layer of the target detection model to obtain the fourth convolutional feature map.

[0074] Figure 5 This is a flowchart illustrating a method for obtaining a detected target using a target detection model based on the fused feature map, as provided in this embodiment of the disclosure. The method provided in this embodiment can be... Figure 1 The terminal or server in the embodiments executes the commands, or the commands are executed by the interaction between the terminal and the server.

[0075] like Figure 5 As shown, the method provided in this disclosure embodiment may include the following steps.

[0076] In step S510, a spatial channel attention mechanism is arranged in the ninth convolutional layer of the target detection model to obtain the ninth convolutional spatial channel attention feature map.

[0077] In this step, the terminal or server deploys a spatial channel attention mechanism in the ninth convolutional layer of the target detection model to obtain the ninth convolutional spatial channel attention feature map.

[0078] In step S520, a spatial channel attention mechanism is arranged in the tenth convolutional layer of the target detection model to obtain the tenth convolutional spatial channel attention feature map.

[0079] In this step, the terminal or server deploys a spatial channel attention mechanism in the tenth convolutional layer of the target detection model to obtain the tenth convolutional spatial channel attention feature map.

[0080] Figure 6 This is a flowchart illustrating a method for obtaining a ninth convolutional spatial channel attention feature map by arranging a spatial channel attention mechanism in the ninth convolutional layer of the target detection model, as provided in this embodiment. The method provided in this embodiment can be... Figure 1 The terminal or server in the embodiments executes the commands, or the commands are executed by the interaction between the terminal and the server.

[0081] like Figure 6 As shown, the method provided in this disclosure embodiment may include the following steps.

[0082] In step S610, the spatial dimension of the ninth convolutional feature map extracted from the ninth convolutional layer of the target detection model is compressed using average pooling and max pooling.

[0083] In this step, the terminal or server uses average pooling and max pooling to compress the spatial dimension of the feature map extracted from the ninth convolutional layer of the target detection model.

[0084] In step S620, the results of average pooling and max pooling are summed using a shared fully connected layer.

[0085] In this step, the terminal or server adds the results of average pooling and max pooling using a shared fully connected layer.

[0086] In step S630, the weights of each channel of the input feature layer are obtained through the sigmoid function.

[0087] In this step, the terminal or server obtains the weights of each channel of the input feature layer through the sigmoid function.

[0088] In step S640, a channel attention feature map is obtained based on the weight of each channel.

[0089] In this step, the terminal or server obtains the channel attention feature map based on the weight of each channel.

[0090] In step S650, the channel attention feature map is subjected to global max pooling and global average pooling to obtain two different feature layers.

[0091] In this step, the terminal or server performs global max pooling and global average pooling on the channel attention feature map to obtain two different feature layers.

[0092] In step S660, the two different feature layers are stacked to obtain a stacked feature map.

[0093] In this step, the terminal or server stacks the two different feature layers to obtain a stacked feature map.

[0094] In step S670, the number of channels is adjusted by a single convolution with 1 channel based on the stacked feature map, and the weight of each feature point in the input feature layer is obtained by using the sigmoid function.

[0095] In this step, the terminal or server adjusts the number of channels using a single convolution with 1 channel based on the stacked feature map, and uses the sigmoid function to obtain the weight of each feature point in the input feature layer.

[0096] In step S680, the ninth convolutional spatial channel attention feature map is obtained based on the weight of each feature point.

[0097] In this step, the terminal or server obtains the ninth convolutional spatial channel attention feature map based on the weight of each feature point.

[0098] Figure 7 This is a schematic diagram of the spatial channel attention mechanism module provided in an embodiment of this disclosure. Figure 7 The left side shows the convolutional feature map obtained from the previous convolutional layer, and the right side shows the convolutional feature map after being processed by the spatial channel attention mechanism module.

[0099] refer to Figure 7The spatial channel attention module is composed of a channel attention module and a spatial attention module. The input feature layer first enters the channel attention module, obtaining feature maps with different channel weights, and then enters the spatial attention module. Important features on the resulting output feature layer are amplified and learned by the network. The weight distribution differences for important information are more pronounced in the feature layers located at the ninth and tenth convolutional layers. This application considers adding a spatial channel attention module to these two layers to retain more feature information of the object to be detected in the B-scan image and better grasp global information.

