Multi-order dispersion curve automatic picking method and device based on pose estimation network

Abstract

The embodiment of the application provides a kind of based on pose estimation network's multi-order dispersion curve automatic picking method and device, the method comprises: the normalized surface wave dispersion energy matrix of frequency-phase velocity domain is input into trained pose estimation network, and the category of each order dispersion curve instance, confidence and its multiple key point coordinates in image coordinate system are output;Wherein, each order dispersion curve is defined as independent instance target, instance category represents curve order, and multiple key points correspond to the phase velocity value at the preset frequency anchor point respectively;Based on instance confidence and key point visibility identification, filter effective curve instance;Effective key point image coordinates are inversely mapped into the physical coordinates of frequency and phase velocity, and the physical coordinate sequence of multi-order dispersion curve is output according to order;The pose estimation network is preferably YOLO-Pose network, or single-stage pose estimation network with target detection branch and key point regression branch.

Description

Technical Field

[0001] This application relates to the field of geophysical near-surface exploration technology, specifically to a method and apparatus for automatically picking multi-order dispersion curves based on attitude estimation networks. Background Technology

[0002] In the field of near-surface geophysical exploration, surface wave exploration technology is a crucial method for detecting subsurface structures. The core data processing flow of this technology first requires converting the seismic surface wave signals received by the detector into the frequency-velocity domain, thereby generating a dispersion energy map. In the dispersion energy map, the horizontal axis typically represents frequency or period, and the vertical axis represents phase velocity. The continuous curve connecting the regions with the strongest energy in the map is the surface wave dispersion curve. Accurate extraction of the dispersion curve is a critical step in surface wave exploration. By inverting the obtained dispersion curve, a detailed subsurface shear wave velocity structure can be obtained. In recent years, with the improvement of exploration equipment performance and the increase in exploration resolution requirements, especially for the acquisition of background noise data, the large number of stations has resulted in massive amounts of surface wave dispersion data. Faced with such a massive data volume, the automatic acquisition of dispersion curves has become an essential requirement for improving data processing efficiency and exploration accuracy. Due to the physical characteristics of surface waves propagating underground, dispersion curves typically exhibit multi-order characteristics, including a primary fundamental order with low phase velocity located below the energy map, and higher orders (such as the first and second higher orders) distributed above the fundamental order with higher phase velocity. However, actual exploration environments are extremely complex, and dispersion energy maps are often accompanied by strong background noise. Furthermore, dispersion curves of different orders frequently exhibit energy loss, energy overlap, or order crossover in specific frequency bands. How to accurately, continuously, and automatically extract the precise coordinates of dispersion curves of different orders from such massive, low signal-to-noise ratio energy maps with multi-order crossovers has always been a core technological bottleneck that urgently needs to be overcome in this field.

[0003] Currently, there are three main technical solutions for extracting surface wave dispersion curves. The first is the manual interactive picking method. This method relies on professional technicians to analyze the surface wave dispersion energy spectrum, determine the distribution pattern of energy clusters based on experience, and interactively pick the extreme points of phase velocity at each frequency to finally fit and form a complete dispersion curve. This method has extremely low processing efficiency and is completely unable to cope with the massive exploration data generated by modern long-term, multi-station monitoring; at the same time, this process is highly dependent on the subjective experience of the operators, and the picking results of different personnel often have large deviations, especially when the energy of high-order curves is weak or interfered with by noise, which can easily lead to subjective misjudgment.

[0004] The second existing technical solution is an algorithm based on mathematical statistics and local feature tracking. This type of method covers the traditional maximum search ridge tracking method, as well as a series of improved schemes based on local mathematical optimization that have emerged in recent years. The common feature of the above schemes is that they all treat the dispersion energy map as a two-dimensional matrix, and identify the energy center by searching for energy extreme points in a local range, applying K-means clustering to partition the data points, or using Gaussian filtering combined with quantile processing. To improve tracking stability, some schemes also introduce forward and backward greedy algorithms or time series prediction models to connect discrete points of adjacent frequencies. However, such schemes are essentially still at the level of local feature extraction at the image pixel level, lacking a global macroscopic topological understanding of the dispersion curve as a complete physical entity instance. When dealing with complex seismic data where multi-order dispersion curves have severe spatial aliasing and trajectory intersections in the energy spectrum, clustering algorithms are prone to misclassifying multi-order overlapping energy blocks as the same cluster, resulting in a serious shift in the final pick point; at the same time, the tracking algorithm is also prone to path loss at energy intersection points, resulting in a hierarchical phenomenon of mutual confusion between the basic order and higher orders.

