A method for adaptive hierarchical threshold generation oriented to boundary preservation

Abstract

The application discloses a boundary-keeping-oriented adaptive hierarchical threshold generation method, relates to the technical field of underwater acoustic detection and signal processing, and is used for generating an adaptive hierarchical threshold. The method comprises the following steps: taking a detected original seabed line as input, combining a reference boundary ridge line constraint, constructing suppression demand and protection demand in a wavelet domain for different scale detail layers respectively, generating a candidate threshold according to the competitive relationship between the suppression demand and the protection demand, and obtaining a final seabed line by means of boundary feasibility screening and smoothing quality sorting, so that the final seabed line can give consideration to boundary fitting, abnormality suppression and continuity recovery capability. While guaranteeing effective suppression of noise and abnormal jump points, the application significantly improves the boundary keeping capability and the continuity recovery capability of the seabed line, and is suitable for complex underwater environment and fine seabed topography detection scenes.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic detection and signal processing technology, and more particularly to the field of post-processing technology for side-scan sonar seabed echo data, specifically referring to an adaptive hierarchical threshold generation method oriented towards boundary preservation. Background Technology

[0002] In actual side-scan sonar detection environments, the original seabed line detection results often contain numerous anomalous jumps due to factors such as underwater acoustic propagation multipath effects, environmental noise, bubble disturbances, biological activity, and system electronic noise. To obtain a smooth and continuous seabed line, existing technologies generally employ filtering or smoothing methods to process the original detection results. Among these, wavelet transform-based multi-scale decomposition and thresholding methods have attracted attention because they can simultaneously handle different frequency band components. In wavelet domain seabed line smoothing, the selection of the threshold is a crucial step. Traditional methods typically use a globally uniform threshold, applying the same threshold value to detail coefficients at different wavelet scale layers or different spatial locations for reduction or zeroing. However, the actual seabed line exhibits local variability in morphology: in flat seabed areas, the boundary trend is gentle, requiring high suppression of noise and small fluctuations; while in locations with real topographic relief, reef edges, gully bends, or target boundaries, local curvature changes and abrupt detail changes in the seabed line are critical boundary features that need protection.

[0003] In the smoothing of the seabed line of side-scan sonar, a uniform threshold can easily lead to two types of problems: one is that the threshold is too small, resulting in insufficient suppression of high-frequency noise, bubble interference and local jump points; the other is that the threshold is too large, resulting in excessive weakening of the true boundary ridge, key turning point and overall trend, thereby causing boundary drift, turning point blunting and a decrease in overall edge-fitting ability.

[0004] Therefore, there is an urgent need for a method that can adaptively generate differentiated thresholds based on the local morphological features of the seabed line and reference boundary information, so as to balance the effective suppression of abnormal jump points with the preservation of the integrity of the true boundary features. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive hierarchical threshold generation method for boundary preservation, so as to solve the problems of noise residue and boundary drift caused by uniform threshold in the prior art, and achieve adaptive boundary preservation smoothness.

[0006] To achieve the above objectives, the present invention provides an adaptive hierarchical threshold generation method for boundary preservation, characterized by comprising: S1. Read and parse the raw side-scan sonar data containing port and starboard echo sampling sequences, construct port and starboard waterfall diagram matrices, enhance and optimize the matrices, and generate a detection enhancement matrix; S2. Perform row-by-row boundary tracking on the detection enhancement matrix, extract the original seabed line sequence, and perform light pre-cleaning on the original seabed line sequence through interpolation repair to obtain the pre-cleaned seabed line sequence. S3. Based on the original seabed line sequence, extract reference boundary ridges in the neighborhood of the original seabed line, and construct constraint quantities for the reference boundary ridges. The constraint quantities include boundary existence degree, boundary locking degree, and key turning point importance. S4. Perform multi-scale discrete wavelet decomposition on the pre-cleaned seabed line sequence to obtain approximate components and detail components at different scales, and calculate the basic threshold of the detail components at each scale. S5. Calculate the suppression correlation and protection correlation of detail components at each scale layer, normalize and weight the suppression correlation and protection correlation respectively, construct the suppression demand driver and basic protection demand driver at each scale layer, correct the basic protection demand driver at each scale layer through approximation components, and construct the protection demand driver at each scale layer. S6. Construct a balanced score based on the suppression demand drive and the protection demand drive, generate multiple sets of candidate threshold offsets, and perform inter-layer consistency correction on each set of candidate threshold offsets. Calculate the candidate threshold weights based on the corrected threshold offsets, and process the basic thresholds in S4 using the candidate threshold weights to obtain the stratified thresholds.

[0007] Original submarine line sequence for: ; In the formula, Indicates the sequence number of the ping sequence. Indicates the first The original seabed line position values ​​corresponding to each ping sequence. Indicates the total length of the submarine line; Anomaly queries were performed on the original seabed line sequence. Ping sequences exhibiting both abnormal trend deviations and anomalous jumps were marked as pre-cleaned anomalies. The original seabed line position values ​​of these pre-cleaned anomalies were then repaired using interpolation to obtain the pre-cleaned seabed line sequence. : ; In the formula, Indicates the first time after cleaning The original submarine line position value corresponding to each ping sequence.