[0100] Figure 8 This is a schematic diagram of the channel attention mechanism module provided in an embodiment of this disclosure.

[0101] refer to Figure 8 This is a typical structure for the channel attention module. This module uses average pooling and max pooling to compress the spatial dimensionality of the input feature map. Max pooling only considers the largest element, ignoring other elements within the region, thus preserving more texture information of the target in the B-scan image. Average pooling calculates the average value of all elements in the pooling region to retain more background information of the target. The results of average and max pooling are then summed using a shared fully connected layer. Finally, the sigmoid function is used to obtain the weight (between 0 and 1) for each channel of the input feature layer. This weight is then multiplied by the original input feature layer. Here, H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels.

[0102] Figure 9 This is a schematic diagram of the spatial attention mechanism module provided in an embodiment of this disclosure.

[0103] refer to Figure 9 This is a typical structure for a spatial attention module. The feature layers obtained through the channel attention mechanism are subjected to global max pooling and global average pooling on each feature plane, resulting in two different feature layers. These two results are then stacked, and the number of channels is adjusted using a single convolution with a channel count of 1. The sigmoid activation function is used to obtain the weights (between 0 and 1) for each feature point in the input feature layer. After obtaining these weights, they are multiplied by the original input feature layer. Finally, the output feature information of the target object in the B-scan image, containing relatively comprehensive feature information, is obtained.

[0104] In one embodiment, the ground-penetrating radar target detection method further includes: arranging prior boxes in the fourth, seventh, eighth, ninth, tenth, and eleventh convolutional layers of the target detection model; and optimizing the prior boxes of each convolutional layer.

[0105] The feature maps of the fourth, seventh, eighth, ninth, tenth, and eleventh convolutional layers are used. For each feature map cell, the regression module predicts the location of the target box appearing in the feature map cell, and the classification module determines the category of the object appearing in the feature map cell, as well as the predicted target box location and the corresponding object category.

[0106] After the model design is completed, parameter optimization training is performed. Model training and tuning are conducted according to different application scenarios and detection requirements of the GPR collected data. This includes setting the scale, aspect ratio, and number of prior boxes, the learning rate, weights, and activation functions. After generating prior boxes with different proportions and numbers, the smooth L1 loss function is used to perform position coordinate regression correction on the target prior boxes, and the softmax loss function is used to classify the targets based on confidence. Finally, non-maximum suppression (NMS) is used to remove duplicate and redundant prior boxes, ultimately obtaining detection results that meet the requirements. The network model can be trained and optimized by adjusting different parameter settings.

[0107] Figure 10 This is a schematic diagram of the structure of a ground-penetrating radar target detection device provided in an embodiment of this disclosure.

[0108] like Figure 10 As shown, the ground-penetrating radar target detection device 1000 provided in this embodiment may include:

[0109] The acquisition module 1010 is used to acquire a two-dimensional profile image detected by ground penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image;

[0110] The fusion module 1020 is used to perform multimodal fusion of the two-dimensional profile and the multiple single-channel waveforms to obtain a fused feature map;

[0111] The detection module 1030 is used to obtain the detection target based on the fused feature map through a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model.

[0112] The ground-penetrating radar target detection device of this application includes an acquisition module for acquiring a two-dimensional profile image detected by the ground-penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image; a fusion module for performing multimodal fusion of the two-dimensional profile image and the multiple single-channel waveform images to obtain a fused feature map; and a detection module for obtaining a target to be detected based on the fused feature map through a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model, which can realize target detection by ground-penetrating radar.

[0113] In one embodiment, the fusion module 1020 is further configured to encode and decode the plurality of single-channel waveforms to obtain a plurality of corresponding feature matrices; convert the two-dimensional profile into a PASCAL VOC dataset; pass the PASCAL VOC dataset through the third convolutional layer of the object detection model to obtain a third convolutional feature map; and fuse the plurality of corresponding feature matrices with the corresponding parts of the third convolutional feature map respectively.

[0114] In one embodiment, the detection module 1030 is further configured to input the fused feature map into the fourth convolutional layer of the target detection model to obtain the fourth convolutional feature map.