[0005] The third existing technical solution, which is also the closest to this application, employs a semantic segmentation-based deep learning method, such as the widely used U-Net model. This solution transforms the curve picking task into a binary classification problem of image pixels. After receiving the dispersion energy map, the model outputs a binary classification prediction map of the same size, where pixels identified as the location of the dispersion curve are labeled as foreground pixels, and the rest are labeled as background pixels. Although this solution introduces deep learning, it still suffers from two major technical drawbacks. First, this solution is essentially a non-end-to-end two-step architecture. The output of the image segmentation model is a strip-shaped region with a certain pixel width, rather than direct physical coordinate points. Therefore, complex post-processing algorithms such as image skeleton extraction and local maximum search must be introduced to forcibly refine the wide strip into single-pixel lines before converting them into numerical coordinates. This cumbersome process not only significantly increases the computational cost but also easily causes geometric center shift during the refinement process, thus introducing uncontrollable secondary calculation errors. Second, existing segmentation models lack the ability to identify the instance attributes and order attributes of curves. Semantic segmentation can only determine whether a pixel belongs to a curve region; it cannot directly distinguish between basic and higher-order curves during the detection phase. When dispersion curves of different orders spatially overlap, alias, or fuse energy in a specific frequency band, the segmentation model will identify them as a single connected pixel block. Due to the lack of global instance information, subsequent clustering analysis or skeleton stripping algorithms are prone to order classification errors or path loss when dealing with such adhesion phenomena, ultimately resulting in poor separation of multi-order dispersion curves in high-frequency bands.

[0006] Therefore, in the current surface wave exploration data processing, the efficiency of manually extracting dispersion curves is extremely low, and the existing automated algorithms are prone to cascading, cumbersome post-processing, and inability to effectively separate independent multi-order curves under complex noise interference and multi-order dispersion curve cross-adhesion. Summary of the Invention

[0007] To address one of the aforementioned technical deficiencies, this application provides a method and apparatus for automatically picking multi-order dispersion curves based on attitude estimation networks.

[0008] According to a first aspect of the embodiments of this application, an automatic picking method for multi-order dispersion curves based on a pose estimation network is provided, comprising the following steps: The surface wave dispersion energy map or normalized dispersion energy matrix in the frequency-phase velocity domain is input into the trained attitude estimation network model, and the model outputs at least one dispersion curve instance target and its corresponding key point coordinates in the image coordinate system, instance category and instance confidence; wherein, each order dispersion curve in the dispersion energy map is identified as an independent instance target, the instance category is used to characterize the order of the dispersion curve, and the phase velocity value corresponding to the preset frequency anchor point on each order dispersion curve is defined as the key point of the instance target; Based on the instance confidence level and / or key point visibility identifier, the validity of each dispersion curve instance target is determined; when there is no valid dispersion curve instance target with a confidence level higher than the preset threshold, the input surface wave dispersion energy map is determined to be an invalid signal and rejected. When there are valid dispersion curve instances, the image coordinates of the valid key points are inversely mapped to physical coordinates of frequency and phase velocity based on a preset coordinate mapping relationship, and a multi-order dispersion curve physical coordinate sequence is output according to the order corresponding to the instance category.

[0009] In an optional embodiment of this application, defining the phase velocity value corresponding to a preset frequency anchor point on each dispersion curve as a key point of the instance target includes: For each dispersion curve, N sampling frequency points are selected along the frequency axis according to linear, logarithmic or piecewise non-uniform frequency anchoring strategies. Each sampling frequency point is defined as a key point by the corresponding phase velocity position on the dispersion energy map. Each instance target has an order category label, and the i-th key point corresponds to the i-th preset frequency anchor point. Each key point also includes a visibility indicator. When the dispersion energy at the corresponding frequency anchor point is lower than the threshold, the curve exceeds the preset velocity range, or the curve of that order is not recognizable, the key point is marked as invisible.

[0010] In an optional embodiment of this application, the pose estimation network model is a pose estimation network with an object detection branch and a keypoint regression branch, which includes at least: A backbone feature extraction network is used to extract deep semantic features from the dispersion energy map; The feature coupling module is used to fuse feature maps of different scales from the backbone network, enhancing the model's ability to perceive weak dispersion features. The prediction head is decoupled for parallel execution of instance target classification, localization, and key point coordinate regression.

[0011] In an optional embodiment of this application, the instance confidence is obtained by the attitude estimation network model evaluating and scoring the global topology and local energy characteristics of the dispersion curve instance target, and the key point visibility identifier is used to characterize whether there is an effective phase velocity point that can be used for output at the corresponding frequency anchor point.

