Ground penetrating radar image recognition method based on adaptive threshold segmentation

Abstract

This invention discloses a ground-penetrating radar (GPR) image recognition method based on adaptive threshold segmentation, relating to the fields of geophysical exploration and digital image processing. The method includes: S1, acquiring raw image data obtained from GPR detection and preprocessing it to enhance target signals and suppress noise; S2, generating an individual optimal threshold for each preprocessed GPR grayscale image, resulting in an optimal threshold set; S3, performing mean processing on this set to calculate a globally unified threshold that reflects the segmentation level of the entire detection area; S4, using the globally unified threshold to binarize all raw input grayscale images, and then performing morphological opening operations on the binarized images; S5, performing connected component analysis on the opened images and applying area thresholding again to filter out connected components smaller than a second preset area threshold, outputting the final segmentation result image; This method features fast and accurate recognition.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration and digital image processing technology, specifically to a ground-penetrating radar image recognition method based on adaptive threshold segmentation. Background Technology

[0002] Ground-penetrating radar (GPR), as a highly efficient, continuous, and non-destructive geophysical exploration technology, has been widely applied in fields such as underground defect detection, underground pipeline investigation, and geotechnical engineering surveys. This method is also gradually being applied to the non-destructive testing of existing building foundations. Its basic principle is to transmit high-frequency electromagnetic waves underground and infer the structure and location of underground targets by receiving and analyzing the waveforms, amplitudes, and time information reflected back from different media interfaces. Currently, there are publicly disclosed patent technologies related to the interpretation and target identification of GPR images. For example: Chinese invention patent CN118644691B, entitled "A Ground Penetrating Radar Image Feature Description and Matching Location Method Based on Shape Context," describes a technique that combines two-dimensional B-scan images acquired by ground penetrating radar along a road direction with mileage and location information to perform spatial gridding, thereby aligning the image with mileage and location labels. Furthermore, it designs a ground penetrating radar image structural feature extraction method based on shape context, extracting image feature points through SURF and constructing shape context feature descriptors to obtain image structural features. Finally, it achieves carrier location through feature matching.

[0003] Chinese invention patent CN115343703B, entitled "A Pipeline Recognition Method Based on Self-Trained 3D-CNN Ground Penetrating Radar 3D Images," first preprocesses the obtained ground penetrating radar 3D echo images, dividing the images into training and validation sets, and manually annotating all data in the validation set and a portion of data in the training set; then, it uses a 3D-CNN neural network model combined with an attention mechanism for self-training to obtain a trained weight model, and finally uses this model to identify pipeline targets in the ground penetrating radar 3D echo images.

[0004] Chinese invention patent application CN114187516A, entitled "A Method for Identifying Underground Cavities Based on Ground Penetrating Radar Using BP Neural Network," describes a technique that first uses robust principal component analysis to suppress clutter in ground penetrating radar B-Scan images, then extracts time-domain, frequency-domain, and wavelet-domain features of the A-Scan signal; next, it uses a BP neural network to perform overall identification of the A-Scan signal to obtain the target's horizontal region; then, it segments the A-Scan signal within the target's horizontal region to obtain the target's vertical region, ultimately determining the target's location.

[0005] The existing technology has the following drawbacks: Firstly, threshold selection is highly subjective and lacks adaptability: interpretation methods based primarily on human experience lack objective standards, resulting in varying outcomes. Traditional global threshold segmentation methods, such as the Otsu method, perform well when processing images with a single, uniform background, but when faced with ground-penetrating radar images containing complex geological noise, signal attenuation with depth, and multi-scale target structures, their segmentation results are unstable and they struggle to adaptively extract complete target features.

[0006] Secondly, the image interpretation accuracy is insufficient: In strip foundation surveying, traditional methods typically segment each image independently when processing multi-line data, ignoring the inherent correlation between images of different lines within the same survey area. This leads to significant fluctuations in segmentation results across different lines, discontinuous image features, and a tendency for phenomena such as broken foundation reflective layers, blurred or even lost internal steel reinforcement mesh to occur, affecting the accurate identification and quantitative interpretation of the foundation structure (such as top surface depth, width, and internal structure).

