An intelligent real-time recognition and positioning method for underground targets based on ground penetrating radar and deep learning

Abstract

The present invention relates to an intelligent real-time recognition and positioning method for underground targets based on ground penetrating radar and deep learning, comprising the following steps: Step 1. Collect ground penetrating radar image samples, and then use a label generator to mark the image; Step 2. Perform K-means screening on the real frame size results of the image mark; Step 3. Use the improved non-maximum suppression principle and YOLO v3 algorithm to estimate the recognition frame; Step 4. Encapsulate the algorithm and import it into the visual system, and finally perform recognition and positioning of the ground penetrating radar image. The principle and technology of the present invention are mature, and the mutual position relationship of multiple electromagnetic waves can be obtained, so as to quickly locate and distinguish the target body in the radar image, and it belongs to the field of detection technology of ground penetrating radar.

Description

Technical Field

[0001] The present invention relates to a detection technology of a ground penetrating radar, and in particular to a method for intelligent real-time recognition and positioning of underground targets based on a ground penetrating radar and deep learning. Background Art

[0002] Ground-penetrating image detection is a hot research direction in the field of electromagnetics. Agricultural machinery is easily affected by hard foreign objects in the cultivated soil during operation. Currently, there is no effective means to provide real-time warning of objects in farmland soil.

[0003] For object detection in farmland soil, it is usually manifested as parabolic electromagnetic waves in ground penetrating radar images. At present, the analysis of ground penetrating radar measured data in most engineering examples mainly relies on empirical analysis or manual judgment, which is undoubtedly prone to produce low-precision detection results and cannot achieve real-time analysis of detection results. Therefore, qualitative analysis of electromagnetic wave signals is a current research problem, especially after electromagnetic wave signals are distorted or decayed in soil media. Summary of the invention

[0004] In view of the technical problems existing in the prior art, the purpose of the present invention is to provide an intelligent real-time identification and positioning method for underground targets based on ground penetrating radar and deep learning, which can greatly improve the real-time analysis and detection results of ground penetrating radar.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] An intelligent real-time recognition and positioning method for underground targets based on ground penetrating radar and deep learning is characterized by comprising the following steps: step 1. collecting ground penetrating radar image samples, and then using a label generator to mark the image; step 2. performing K-means screening on the real frame size results of the image mark; step 3. using the improved non-maximum suppression principle and the YOLO v3 algorithm to estimate the recognition frame; step 4. packaging the algorithm and importing it into a visual system, and finally performing recognition and positioning of the ground penetrating radar image.

[0007] As a preferred embodiment, in step 1, according to the ground penetrating radar image features, the apex of the electromagnetic wave parabola of the target body is used as the main feature point for detection and positioning to estimate the waveform posture; then the K-means method is used to screen and train the parabolic waveforms with higher occurrence frequency to obtain the optimal rectangular recognition frame size; then, for cases with electromagnetic wave oscillation signals or waveform overlap in the ground penetrating radar, the oscillation signal similarity (Intersection over Union) IoU merging principle (Vibration IoU) V-IoU is used for non-maximum suppression; finally, it is input into the YOLO v3 framework for training and output model; finally, the IoU threshold module and the YOLO v3 trained model are imported into the user's computer, and after the user selects the appropriate parameters in the pre-compiled graphical user interface, the interface software port is connected to the mobile camera-computer system using the USB interface for real-time detection. Among them, the K-means method is used to screen and train the parabolic waveforms with higher frequency to obtain the optimal rectangular recognition box size. Higher frequency refers to the size in which the parabola is selected in the rectangular box with the largest length-to-width ratio in the sample; after the user selects the appropriate parameters, the appropriate parameters refer to the IoU threshold that can complete the current actual working conditions and the probability measurement value of judging as the target object. For example, in the logistic regression calculation, the probability greater than 0.5 is automatically considered by the algorithm to be the target body.

[0008] As a preferred embodiment, in step 2, ground penetrating radar images are collected before deep learning training in the following manner: the physical properties of hard foreign matter in farmland soil are sensitive to electromagnetic waves, and appear as a parabola opening downward in the ground penetrating radar image; the ground penetrating radar passes high-frequency radio waves on the soil surface to generate very high-resolution radar data in a short time; the ground penetrating radar moves along the survey line and continuously collects a series of trajectory A-scans to form an electromagnetic wave image.

[0009] As a preferred option, the overall framework of the convolutional layer of the deep learning classic recognition algorithm YOLO v3 outputs feature maps in three different stages to quickly identify and locate ground penetrating radar images of different sizes.