[0012] In an optional embodiment of this application, the coordinates of the key point in the image coordinate system are any of the following: The first method is to directly output the coordinate offset values ​​of the key points; The second approach involves predicting a probability heatmap for each key point and using the probability peak position in the heatmap as the coordinates of that key point.

[0013] In an optional embodiment of this application, the pose estimation network model is a YOLO-Pose network, or a single-stage pose estimation network with an object detection branch and a keypoint regression branch; the pose estimation network model includes a backbone feature extraction network, a feature fusion network, and a decoupled prediction head; the decoupled prediction head is used to output instance category, target region, instance confidence, keypoint coordinates, and keypoint visibility identifier in parallel.

[0014] In an optional embodiment of this application, during the training phase of the pose estimation network model, its loss function includes at least instance target classification loss, instance target bounding box regression loss, keypoint coordinate regression loss, and keypoint visibility loss; wherein, the loss function further includes a prior constraint term on the physical continuity of the dispersion curve, the prior constraint term including at least one of adjacent keypoint velocity difference constraint, curvature smoothing constraint, and inter-order velocity order constraint.

[0015] According to a second aspect of the embodiments of this application, an automatic multi-order dispersion curve picking device based on a pose estimation network is provided, comprising: The input module is used to input the surface wave dispersion energy map or normalized dispersion energy matrix in the frequency-phase velocity domain into the trained attitude estimation network model, and output at least one dispersion curve instance target and its corresponding key point coordinates in the image coordinate system, instance category and instance confidence; wherein, each order dispersion curve in the dispersion energy map is identified as an independent instance target, the instance category is used to characterize the order of the dispersion curve, and the phase velocity value corresponding to the preset frequency anchor point on each order dispersion curve is defined as the key point of the instance target; The validity determination module is used to determine the validity of each dispersion curve instance target based on the instance confidence level and / or key point visibility identifier; when there is no valid dispersion curve instance target with a confidence level higher than a preset threshold, the input surface wave dispersion energy map is determined to be an invalid signal and rejected. The coordinate inverse mapping and output module is used to inversely map the image coordinates of the effective key points into physical coordinates of frequency and phase velocity based on a preset coordinate mapping relationship when there are effective dispersion curve instance targets, and output a multi-order dispersion curve physical coordinate sequence according to the order corresponding to the instance category.

[0016] According to a third aspect of the embodiments of this application, a computer device is provided, comprising: Memory; Processor; and Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the automatic picking method for multi-order dispersion curves based on the attitude estimation network.

[0017] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided having a computer program stored thereon; the computer program is executed by a processor to implement the automatic picking method for multi-order dispersion curves based on attitude estimation networks.

[0018] The automatic picking method for multi-order dispersion curves based on attitude estimation networks provided in this application uses dispersion curves as instance targets and phase velocity values ​​at preset frequency anchor points as key points. It combines instance confidence, key point visibility identification, and physical coordinate inverse mapping to achieve end-to-end output from dispersion energy matrix to physical coordinate sequence of multi-order dispersion curves. This overcomes the problems of low efficiency of manual picking, easy cascading of traditional tracking algorithms, and complex post-processing and difficulty in separating multi-order curves in semantic segmentation methods. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of the automatic picking method for multi-order dispersion curves based on attitude estimation networks provided in the embodiments of this application; Figure 2 This is a structural diagram of the pose estimation network model provided in the embodiments of this application; Figure 3 A flowchart illustrating the application of the automatic picking method for multi-order dispersion curves based on attitude estimation networks provided in this embodiment of the application. Figure 4 This is a structural diagram of an automatic multi-order dispersion curve picking device based on an attitude estimation network provided in an embodiment of this application. Detailed Implementation

[0020] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0021] See Figure 1 This application provides an automatic picking method for multi-order dispersion curves based on attitude estimation networks, including the following steps: S110, input the surface wave dispersion energy map or normalized dispersion energy matrix in the frequency-phase velocity domain into the trained attitude estimation network model, and output at least one dispersion curve instance target and its corresponding key point coordinates in the image coordinate system, instance category and instance confidence; wherein, each order dispersion curve in the dispersion energy map is identified as an independent instance target, the instance category is used to characterize the order of the dispersion curve, and the phase velocity value corresponding to the preset frequency anchor point on each order dispersion curve is defined as the key point of the instance target; S120, based on the instance confidence level and / or key point visibility identifier, the validity of each dispersion curve instance target is determined; when there is no valid dispersion curve instance target with a confidence level higher than the preset threshold, the input surface wave dispersion energy map is determined to be an invalid signal and rejected. S130, when there is an effective dispersion curve instance target, based on the preset coordinate mapping relationship, the image coordinates of the effective key points are inversely mapped to the physical coordinates of frequency and phase velocity, and the multi-order dispersion curve physical coordinate sequence is output according to the order corresponding to the instance category.