[0007] Thirdly, it has weak anti-interference capabilities: the on-site detection environment is complex, and radar images often contain direct waves, multiple reflected waves, and clutter interference from the surrounding environment (such as iron gates, shafts, etc.). Traditional methods are sensitive to these interferences and are prone to misidentifying interference signals as targets or filtering out target signals as noise, thus reducing the reliability of identification. Summary of the Invention

[0008] This invention provides a ground-penetrating radar image recognition method based on adaptive threshold segmentation to solve the problems of strong subjectivity in threshold selection, poor adaptability, insufficient image interpretation accuracy and weak anti-interference ability in the prior art, and to achieve rapid, accurate and automated identification and quantitative interpretation of strip-shaped foundation structures.

[0009] This invention provides the following technical solution: a ground-penetrating radar image recognition method based on adaptive threshold segmentation, comprising: S1. Acquire the raw image data obtained by ground penetrating radar detection, and perform preprocessing to enhance the target signal and suppress noise; S2. For each ground-penetrating radar grayscale image obtained after preprocessing, generate an individual optimal threshold and obtain an optimal threshold set. S3. Perform mean processing on the set to calculate a globally unified threshold that reflects the segmentation level of the entire detection area. S4. Using a globally unified threshold, binarize all original input grayscale images, and then perform morphological opening operations on the binarized images again. S5. Perform connected component analysis on the image after the opening operation, and apply the area threshold again for filtering to remove connected components smaller than the second preset area threshold, and output the final segmentation result image.

[0010] As a further improvement of the present invention, in S1, the preprocessing step includes: Linear gain is used to perform time-varying compensation on radar data to balance the signal strength between shallow and deep layers. The amplitude data is linearly normalized to the grayscale range and converted into a standard grayscale image; The invalid data areas at both ends of the image are cropped, the core range where the target may appear is retained, and the ground penetrating radar grayscale image is output.

[0011] As a further improvement of the present invention, in S2, the step of generating the individual optimal threshold includes: S2.1. Within a preset grayscale value range, generate a threshold sequence with a fixed step size; S2.2 For each candidate threshold in the threshold sequence, the grayscale image is binarized, and morphological opening operation and connected component analysis are performed on the obtained binary image to filter out regions with an area smaller than the first preset area threshold, and the number of retained connected components is counted. S2.3. Based on all candidate thresholds and their corresponding number of connected regions, generate a "threshold-number of regions" response curve, and determine the candidate threshold that makes the number of connected regions reach the maximum value as the individual optimal threshold for the grayscale image.

[0012] As a further improvement of the present invention, in step S2.2, "filtering out regions with an area smaller than the first preset area threshold" and in step S5, "removing connected regions with an area smaller than the second preset area threshold", the first preset area threshold and / or the second preset area threshold are adaptively determined based on the image resolution or a priori noise model.

[0013] As a further improvement of the present invention, in step S2.3, the candidate threshold that makes the number of connected regions reach the maximum value is determined as the individual optimal threshold of the grayscale image, including: identifying the global maximum value of the number of connected regions from the response curve, and determining the corresponding candidate threshold as the individual optimal threshold.

[0014] As a further improvement of the present invention, in step S2.3, the candidate threshold that makes the number of connected regions reach the maximum value is determined as the individual optimal threshold of the grayscale image, including: identifying the global maximum value of the number of connected regions from the response curve, and determining the corresponding candidate threshold as the individual optimal threshold.

[0015] As a further improvement of the present invention, the calculation formulas for the opening operation image and the individual optimal threshold are as follows: ; ; In the formula: B openFor the opening operation image, S is a predefined structuring element that determines the shape and size of the neighborhood of the operation; For erosion operations, S slides inside the image, and its center point is preserved only when S is completely contained within the target area; For dilation, it means sliding S across the image again, and when the center point of S sweeps across all remaining pixels, the area it covers is set as the target region; T opt The optimal threshold for an individual.