[0010] As a preferred embodiment, in step 3, for the structure of ground penetrating radar image feature extraction, first convert the input ground penetrating radar image size to a specified value, then the original image will be divided into a fixed size of 13×13 by the default computer parameters; the division cell where the vertex of the parabola in the ground penetrating radar image is located is responsible for predicting the target; when marking the ground penetrating radar image sample, the prediction box rectangle A1B1C1D1 corresponds to the position of the parabola vertex, and the rectangular box in Figure 2(a) contains the parabola-shaped electromagnetic wave vertex; Figure 2(b) shows the sample encoding method of YOLO v3, point A is first defined as the upper left corner vertex of the rectangular box, tx and ty are the pixel positions of point A in the radar image; Cx and Cy then describe the position information of the parabola of the ground penetrating radar image according to the pixel width and height values ​​of each cell unit respectively. The width of B1D1 is tx and the height of C1D1 is th. Figure 2(c); finally, the type and position information of the detected target are converted into an encoding method similar to one hot.

[0011] As a preference, the ground penetrating radar image has a waveform which is only a parabola.

[0012] As a preferred embodiment, in step 4, non-maximum suppression is used. In the ground penetrating radar image, for the same soil target body, there will be a large number of predicted rectangular boxes that intersect each other or contain the center point of the parabola. The goal of the non-maximum suppression method is to remove redundant detection boxes and retain the best one. IoU is a standard performance metric for image category segmentation problems. For a given set of images, IoU gives the ratio of the intersection and union of the predicted box and the true box. In the algorithm, the non-maximum suppression algorithm calculates and sorts the list of rectangular boxes and confidence of the candidate area, and selects the inspection box with the maximum score. Then, the IoU coefficient of the inspection box with the maximum score and the current box is calculated. If the IoU is greater than the set threshold, it is deleted. This is an iterative process. Then, the second iteration will select the box with the largest score from the remaining boxes and delete those that exceed the set IoU threshold until all possible targets in the image are obtained. Assume that three parabolas are generated by a strongly conductive target body in the soil depth direction. The algorithm will identify them as three adjacent targets, but this is only one target, although their IoU thresholds are already included in the pre-set range. The V-IoU merging principle based on the oscillation signal similarity of GPR images. Obviously, we set a horizontal critical value β so that; at the same time, we set a vertical critical value α. When the N, P and Q points at the soil depth, that is, the midpoint of the parabola, meet the horizontal critical value and the vertical critical value, the algorithm will be freed from the limitation of the IoU threshold and merge the prediction boxes.

[0013] The principle of the present invention is to improve the convolutional neural network algorithm, transmit the ground penetrating radar image in real time to a mobile device with a camera function, connect it to the user's computer through a USB interface, and then identify and select the target presented as a parabolic electromagnetic wave in real time according to the detection target type set by the user. Finally, the process of human-computer interaction and output of ground penetrating radar detection results is completed.

[0014] The present invention has the following advantages over the prior art:

[0015] 1. The principle technology is mature and can obtain the relative position relationship of multiple electromagnetic waves, so as to quickly locate and distinguish the target in the radar image.

[0016] 2. The algorithm framework YOLO v3 can sample in different feature maps, so even if the ground penetrating radar image contains targets of different sizes, it can still be identified and located.

[0017] 3. Simply connect a mobile device or a computer with a camera function to complete the real-time input and output framework process.

[0018] 4. For the detection of multiple objects, this method can also be extended to objects of different shapes. As long as the propagation law of electromagnetic waves on its surface can be obtained by calculation, this method can extract the geometric features of the object efficiently and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a real-time recognition framework diagram for the deep learning detection algorithm YOLO v3.

[0020] Figure 2a-2c is a sample encoding diagram of a parabolic electromagnetic wave signal, where Figure 2a This is the YOLO v3 feature extraction diagram. Figure 2b is a real frame diagram. Figure 2c For encoding method.

[0021] Figure 3 Structural design diagram of non-maximum suppression method in ground penetrating radar images.

[0022] Figure 4 Implementation flow chart of the system as a whole.

[0023] Figure 5 To achieve the effect diagram of identification and positioning. DETAILED DESCRIPTION

[0024] The present invention will be further described in detail below in conjunction with specific implementation methods.

[0025] A method for intelligent real-time recognition and positioning of underground targets based on ground penetrating radar and deep learning, specifically comprising the following steps:

[0026] Step 1: According to the characteristics of the ground penetrating radar image, the apex of the electromagnetic wave parabola of the target body is used as the main feature point for detection and positioning to estimate the waveform pose; then the K-means method is used to screen and train the parabolic waveforms with higher frequency to obtain the optimal rectangular recognition frame size; then, for cases with electromagnetic wave oscillation signals or waveform overlap in the ground penetrating radar, the oscillation signal similarity IoU merging principle (V-IoU) is used for non-maximum suppression; finally, it is input into the YOLO v3 framework for training and output model; finally, the IoU threshold module and the YOLO v3 trained model are imported into the user's computer. After the user selects the appropriate parameters, it can be connected to the mobile camera system for real-time detection.