[0022] In a specific embodiment, if the image width is W, the height is H, the frequency range is [fmin, fmax], and the phase velocity range is [cmin, cmax], then when the frequency axis adopts a linear mapping, the frequency f = fmin + x / (W-1) × (fmax - fmin); when the frequency axis adopts a logarithmic mapping, log(f) = log(fmin) + x / (W-1) × (log(fmax) - log(fmin)); and the phase velocity c = cmax - y / (H-1) × (cmax - cmin), where x and y are the horizontal and vertical coordinates of the key point in the image coordinate system.

[0023] This application presents an automatic multi-order dispersion curve picking method based on pose estimation networks, overcoming the technical bottleneck of the inability to separate intersecting and adhered multi-order curves, and significantly improving the picking completeness of high-frequency bands. Current existing technologies essentially rely on mathematical statistics of "locally adjacent pixels." When the basic and higher-order curves are extremely close in certain frequency bands, clustering algorithms inevitably classify overlapping areas into the same category, and greedy algorithms may lose their way, leading to order confusion. Existing image semantic segmentation schemes, lacking instance-level recognition capabilities, also output adhered regions as a large connected fusion block, resulting in picking failure. In contrast, this application treats each order curve as an independent instance target. Even with severe adhesion in the image, the model can utilize the full-band physical structure features learned from the global network and, through a keypoint allocation mechanism, independently and accurately regress the node coordinates of the basic and higher-order curves, making the extraction of high-frequency multi-order adhered curves possible.

[0024] This application presents an automatic multi-order dispersion curve picking method based on a pose estimation network, achieving end-to-end direct physical coordinate picking and avoiding the complex post-processing and secondary errors introduced by semantic segmentation methods. Existing semantic segmentation schemes typically output a binary prediction or probability map with a certain pixel width. To obtain the precise coordinates of frequency and phase velocity required for actual inversion and to distinguish different orders, it is often necessary to rely on cumbersome post-processing algorithms such as local extremum search, cluster analysis, and image skeleton extraction. This process not only significantly increases computational costs but also easily causes geometric center shifts or clustering errors during post-processing, thus introducing secondary computational errors. In contrast, this application innovatively adopts a pose estimation architecture, defining dispersion points as keypoints with instance attributes. The network directly regresses and outputs high-precision image coordinates for each node end-to-end and maps them to physical coordinates in one step. This method eliminates the complex post-processing processes such as extremum search, cluster separation, and morphological thinning in the semantic segmentation route, fundamentally ensuring the original accuracy of coordinate extraction and the efficiency of multi-order separation.

[0025] This application presents an automatic multi-order dispersion curve picking method based on attitude estimation networks, which possesses global confidence assessment and automatic rejection capabilities, effectively avoiding the picking of spurious signals in environments with extremely low signal-to-noise ratios. Existing automated algorithms are often coercive. Regardless of how low the signal-to-noise ratio of the input dispersion energy map is, they forcibly search for local energy maximums or color patches in the image and output a spurious curve, resulting in a large amount of poorly denoised noise data being mixed into the subsequent inversion process, relying on extremely high manual screening costs. The attitude estimation model in this application incorporates a confidence evaluation mechanism for target detection, scoring the global topological structure of the curve with confidence before regression. Images that cannot present effective physical features due to noise interference are directly rejected. This intelligent automatic rejection mechanism endows the system with strong anti-interference capabilities and algorithm stability, ensuring the overall quality of surface wave exploration data.

[0026] The automatic picking method for multi-order dispersion curves based on attitude estimation networks proposed in this application breaks through the limitations of traditional local statistical processing and semantic segmentation approaches, and realizes efficient, accurate and fully end-to-end automated coordinate picking of massive surface wave dispersion data, providing a reliable and high-quality data foundation for the refined inversion of subsurface shear wave velocity structure.

[0027] This application presents an automatic multi-order dispersion curve picking method based on a pose estimation network. Addressing the limitations of existing algorithms such as traditional maximum search, filtering clustering, and greedy tracking in noise resistance and the tendency for cascading errors, this application achieves significant technical improvements by introducing a target detection and pose estimation framework. Instead of focusing on local pixel thresholds or connected component partitioning, this application treats the overall spatial trend of the multi-order dispersion curve as the instance recognition target for the deep learning model. This allows the model to adaptively extract full-band physical structure features from the dispersion energy map, accurately tracking and independently identifying the basic and higher-order dispersion curves. This effectively eliminates the cascading errors frequently occurring in traditional clustering or tracking algorithms when multiple energy levels are close together. Furthermore, traditional algorithms, when faced with dispersion spectra with extremely low signal-to-noise ratios, often lack global judgment capabilities and forcibly pick and fit pseudo-curves composed of noise energy. In contrast, this application's method possesses powerful target detection and confidence assessment capabilities. When the target features of the dispersion image are below a preset confidence threshold, the model automatically determines that the image does not contain valid dispersion signals and therefore does not pick it up. This mechanism effectively avoids the introduction of invalid noise data and eliminates the need for tedious secondary manual screening.