[0016] As a further improvement of the present invention, in step S2.2, after performing morphological opening operation on the obtained binary image, it is also necessary to identify all connected regions in the image after the opening operation and calculate the pixel area of ​​each region.

[0017] As a further improvement of the present invention, the calculation formula for the globally unified threshold image is as follows: , , In the formula: T uni For a globally uniform threshold, For each input image, B is the individual optimal threshold. uni-i I(x) is the image, and I(x) is the grayscale image of the ground penetrating radar.

[0018] As a further improvement of the present invention, the calculation formula for a binary image is as follows: , ; Among them, T k T is the lower limit of grayscale. low Gray values ​​after k steps ΔT; B k It is a binary image.

[0019] As a further improvement of the present invention, it also includes S6: establishing a theoretical model containing the strip foundation using the finite-difference time-domain method and performing forward modeling; inputting the simulated radar image into steps S1 to S4 for processing; and verifying the effectiveness of the method by comparing the segmentation results with the structure of the theoretical model.

[0020] The present invention has the following beneficial effects: 1. This invention can clearly and completely present the three-layer waveform structure of the top surface, bottom surface and internal steel reinforcement of a strip foundation; in forward modeling, this method can accurately segment the reflection interfaces corresponding to the top, middle and bottom layers of the foundation, and can meticulously present the distribution pattern of the internal steel reinforcement mesh, with image features highly consistent with the model structure.

[0021] 2. By combining morphological opening operation and connected region area filtering, the method of the present invention can effectively remove isolated pixels and small regions caused by geological clutter, environmental interference, etc., and smooth the target boundary. Thus, it can still stably extract the main reflection features of the strip foundation in complex detection environment, significantly improving the reliability of the segmentation results.

[0022] 3. Based on the maximization criterion of the "threshold-number of regions" response curve, this invention adaptively selects the individual optimal threshold for each image, effectively avoiding the subjectivity of manual experience selection; furthermore, it obtains a globally unified threshold for the detection area through averaging processing, so that the threshold level of each measurement line segmentation result in the same area remains consistent, enhancing the continuity and stability of multi-measurement line image interpretation.

[0023] 4. In engineering case verification, after processing with the method of this invention, the depth and width of the top surface of the strip foundation were successfully identified. The detection results had a small error with the parameters in the design drawings, which can provide reliable data support for the quantitative assessment of existing building foundations. It has the advantages of accurate interpretation results and high degree of quantification. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention.

[0025] Figure 2 This is a flowchart of the preprocessing method in this invention.

[0026] Figure 3 This is a flowchart of the method for generating the optimal threshold for an individual in this invention. Detailed Implementation

[0027] The technical solutions of the embodiments of this specification will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of this specification and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of this specification.

[0028] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0029] In the following description, terms such as “inner,” “outer,” “upper,” “lower,” “left,” and “right” are used only to facilitate the description of the embodiments and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this specification.

[0030] All data involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions. Example 1

[0031] Please see Figure 1-3 As shown, a ground-penetrating radar image recognition method based on adaptive threshold segmentation is characterized by comprising: Step 1: Data Acquisition and Image Preprocessing S1. Acquire the raw image data obtained by ground penetrating radar and perform preprocessing to enhance the target signal and suppress noise.

[0032] The preprocessing steps include: S1.1. Use linear gain to perform time-varying compensation on radar data to balance the signal strength between shallow and deep layers.

[0033] Linear gain, simply put, is the gain that is strictly proportional to the input and output. If the input increases by a certain factor, the output increases by the same factor in a constant ratio, without distortion or compression. The image is a straight line passing through the origin, without curvature, saturation, or thresholding. The formula can be expressed as: y = kx, where k is the linear gain (gain coefficient).

[0034] S1.2. Linearly normalize the amplitude data to the grayscale range of [0, 255] and convert it into a standard grayscale image.