[0115] In one embodiment, the detection module 1030 is further configured to deploy a spatial channel attention mechanism in the ninth convolutional layer of the target detection model to obtain a ninth convolutional spatial channel attention feature map; and to deploy a spatial channel attention mechanism in the tenth convolutional layer of the target detection model to obtain a tenth convolutional spatial channel attention feature map.

[0116] In one embodiment, the detection module 1030 is further configured to: compress the spatial dimension of the feature map extracted from the ninth convolutional layer of the target detection model using average pooling and max pooling; sum the results of average pooling and max pooling using a shared fully connected layer; obtain the weight of each channel of the input feature layer using the sigmoid function; obtain a channel attention feature map based on the weight of each channel; perform global max pooling and global average pooling on the channel attention feature map to obtain two different feature layers; stack the two different feature layers to obtain a stacked feature map; adjust the number of channels using a single convolution with 1 channel based on the stacked feature map; obtain the weight of each feature point of the input feature layer using the sigmoid function; and obtain the ninth convolutional spatial channel attention feature map based on the weight of each feature point.

[0117] In one embodiment, the system further includes an optimization module for arranging prior boxes in the fourth, seventh, eighth, ninth, tenth, and eleventh convolutional layers of the target detection model and optimizing the prior boxes of each convolutional layer.

[0118] See Figure 11 , Figure 11 This is a schematic diagram of the structure of a computer device 1100 provided in an embodiment of this disclosure. For example... Figure 11As shown, the computer device in this embodiment may include one or more processors 1101, a memory 1102, and an input / output interface 1103. The processor 1101, memory 1102, and input / output interface 1103 are connected via a bus 1104. The memory 1102 stores a computer program, which includes program instructions. The input / output interface 1103 receives and outputs data, such as for data interaction between the host machine and the computer device, or for data interaction between various virtual machines within the host machine. The processor 1101 executes the program instructions stored in the memory 1102.

[0119] The processor 1101 can perform the following operations:

[0120] A two-dimensional profile image detected by ground penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image are obtained; the two-dimensional profile image and the multiple single-channel waveform images are fused in a multi-modal manner to obtain a fused feature map; based on the fused feature map, a target detection model is used to obtain the detected target, wherein the target detection model is a single-shot multi-frame detection neural network model.

[0121] In some feasible implementations, the processor 1101 may be a central processing unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0122] The memory 1102 may include read-only memory and random access memory, and provides instructions and data to the processor 1101 and the input / output interface 1103. A portion of the memory 1102 may also include non-volatile random access memory. For example, the memory 1102 may also store device type information.

[0123] In practice, the computer device can execute the implementation methods provided by the steps in the above embodiments through its built-in functional modules. For details, please refer to the implementation methods provided by the steps in the above embodiments, which will not be repeated here.

[0124] This disclosure provides a computer device including a processor, an input / output interface, and a memory. The processor retrieves a computer program from the memory and executes the steps of the method shown in the above embodiments to perform a transmission operation.

[0125] This disclosure also provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and execute the methods provided in the steps of the above embodiments. Specific implementations of the steps in the above embodiments can be found therein and will not be repeated here. Furthermore, the beneficial effects of using the same method will not be repeated here either. For technical details not disclosed in the embodiments of the computer-readable storage medium involved in this disclosure, please refer to the description of the method embodiments of this disclosure. As an example, the computer program can be deployed to execute on a single computer device, or on multiple computer devices located in one location, or on multiple computer devices distributed across multiple locations and interconnected via a communication network.

[0126] The computer-readable storage medium can be the apparatus provided in any of the foregoing embodiments or the internal storage unit of the computer device, such as the hard disk or memory of the computer device. The computer-readable storage medium can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the computer device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of the computer device. The computer-readable storage medium is used to store the computer program and other programs and data required by the computer device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0127] This disclosure also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative embodiments described above.

[0128] The terms "first," "second," etc., used in the specification, claims, and drawings of this disclosure are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or modules, but may optionally include steps or modules not listed, or may optionally include other step units inherent to these processes, methods, apparatuses, products, or devices.