[0028] This application presents an automatic multi-order dispersion curve picking method based on a pose estimation network. Addressing the core technical shortcomings of existing semantic segmentation-based deep learning schemes, such as cumbersome post-processing steps and difficulties in separating adhered multi-order curves, this application overcomes these shortcomings through an innovative keypoint direct regression mechanism within the pose estimation network. Instead of outputting binary-class pixel images, this application transforms the phase velocity points corresponding to specific frequencies on the dispersion curve into keypoints directly predicted by the network. This mechanism enables the computer to directly output extremely accurate physical coordinates of each order of dispersion curve end-to-end, completely eliminating the morphological thinning post-processing step, which is prone to introducing secondary calculation errors, thus significantly improving computational efficiency. More importantly, thanks to the pose estimation network's ability to distinguish and regress the inherent structures of different instance targets, even if basic and higher-order dispersion curves exhibit severe spatial adhesion due to energy fusion on the energy map, this application can still accurately identify their inherent independent trajectories and directly output the specific node coordinates of each order curve. This breakthrough provides a practical and reliable approach to solving the technical challenge of separation failure when traditional algorithms encounter adhered curves in the high-frequency band, significantly improving the completeness and accuracy of automatic multi-order dispersion curve picking.

[0029] In an optional embodiment of this application, defining the phase velocity value corresponding to a preset frequency anchor point on each dispersion curve as a key point of the instance target includes: For each dispersion curve, N sampling frequency points are selected along the frequency axis according to linear, logarithmic or piecewise non-uniform frequency anchoring strategies. Each sampling frequency point is defined as a key point by the corresponding phase velocity position on the dispersion energy map. Each instance target has an order category label, and the i-th key point corresponds to the i-th preset frequency anchor point. Each key point also includes a visibility indicator. When the dispersion energy at the corresponding frequency anchor point is lower than the threshold, the curve exceeds the preset velocity range, or the curve of that order is not recognizable, the key point is marked as invisible.

[0030] This application employs linear, logarithmic, or piecewise non-uniform frequency anchor point strategies; it densifies key points in frequency bands with large curvature changes, drastic energy changes, modal cross-adhesion, or sensitivity to picking errors, and sparsely defines key points in frequency bands with relatively smooth trends, thereby improving the accuracy of curve coordinate regression in key frequency bands while ensuring computational efficiency.

[0031] This application breaks away from the traditional approach of clustering tracking and image binary classification semantic segmentation, introducing for the first time a pose estimation architecture from computer vision into the field of surface wave dispersion curve picking. Each order dispersion curve in the dispersion energy map is defined as an independent instance target with a specific topological structure, and the phase velocity coordinates corresponding to a specific frequency on the curve are defined as key points of that instance, thus transforming the picking problem into a key point coordinate regression problem. Utilizing the feature extraction capability of the pose estimation network for independent instances, when there is energy overlap or visual adhesion between basic and higher-order dispersion curves, the model can independently distinguish and regress key points belonging to different orders of curves based on the learned full-band physical structure features, completely solving the "order cascading" and separation difficulties of traditional clustering partitioning and maximum search algorithms.

[0032] In this application, the network model directly regresses and outputs the precise image pixel coordinates of each node of multiple curves, and directly transforms them into a physically meaningful sequence of frequency and phase velocity coordinates through an inverse mapping function. This process completely eliminates the cumbersome and error-prone morphological post-processing steps such as image skeleton thinning that are essential in traditional deep learning segmentation schemes.

[0033] This application introduces target confidence evaluation logic during the inference stage. When the signal-to-noise ratio of the input dispersion energy map is extremely low, causing the effective features extracted by the network to be lower than the preset confidence threshold, the system automatically determines that the image does not contain effective dispersion signals and does not pick it up. This mechanism effectively avoids the technical defects of traditional mathematical extraction algorithms that forcibly fit pseudo-curves in strong noise backgrounds.