[0035] The original amplitude data varied in size and could not be directly used to generate an image. By stretching the data proportionally, the smallest amplitude was made into pure black, the largest amplitude into pure white, and the amplitudes in between were made into different shades of gray, ultimately resulting in a standard grayscale image.

[0036] S1.3. Crop the invalid data areas at both ends of the image, retain the core range where the target may appear, and output the ground penetrating radar grayscale image.

[0037] The beginning and end of a ground-penetrating radar profile image usually indicate that the antenna has not entered / left the detection area, there is poor coupling, or there is system noise. These areas do not contain effective downlink information and are considered invalid data, which need to be cut off directly.

[0038] Step 2: Determining the optimal threshold for each individual image: S2. For each ground-penetrating radar grayscale image obtained after preprocessing, generate an individual optimal threshold and obtain an optimal threshold set.

[0039] The steps for generating the optimal threshold for an individual include: S2.1. Within the preset grayscale value range, generate a threshold sequence with a fixed step size.

[0040] Within a set grayscale range (e.g., 50 to 200), a series of candidate thresholds are generated at fixed intervals (e.g., every 5 levels). For example: 50, 55, 60, 65... all the way up to 200.

[0041] S2.2 For each candidate threshold in the threshold sequence, the grayscale image is binarized, and morphological opening and connected component analysis are performed on the resulting binary image to filter out regions with an area smaller than a first preset area threshold, and the number of retained connected components is counted.

[0042] For each candidate threshold in the threshold sequence, do three things: The first step is binarization: converting the image into black and white.

[0043] The second is the opening operation: removing small noise and making the edges cleaner.

[0044] The third step is connected region analysis: find the connected white areas in the graph, discard the smaller ones, and keep only the valid areas.

[0045] Finally, count how many areas are left.

[0046] S2.3. Based on all candidate thresholds and their corresponding number of connected regions, generate a "threshold-number of regions" response curve, and determine the candidate threshold that makes the number of connected regions reach the maximum value as the individual optimal threshold for the grayscale image.

[0047] Plot each threshold along with the corresponding number of regions retained as a curve. The threshold with the highest number of regions retained is the optimal threshold for this graph.

[0048] In step S2.2, "filtering out regions with an area smaller than the first preset area threshold" and in step S5, "removing connected regions with an area smaller than the second preset area threshold", the first preset area threshold and / or the second preset area threshold are adaptively determined based on the image resolution or a priori noise model.

[0049] In step S2.3, the candidate threshold that makes the number of connected regions reach the maximum value is determined as the individual optimal threshold of the grayscale image. This includes: identifying the global maximum value of the number of connected regions from the response curve and determining the corresponding candidate threshold as the individual optimal threshold.

[0050] In step S2.3, the candidate threshold that makes the number of connected regions reach the maximum value is determined as the individual optimal threshold of the grayscale image. This includes: identifying the global maximum value of the number of connected regions from the response curve and determining the corresponding candidate threshold as the individual optimal threshold.

[0051] In step S2.2, after performing morphological opening operations on the obtained binary image, it is also necessary to identify all connected regions in the image after the opening operation and calculate the pixel area of ​​each region.

[0052] The formula for calculating pixel area is: S 总像素 = Width in pixels × Height in pixels.

[0053] S2.1~S2.3 can also be described in the following forms: For a single ground-penetrating radar grayscale image I, each grayscale value is represented as an I(x) histogram.

[0054] S201, from the lower limit of grayscale value T low Initially, the grayscale value is increased upwards with a fixed step size ∆T until the upper limit T of the grayscale value is reached. high Generate a set of threshold sequences {T n}

[0055] S202, For each threshold T in the sequence n The radar grayscale image is binarized, and the grayscale value is less than T. n All pixels are uniformly set to a grayscale value of 0, and grayscale values ​​greater than T are also set to 0. n All pixels are uniformly set to a grayscale value of 255 to generate a binary image B. k .