[0027] Step 2: Use the ground penetrating radar to move along the survey line. Because the physical properties of hard foreign objects in farmland soil are often sensitive to electromagnetic waves, most of them appear as downward-opening parabolas in ground penetrating radar images, so the radar controller can continuously collect a series of trajectories (A-scan) to form electromagnetic wave images. The collected radar images (rd3 or rd7 format) are marked and classified, and then encoded using a rectangular mark method with the parabola feature point as the midpoint.

[0028] Step 3: The input GPR image size is converted to the specified value, then the original image will be divided into a fixed size of 13×13 by the default computer parameters, such as Figure 2a As shown in FIG. 1 , the cell where the parabola vertex M in the GPR image is located is responsible for predicting the target. When labeling the GPR image samples, the rectangle A1B1C1D1 corresponds to the position of the parabola vertex. Figure 2a The rectangular box in contains the vertex of the parabola-shaped electromagnetic wave; Figure 2b The sample encoding method of YOLO v3 is shown. Point A is the upper left corner of the rectangular box. x and t y is the pixel position of point A in the radar image; C x and C y are the pixel width and height of each cell unit. The width of B1D1 is tx and the height of C1D1 is t h . Figure 2c Following a one hot encoding method, the waveform of the ground penetrating radar image targeted by this patent is only a parabola.

[0029] Step 4: Use non-maximum suppression to remove redundant detection boxes and retain the best one. IoU (Intersection over Union) is a standard performance metric for image category segmentation problems. For a given set of images, IoU gives the ratio of the intersection and union of the predicted box and the true box. In the algorithm implemented in this patent, NMS calculates and sorts the list of rectangular boxes and confidence levels of the candidate area, and selects the inspection box with the maximum score; then calculates the IoU coefficient of other scored boxes with the current box, and deletes it if the IoU is greater than the set threshold. This is an iterative process, and then the second iteration will select the boxes with the largest scores from the remaining boxes and delete those that exceed the set IoU threshold until all possible targets in the image are obtained. Figure 3 As shown, N, P and Q represent three parabolas generated by a strong conductive target in the soil depth direction; the algorithm will identify them as three adjacent targets, but this is only one target, although their IoU thresholds are already included in the pre-set range. V-IoU merging principle based on the similarity of oscillation signals of GPR images. Obviously, we set a horizontal critical value β so that; at the same time, we set a vertical critical value α. When the N, P and Q points at the soil depth, that is, the midpoint of the parabola, meet the horizontal critical value and the vertical critical value, the algorithm will be freed from the limitation of the IoU threshold and merge the prediction boxes.

[0030] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for intelligent real-time recognition and positioning of underground targets based on ground penetrating radar and deep learning, characterized by: (1) Based on the characteristics of the ground penetrating radar image, the apex of the target's electromagnetic wave parabola is used as the main feature point for detection and positioning to estimate the waveform's pose; (2) Then, the K-means method is used to screen and train the parabolic waveforms with higher occurrence frequency to obtain the optimal rectangular recognition frame size; (3) Then, for cases with electromagnetic wave oscillation signals or waveform overlap in the ground penetrating radar, non-maximum suppression is performed using the oscillation signal similarity IoU merging principle V-IoU; finally, it is input into the YOLO v3 framework for training and output model; The final input is to the YOLOv3 framework for training and output model, specifically: For the structure of GPR image feature extraction, the input GPR image size is first converted to a specified value, then the original image will be divided into a fixed size of 13 × 13 by the default computer parameters; The partitioned cell where the vertex of the parabola in the GPR image is located is responsible for predicting the target; When annotating the ground penetrating radar image sample, the prediction box rectangle corresponds to the vertex position of the parabola, and the rectangular box contains the vertex of the parabola-shaped electromagnetic wave; The sample encoding method of YOLO v3 first defines the upper left corner vertex of the rectangular box, and then defines the pixel position of the point in the radar image; then the position information of the parabola of the ground penetrating radar image is described according to the pixel width and height value of each cell unit; finally, the type and position information of the detected target are converted into a one hot encoding method; (4) Finally, the IoU threshold module and the YOLO v3 trained model are imported into the user's computer. After the user selects the appropriate parameters in the pre-compiled graphical user interface, the interface software port is connected to the mobile camera-computer system via the USB interface for real-time recognition and positioning.

2. The method for intelligent real-time identification and positioning of underground targets based on ground penetrating radar and deep learning according to claim 1, characterized in that: Before deep learning training, ground penetrating radar images were collected in the following way: the physical properties of hard foreign objects in farmland soil are sensitive to electromagnetic waves, and appear as a parabola opening downward in the ground penetrating radar image; the ground penetrating radar passes high-frequency radio waves on the soil surface to generate very high-resolution radar data in a short period of time; the ground penetrating radar moves along the survey line and continuously collects a series of trajectory A-scans to form an electromagnetic wave image.

Images