[0034] In an optional embodiment of this application, see [link to relevant documentation]. Figure 2 The pose estimation network model is a pose estimation network with an object detection branch and a keypoint regression branch, and it includes at least: A backbone feature extraction network is used to extract deep semantic features from the dispersion energy map; The feature coupling module is used to fuse feature maps of different scales from the backbone network, enhancing the model's ability to perceive weak dispersion features. The prediction head is decoupled for parallel execution of instance target classification, localization, and key point coordinate regression.

[0035] In an optional embodiment of this application, a keypoint detection architecture is used to reconstruct the curves of each order in the dispersion energy map as a pose target with a specific spatial distribution shape and physical topology. The overall technical architecture of this application mainly consists of a Coordinate Topological Mapping (CTM) module, a Multi-scale Feature Coupling (MFC) module, a Decoupled Prediction Head (DPH) module, and a Coordinate Inverse Mapping (CIM) module based on confidence assessment. In the underlying mapping stage of the specific implementation, this application breaks through the limitation of the traditional scheme that treats the dispersion curve as an isolated pixel or local energy block, and creatively defines the dispersion curve as an instance target and a keypoint. This method defines each independent dispersion curve of each order in the energy spectrum as a complete independent instance target, and defines the phase velocity value corresponding to each frequency as the keypoint of the instance according to a preset sampling sequence on the frequency axis. Each instance target carries a category label of its order, and the i-th keypoint corresponds to the i-th preset frequency anchor point. This definition method, based on underlying physical logic, enables neural networks to deeply learn the physical evolution laws and macroscopic continuity constraints of dispersion curves across the entire frequency band. Therefore, in subsequent inference, even in complex regions where multi-order energies are severely overlapped or intersecting in space, the model can accurately trace and distinguish the order of key points based on global topological relationships, fundamentally overcoming the underlying deficiency of traditional clustering algorithms in effectively classifying orders in contiguous regions. In the automatic picking inference implementation stage, the measured dispersion energy map to be processed is first normalized and preprocessed before being input into an improved multi-scale feature coupling network with pre-trained weights. In this stage, the network achieves layer-by-layer feature extraction and mapping through forward propagation. Specifically, the backbone network deeply extracts multi-dimensional semantic features of the energy spectrum, and then performs multi-scale feature fusion through a top-down path aggregation network. Compared to Gaussian filtering or extreme value search operations in existing technologies that can only handle local statistical features, the multi-scale feature fusion of this scheme significantly enhances the model's ability to perceive weak energy signals in high-frequency bands and weak dispersion features masked by strong noise. The deeply fused feature maps are then fed into a decoupled multi-task prediction head, where the network performs classification, localization, and regression inference simultaneously and in parallel. The model not only predicts the category and spatial bounding box of each dispersed instance target, but more importantly, it directly regresses the high-precision position coordinates of each node in the pixel coordinate system end-to-end through keypoint branches.This end-to-end direct output mechanism completely eliminates the cumbersome post-processing procedures such as image skeleton thinning and secondary search for local maxima required by existing semantic segmentation schemes, and thoroughly eliminates the geometric center offset and secondary calculation errors caused by morphological thinning algorithms. In the final inference and automatic screening stage, this application introduces an intelligent rejection mechanism based on confidence scoring. The model outputs a global confidence score for each predicted candidate dispersion curve. When the score is lower than a preset safety threshold, the system determines that the image has lost its physical analysis value due to excessive background noise, and automatically rejects the picking operation. This mechanism effectively avoids the technical blind spot of traditional forward-backward greedy tracking algorithms, which are prone to forcibly tracking noise pseudo-signals in low signal-to-noise ratio environments, and ensures the purity of the inversion data from the source. For dispersion targets that are determined to be valid by confidence, the system will directly restore the pixel coordinates output by the network to the real coordinates of frequency and phase velocity in physical space according to the preset linear inverse mapping operator, and finally output a continuous and high-precision dispersion curve physical coordinate sequence file classified by order.

[0036] In an optional embodiment of this application, the instance confidence is obtained by the attitude estimation network model evaluating and scoring the global topology and local energy characteristics of the dispersion curve instance target, and the key point visibility identifier is used to characterize whether there is an effective phase velocity point that can be used for output at the corresponding frequency anchor point.

[0037] In an optional embodiment of this application, the coordinates of the key point in the image coordinate system are any of the following: The first method is to directly output the coordinate offset values ​​of the key points; The second approach involves predicting a probability heatmap for each key point and using the probability peak position in the heatmap as the coordinates of that key point.

[0038] In an optional embodiment of this application, the pose estimation network model is a YOLO-Pose network, or a single-stage pose estimation network with an object detection branch and a keypoint regression branch; the pose estimation network model includes a backbone feature extraction network, a feature fusion network, and a decoupled prediction head; the decoupled prediction head is used to output instance category, target region, instance confidence, keypoint coordinates, and keypoint visibility identifier in parallel.