[0056] in, , In the formula: n is the number of iteration steps. S203, Regarding B k Perform morphological opening operation

[27] To eliminate small isolated pixels caused by random noise and smooth region boundaries, the opening operation image B is obtained. open .

[0057] S204, Based on this, identify B open Calculate the pixel area A of all connected regions in the region. j Remove areas smaller than the preset minimum value A.min Let N(T) be the number of regions that are retained from the given region. k ).

[0058] S205. Traverse the entire threshold sequence. Based on this, a "threshold-number of regions" response curve can be plotted. The point with the most response regions on the curve is the individual's optimal threshold T. opt .

[0059] The calculation process can be expressed as follows: , ; In the formula: S is a predefined structuring element that determines the shape and size of the neighborhood of the operation; For erosion operations, S slides within the image, preserving its center point only when S is completely contained within the target area. This shrinks the object and eliminates isolated points or small protrusions smaller than S. For dilation, the operation involves sliding S across the image again. When the center point of S sweeps across all remaining pixels, the area it covers is designated as the target region. This will enlarge the object, restoring the size of the parts that have shrunk due to erosion but still retain the main structure, but it cannot restore small objects that have been completely eroded.

[0060] S206. Select the threshold in the response curve that maximizes the number of regions as the individual optimal threshold T for this image. opt .

[0061] Step 3: Determine the globally unified threshold for the detection area: S3. Perform mean processing on the set to calculate a globally unified threshold that reflects the segmentation level of the entire detection area.

[0062] In step S3, the arithmetic mean is a weighted arithmetic mean, wherein the weights are determined based on the signal quality index or location information of the corresponding image.

[0063] After exploring the individual optimal thresholds for all input images, the optimal threshold set is obtained. This set represents the sample distribution of adaptive segmentation thresholds for each image under the premise of "maximizing the number of regions". By averaging this set, a globally uniform threshold T that reflects the overall segmentation level can be calculated. uni Using T uni For each original input image I i (x) is binarized to obtain image B. uni-i The process can be expressed as follows: , , , In the formula: Let i be the set of optimal thresholds, i = 1, 2, 3... N. The optimal threshold for each input image.

[0064] Step 4: Unified threshold segmentation and post-processing: S4. Using a globally uniform threshold, binarize all original input grayscale images, and then perform morphological opening operations on the binarized images again.

[0065] S5. Perform connected component analysis on the image after the opening operation, and apply the area threshold A again. min Filter the data to remove connected regions smaller than the second preset area threshold, and output the final segmentation result image.

[0066] Step 5, Verification: S6. Establish a theoretical model containing the strip foundation using the finite-difference time-domain method and perform forward modeling. Input the simulated radar image into steps S1 to S4 for processing. Verify the effectiveness of the method by comparing the segmentation results with the structure of the theoretical model. Example 2

[0067] I. Implementation Scenarios and Equipment Parameters 1. Detection Target The existing multi-story masonry structure has a strip foundation under the wall (conforming to the 05SG811 standard drawing set). The foundation type is reinforced concrete, with a design embedment depth of 1.2m, a width of 1.0m, and an internal steel reinforcement mesh spacing of 0.2m. The detection area is along the south exterior wall of the building, with an area of ​​12m × 8m.

[0068] 2. Equipment Configuration Ground penetrating radar: SIR-3000 (GSSI, USA), center frequency 400MHz, sampling rate 2GHz, scanning speed 0.5m / s; Data processing equipment: Intel Core i7-12700K processor, 16GB memory, running Python 3.9 (OpenCV 4.5.5, NumPy 1.23.5). Auxiliary equipment: GPS positioning module (positioning accuracy ±2cm), level, measuring tape.

[0069] II. Implementation Steps Step 1: Data Acquisition and Image Preprocessing 25 parallel survey lines (0.5m spacing between lines, 12m in total length) are laid out along the detection area. The radar antenna moves at a constant speed close to the ground to collect raw radar data (format: .dzt) and simultaneously record GPS positioning information.