[0129] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0130] The methods and related apparatuses provided in this disclosure are described with reference to the method flowcharts and / or structural diagrams provided in this disclosure. Specifically, each block of the method flowchart and / or structural diagram, as well as combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions are provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable transmission device to create a machine, such that the instructions, which execute via the processor of the computer or other programmable transmission device, generate instructions for implementing the process. Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable transmission device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable transmission device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 A process or multiple processes and / or structures illustrate the steps of the functions specified in one or more boxes.

[0131] The above-disclosed embodiments are merely preferred embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Therefore, any equivalent variations made in accordance with the claims of this disclosure shall still fall within the scope of this disclosure.

Claims

1. A target detection method for ground-penetrating radar, characterized in that, include: Obtain a two-dimensional profile image detected by ground penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image; The two-dimensional profile and the multiple single-channel waveforms are fused using multimodal methods to obtain a fused feature map. Based on the fused feature map, the target is obtained through a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model. The multimodal fusion of the two-dimensional cross-sectional view and the multiple single-channel waveforms includes: The multiple single-channel waveforms are encoded and decoded to obtain multiple corresponding feature matrices; The two-dimensional profile is converted into the PASCAL VOC dataset. The PASCAL VOC dataset is passed through the third convolutional layer of the object detection model to obtain the third convolutional feature map; The multiple corresponding feature matrices are respectively fused with the corresponding parts of the third convolutional feature map; The step of obtaining the detection target through the target detection model based on the fused feature map includes: The fused feature map is input into the fourth convolutional layer of the target detection model to obtain the fourth convolutional feature map.

2. The method according to claim 1, characterized in that, Based on the fused feature map, the target to be detected is obtained through the target detection model, including: A spatial channel attention mechanism is deployed in the ninth convolutional layer of the target detection model to obtain the ninth convolutional spatial channel attention feature map. A spatial channel attention mechanism is deployed in the tenth convolutional layer of the target detection model to obtain the tenth convolutional spatial channel attention feature map.

3. The method according to claim 2, characterized in that, A spatial channel attention mechanism is deployed in the ninth convolutional layer of the object detection model to obtain the ninth convolutional spatial channel attention feature map, including: The spatial dimension of the ninth convolutional feature map extracted from the ninth convolutional layer of the target detection model is compressed using average pooling and max pooling. The results of average pooling and max pooling are added together using a shared fully connected layer; The weights of each channel in the input feature layer are obtained using the sigmoid function; Obtain the channel attention feature map based on the weight of each channel; The channel attention feature map is subjected to global max pooling and global average pooling to obtain two different feature layers; The two different feature layers are stacked to obtain a stacked feature map; The number of channels is adjusted using a single convolution with 1 channel based on the stacked feature map, and the weights of each feature point in the input feature layer are obtained using the sigmoid function. The ninth convolutional spatial channel attention feature map is obtained based on the weight of each feature point.

4. The method according to claim 1, characterized in that, Also includes: Prior boxes are arranged in the fourth, seventh, eighth, ninth, tenth and eleventh convolutional layers of the target detection model; The prior boxes for each convolutional layer are optimized.

5. A target detection device for ground-penetrating radar, characterized in that, include: The acquisition module is used to acquire a two-dimensional profile image detected by ground penetrating radar and multiple single-channel waveform images corresponding to the two-dimensional profile image; The fusion module is used to perform multimodal fusion of the two-dimensional profile and the multiple single-channel waveforms to obtain a fused feature map; The detection module is used to obtain the detection target based on the fused feature map through a target detection model, wherein the target detection model is a single-shot multi-frame detection neural network model. The fusion module is further configured to encode and decode the multiple single-channel waveforms to obtain multiple corresponding feature matrices; convert the two-dimensional profile into a PASCAL VOC dataset; pass the PASCAL VOC dataset through the third convolutional layer of the object detection model to obtain a third convolutional feature map; and fuse the multiple corresponding feature matrices with the corresponding parts of the third convolutional feature map respectively. The detection module is also used to input the fused feature map into the fourth convolutional layer of the target detection model to obtain the fourth convolutional feature map.

6. A computer device, characterized in that, Includes processor, memory, and input / output interfaces; The processor is connected to the memory and the input / output interface respectively, wherein the input / output interface is used to receive data and output data, the memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device executes the method of any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-4.

Images