[0039] In this application, a single-stage pose estimation network with target detection branch and key point regression branch is adopted, which can simultaneously complete dispersion curve instance localization, order category identification, key point coordinate regression and key point visibility determination in one forward propagation process, thereby specifically improving the efficiency of automatic picking of multi-order dispersion curves and the accuracy of coordinate localization.

[0040] In an optional embodiment of this application, during the training phase of the pose estimation network model, its loss function includes at least: instance target classification loss, instance target bounding box regression loss, keypoint coordinate regression loss, and keypoint visibility loss; wherein, the loss function further includes a prior constraint term on the physical continuity of the dispersion curve, the prior constraint term including at least one of adjacent keypoint velocity difference constraint, curvature smoothing constraint, and inter-order velocity order constraint.

[0041] In an optional embodiment of this application, see [link to relevant documentation]. Figure 3 The specific working sequence logic of the automatic picking method for multi-order dispersion curves based on attitude estimation network in this application is as follows: S101, input the measured surface wave dispersion energy matrix; S102, perform normalization, frequency range and phase velocity range setting, and necessary channel copying processing; S103, send the data into the attitude estimation network for feature extraction, wherein the attitude estimation network is preferably the YOLO-Pose network; S104, the network outputs instance category, instance confidence, target area, key point coordinates, and key point visibility identifier in parallel; S105, filter the instance confidence and key point visibility; S106, trigger automatic rejection when there is no valid curve instance; S107, extract candidate key point pixel coordinates when there is a valid curve instance; S108, apply the coordinate inverse mapping operator to restore it to frequency and phase velocity physical coordinates; S109, apply physical constraints such as smoothing, velocity continuity, and inter-order velocity order, and output continuous multi-order dispersion curves according to order.

[0042] This application achieves end-to-end direct mapping from raw energy maps to multi-order physical coordinates by leaping from local pixel statistics to global instance perception. This method not only effectively overcomes the technical bottleneck of multi-order dispersion curves easily becoming misaligned or lost in overlapping regions, but also significantly improves the efficiency of automated processing and the accuracy of raw extraction for massive surface wave exploration data by eliminating cumbersome two-step post-processing steps, thus meeting the practical needs of modern high-precision geophysical exploration.

[0043] See Figure 4 This application provides an automatic multi-order dispersion curve picking device based on a pose estimation network, comprising the following modules: The input module 410 is used to input the surface wave dispersion energy map or normalized dispersion energy matrix in the frequency-phase velocity domain into the trained attitude estimation network model, and output at least one dispersion curve instance target and its corresponding key point coordinates in the image coordinate system, instance category and instance confidence; wherein, each order dispersion curve in the dispersion energy map is identified as an independent instance target, the instance category is used to characterize the order of the dispersion curve, and the phase velocity value corresponding to the preset frequency anchor point on each order dispersion curve is defined as the key point of the instance target; The validity determination module 420 is used to determine the validity of each dispersion curve instance target based on the instance confidence level and / or key point visibility identifier; when there is no valid dispersion curve instance target with a confidence level higher than a preset threshold, the input surface wave dispersion energy map is determined to be an invalid signal and rejected. The coordinate inverse mapping and output module 430 is used to, when there is an effective dispersion curve instance target, inversely map the image coordinates of the effective key points into physical coordinates of frequency and phase velocity based on a preset coordinate mapping relationship, and output a multi-order dispersion curve physical coordinate sequence according to the order corresponding to the instance category.

[0044] This application provides a computer device, including: Memory; Processor; and Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the automatic picking method for multi-order dispersion curves based on the attitude estimation network.

[0045] This application provides a computer-readable storage medium storing a computer program thereon; the computer program is executed by a processor to implement the automatic picking method for multi-order dispersion curves based on attitude estimation networks.

[0046] The device and storage medium provided in this embodiment have the same technical effects as the methods described above.

[0047] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as C, VHDL, Verilog, the object-oriented programming language Java, and the interpreted scripting language JavaScript.