[0070] Time-varying gain compensation: A linear gain algorithm is adopted, with a gain coefficient of 0.02dB / ns, to compensate for the signal attenuation with depth (depth range 0~3m) and balance the signal strength of shallow (0~1m) and deep (1~3m) layers.

[0071] Gray-level normalization: The original amplitude data (range: -512~511) is normalized to the gray-level range [0, 255] through a linear transformation, as shown in the following formula: , Where A(x,y) is the original amplitude value, A min =-512, A max =511.

[0072] 1.4 Invalid Region Cropping: Cropping 50 pixels of invalid data area (no detection signal) from both the left and right ends of the image, outputting 25 preprocessed grayscale images (size: 1024×768 pixels, pixel resolution: 0.0117m / pixel).

[0073] Step 2: Determining the optimal threshold for each individual image 2.1 Threshold sequence generation: Set the threshold range [80, 180], step size 5, and generate the threshold sequence: 80, 85, 90, ..., 180 (a total of 21 thresholds).

[0074] 2.2 Binarization and Morphological Processing For each single image (such as the image of the 10th survey line), perform threshold binarization (set the threshold to 255 if the pixel value is ≥ the threshold, otherwise set it to 0); Perform morphological opening operation: Select a 3×3 square structuring element (kernel=np.ones((3,3),dtype=np.uint8)), erode first and then dilate to eliminate speckle noise; Connected component filtering: Using 8-connected component analysis, tiny regions with an area <50 pixels (corresponding to an actual size <0.0068m) are removed. 2 ).

[0075] 2.3 Response Curve Plotting and Optimal Threshold Selection Count the number of connected regions retained at each threshold and plot the "threshold-number of regions" curve (see Table 1 below). Select the threshold corresponding to the maximum number of regions (the number of regions reaches 86 when the threshold is 135, which is the peak value), and determine the optimal threshold for this individual image as 135.

[0076] Table 1 Step 3: Determine the globally unified threshold for the detection area 3.1 Collect the individual optimal thresholds for 25 images. The results are as follows (partial examples): 128, 130, 132, 135, 133, 131, 134, ..., 136 (the complete set contains 25 values). 3.2 Calculate the mean: The integer value is 133, which is used as the global unified threshold.

[0077] Step 4: Unified Threshold Segmentation and Post-processing 4.1 Global Binarization: A threshold of 133 was used to uniformly binarize the 25 preprocessed images; 4.2 Secondary Morphological Optimization: Repeat the opening operation on the 3×3 square structuring element to further smooth the target boundary; Connected region filtering: Increase the area threshold to 80 pixels (corresponding to an actual size < 0.0108m). 2 ), to remove residual clutter; 4.3 Feature Extraction: Automatically extract basic structural features from the segmentation results, including: Basic top surface: depth 1.18~1.22m (pixel row range 620~640); Base surface: Depth 2.20~2.24m (pixel row range 880~900); Reinforcing mesh: The spacing between horizontally connected areas is 20~22 pixels (corresponding to an actual spacing of 0.19~0.21m).

[0078] III. Verification of Implementation Results 1. Accuracy verification (comparison with design drawings) 2. Anti-interference capability verification Direct wave suppression: No continuous horizontal bright bands (direct wave characteristic) are found in the segmentation results, and the clutter removal rate reaches 92%; Multiple reflection wave filtering: There are no false connected regions inside the foundation, and the effective signal retention rate reaches 96%; Signal-to-noise ratio improvement: The signal-to-noise ratio of the processed image was improved from 18.7dB to 32.4dB (detected using ENVI5.6 software).

[0079] 3. Multi-line consistency verification The depth fluctuation range of the foundation top surface of the 25 survey lines is only ±0.02m, with no characteristic fractures; The connected areas of the steel mesh are continuous between adjacent survey lines, with no misalignment or loss, and the consistency reaches 97%.

[0080] This invention innovatively proposes to select the optimal threshold for a single image by analyzing the response curve of the number of connected regions as a function of a threshold, using the maximization of the number of regions as the criterion. This method can automatically identify the best segmentation threshold that best presents the target structure while effectively suppressing noise, avoiding subjective experience-based selection.

[0081] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A ground-penetrating radar image recognition method based on adaptive threshold segmentation, characterized in that, include: S1. Acquire the raw image data obtained by ground penetrating radar detection, and perform preprocessing to enhance the target signal and suppress noise; S2. For each ground-penetrating radar grayscale image obtained after preprocessing, generate an individual optimal threshold and obtain an optimal threshold set. S3. Perform mean processing on the set to calculate a globally unified threshold that reflects the segmentation level of the entire detection area. S4. Using a globally unified threshold, binarize all original input grayscale images, and then perform morphological opening operations on the binarized images again. S5. Perform connected component analysis on the image after the opening operation, and apply the area threshold again for filtering to remove connected components smaller than the second preset area threshold, and output the final segmentation result image.

2. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 1, characterized in that, In S1, the preprocessing steps include: Linear gain is used to perform time-varying compensation on radar data to balance the signal strength between shallow and deep layers. The amplitude data is linearly normalized to the grayscale range and converted into a standard grayscale image; The invalid data areas at both ends of the image are cropped, the core range where the target may appear is retained, and the ground penetrating radar grayscale image is output.

3. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 1, characterized in that, In S2, the steps for generating the optimal threshold for an individual include: S2.

1. Within a preset grayscale value range, generate a threshold sequence with a fixed step size; S2.2 For each candidate threshold in the threshold sequence, the grayscale image is binarized, and morphological opening operation and connected component analysis are performed on the obtained binary image to filter out regions with an area smaller than the first preset area threshold, and the number of retained connected components is counted. S2.

3. Based on all candidate thresholds and their corresponding number of connected regions, generate a "threshold-number of regions" response curve, and determine the candidate threshold that makes the number of connected regions reach the maximum value as the individual optimal threshold of the grayscale image.

4. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 3, characterized in that, In step S2.2, "filtering out regions with an area smaller than the first preset area threshold" and in step S5, "removing connected regions smaller than the second preset area threshold", the first preset area threshold and / or the second preset area threshold are adaptively determined based on the image resolution or a priori noise model.

5. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 3, characterized in that, In step S2.3, the candidate threshold that makes the number of connected regions reach the maximum value is determined as the individual optimal threshold of the grayscale image, including: identifying the global maximum value of the number of connected regions from the response curve, and determining the corresponding candidate threshold as the individual optimal threshold.

6. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 3, characterized in that, The formulas for calculating the opening operation image and the optimal threshold for each individual are: ； ； In the formula: B open For the opening operation image, S is a predefined structuring element that determines the shape and size of the neighborhood of the operation; For erosion operations, S slides inside the image, and its center point is preserved only when S is completely contained within the target area; For dilation, it means sliding S across the image again, and when the center point of S sweeps across all remaining pixels, the area it covers is set as the target region; T opt The optimal threshold for an individual.

7. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 1, characterized in that, In step S2.2, after performing morphological opening operations on the obtained binary image, it is also necessary to identify all connected regions in the image after the opening operation and calculate the pixel area of ​​each region.

8. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 1, characterized in that, The formula for calculating the globally uniform threshold image is: ， ， In the formula: T uni For a globally uniform threshold, For each input image, B is the individual optimal threshold. uni-i I(x) is the image, and I(x) is the grayscale image of the ground penetrating radar.

9. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 1, characterized in that, The formula for calculating a binary image is: ， ； Among them, T k T is the lower limit of grayscale. low Gray values ​​after k steps ΔT; B k It is a binary image.

10. The ground-penetrating radar image recognition method based on adaptive threshold segmentation according to claim 1, characterized in that, The method also includes S6, which uses the finite-difference time-domain method to establish a theoretical model containing the strip foundation and performs forward modeling. The simulated radar image is then input into steps S1 to S4 for processing. The effectiveness of the method is verified by comparing the segmentation results with the structure of the theoretical model.

Images