[0048] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing 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 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0052] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0053] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0054] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0055] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for automatically picking multi-order dispersion curves based on attitude estimation networks, characterized in that, Includes the following steps: The surface wave dispersion energy map or normalized dispersion energy matrix in the frequency-phase velocity domain is input into the trained attitude estimation network model, and the model outputs at least one dispersion curve instance target and its corresponding key point coordinates in the image coordinate system, instance category and instance confidence; wherein, each order dispersion curve in the dispersion energy map is identified as an independent instance target, the instance category is used to characterize the order of the dispersion curve, and the phase velocity value corresponding to the preset frequency anchor point on each order dispersion curve is defined as the key point of the instance target; Based on the instance confidence level and / or key point visibility identifier, the validity of each dispersion curve instance target is determined; when there is no valid dispersion curve instance target with a confidence level higher than the preset threshold, the input surface wave dispersion energy map is determined to be an invalid signal and rejected. When there are valid dispersion curve instances, the image coordinates of the valid key points are inversely mapped to physical coordinates of frequency and phase velocity based on a preset coordinate mapping relationship, and a multi-order dispersion curve physical coordinate sequence is output according to the order corresponding to the instance category.

2. The method according to claim 1, characterized in that, The step of defining the phase velocity value corresponding to the preset frequency anchor point on each dispersion curve as the key point of the instance target includes: For each dispersion curve, N sampling frequency points are selected along the frequency axis according to linear, logarithmic or piecewise non-uniform frequency anchoring strategies. Each sampling frequency point is defined as a key point by the corresponding phase velocity position on the dispersion energy map. Each instance target has an order category label, and the i-th key point corresponds to the i-th preset frequency anchor point. Each key point also includes a visibility indicator. When the dispersion energy at the corresponding frequency anchor point is lower than the threshold, the curve exceeds the preset velocity range, or the curve of that order is not recognizable, the key point is marked as invisible.

3. The method according to claim 1 or 2, characterized in that, The pose estimation network model is a pose estimation network with an object detection branch and a keypoint regression branch, and it includes at least: A backbone feature extraction network is used to extract deep semantic features from the dispersion energy map; The feature coupling module is used to fuse feature maps of different scales from the backbone network, enhancing the model's ability to perceive weak dispersion features. The prediction head is decoupled for parallel execution of instance target classification, localization, and key point coordinate regression.

4. The method according to claim 1, characterized in that, The instance confidence score is obtained by evaluating and scoring the global topology and local energy characteristics of the dispersion curve instance target by the attitude estimation network model. The key point visibility marker is used to characterize whether there is an effective phase velocity point that can be used for output at the corresponding frequency anchor point.

5. The method according to claim 1, characterized in that, The coordinates of the key point in the image coordinate system are any of the following: The first method is to directly output the coordinate offset values ​​of the key points; The second approach involves predicting a probability heatmap for each key point and using the probability peak position in the heatmap as the coordinates of that key point.

6. The method according to claim 1, characterized in that, The pose estimation network model is the YOLO-Pose network, or a single-stage pose estimation network with an object detection branch and a keypoint regression branch; the pose estimation network model includes a backbone feature extraction network, a feature fusion network, and a decoupled prediction head; the decoupled prediction head is used to output instance category, target region, instance confidence, keypoint coordinates, and keypoint visibility identifier in parallel.

7. The method according to claim 1, characterized in that, During the training phase of the pose estimation network model, its loss function includes at least instance target classification loss, instance target localization box regression loss, key point coordinate regression loss, and key point visibility loss; wherein, the loss function also includes a prior constraint term on the physical continuity of the dispersion curve, the prior constraint term including at least one of adjacent key point velocity difference constraint, curvature smoothing constraint, and inter-order velocity order constraint.

8. An automatic picking device for multi-order dispersion curves based on attitude estimation networks, characterized in that, include: The input module is used to input the surface wave dispersion energy map or normalized dispersion energy matrix in the frequency-phase velocity domain into the trained attitude estimation network model, and output at least one dispersion curve instance target and its corresponding key point coordinates in the image coordinate system, instance category and instance confidence; wherein, each order dispersion curve in the dispersion energy map is identified as an independent instance target, the instance category is used to characterize the order of the dispersion curve, and the phase velocity value corresponding to the preset frequency anchor point on each order dispersion curve is defined as the key point of the instance target; The validity determination module is used to determine the validity of each dispersion curve instance target based on the instance confidence level and / or key point visibility identifier; when there is no valid dispersion curve instance target with a confidence level higher than a preset threshold, the input surface wave dispersion energy map is determined to be an invalid signal and rejected. The coordinate inverse mapping and output module is used to inversely map the image coordinates of the effective key points into physical coordinates of frequency and phase velocity based on a preset coordinate mapping relationship when there are effective dispersion curve instance targets, and output a multi-order dispersion curve physical coordinate sequence according to the order corresponding to the instance category.

9. A computer device, characterized in that, include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the automatic picking method for multi-order dispersion curves based on attitude estimation networks as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program; the computer program is executed by a processor to implement the automatic picking method for multi-order dispersion curves based on attitude estimation networks as described in any one of claims 1 to 7.

Images