[0008] The reference boundary ridge line is: ; In the formula, Indicates the reference boundary ridge line. Indicates the ridge line corresponding to the first Sampling points of a ping sequence; The existence degree of the boundary is: ; In the formula, This indicates that there is a constraint vector at the boundary. for The corresponding number in the middle The components of a ping sequence; Boundary locking degree is: ; In the formula, Represents the boundary locking constraint vector. for The corresponding number in the middle The components of a ping sequence; The importance of key turning points is as follows: ; In the formula, This represents the vector of importance constraints for transitions. for The corresponding number in the middle The components of a ping sequence.

[0009] The basic threshold for detail components at each scale layer is: ; ; In the formula, Indicates the first The base threshold for the detail components of the i-th scale layer is equivalent to the i-th scale layer. The base threshold for layer detail components, Indicates the first The length of the layer detail factor, Let e ​​represent the logarithmic function with the numerical constant e as the base. Indicates the first The noise standard deviation of the detail components at each scale layer. This indicates the calculation of the median absolute deviation. Indicates the first The detail components of each scale layer.

[0010] Suppression-related quantities include noise dominance, cross-scale persistence, and local volatility; protection-related quantities include boundary presence, boundary locking, turning point protection, and trend risk. Normalize and weight the inhibition-related quantities to construct the inhibition demand driver: ; ; In the formula, Indicates the scale layer index. Indicates the total number of scale layers. Indicates the first The need to suppress layer detail components, , , They represent the normalized values ​​respectively. , Indicates the first Noise dominance of layer detail components, Indicates the first The non-discontinuity of layer detail components, Indicates the first Cross-scale persistence of layer detail components, Indicates the first Local fluctuations in layer detail components, , , They represent , , The weights; Normalize and weight the protection-related quantities to calculate the basic protection demand drivers: ; In the formula, Indicates the first Basic protection requirements for layer detail components , , , They represent the normalized values ​​respectively. , , , , Indicates the first The existence degree of the boundary of the layer detail component. Indicates the first Boundary locking degree of layer detail components, Indicates the first The degree of protection of the transition of layer detail components. Indicates the first The trend risk of layer detail components, , , , They represent , , , The weights; Introducing coarse-layer prior protection factors right Make corrections to construct protection requirements-driven mechanisms: ; In the formula, Indicates the first The need for protection of layer detail components, To protect the requirements, a correction function is used.

[0011] The noise dominance is: ; ; In the formula, Indicates the first Noise variance of layer detail components, Here is the regularization constant. For the first Energy of layer detail components, Indicates the position index of the detail component. Indicates the first The total number of sampling points for the layer detail components. Indicating the first detail component Detail coefficients for each sampling point; Cross-scale persistence is: ; ; ; In the formula, For continuously weighted mixing parameters, No. The fundamental cross-scale persistence of layer detail components. Indicates the first Weighted cross-scale persistence of layer detail components, Indicates the coarse-scale amplitude. Indicates the first The first layer of detail component The locking weight of each sampling point; Local volatility is: ; In the formula, express The first-order difference.

[0012] The detail components of each scale layer are reconstructed to the original seabed line length, the detail reconstruction components of each scale layer are obtained, and a normalized amplitude distribution is constructed: ; In the formula, Indicates the first Layer detail reconstruction component corresponds to the first The normalized magnitude of a ping sequence. Indicates the first Layer detail reconstruction component corresponds to the first The reconstruction coefficients of each ping sequence, based on Calculate protection-related quantities; The existence degree of the boundary is: ; Boundary locking degree is: ; The transition protection degree is: ; The trend risk is: ; In the formula, For trend risk function, Indicates the first Layer detail reconstruction components.

[0013] The balance fraction is: ; In the formula, Indicates the first The balance value of layer detail components; right Implement robust centralization: ; In the formula, for The central value, This indicates the value of the median.

[0014] Multiple candidate threshold schemes are generated using a candidate configuration set, which is: : ; In the formula, Indicates the first Group candidate configuration; right Constructing a clipping response: ; In the formula, Indicates the first The response value of the group candidate configuration. for The fractional cropping scale, This is a numerical clipping function; based on Generate candidate threshold offsets: ; In the formula, This indicates the allowable offset in the threshold direction. Indicates the allowable offset in the protection direction; Perform inter-layer consistency correction on the candidate threshold offset to obtain the corrected threshold offset: ; In the formula, Indicates the corrected number Threshold offset for group candidate configurations. Indicates degrees of freedom. Indicates consistency. This is a consistency correction function.

[0015] definition Upper and lower bound constraints are used to calculate candidate threshold weights: ; ; ; In the formula, Indicates the first Candidate threshold weights for group candidate configurations. This indicates the lower limit of the candidate threshold. Indicates the upper limit of the candidate threshold. Let be the lower bound constraint function. For upper limit constraint functions; The stratification threshold is: ; In the formula, Indicates the first Layer detail components in the first Hierarchical threshold under group candidate configuration.

[0016] Compared with the prior art, the present invention has the following advantages: This invention can distinguish between noise components that should be suppressed and boundary components that should be protected in layers of different scales, avoiding the "one-size-fits-all" problem caused by a uniform threshold. This invention establishes boundary preservation constraints by referencing ridge lines, so that seabed line processing no longer only pursues smoothness, but prioritizes ensuring the fit of the true boundary main ridge. This invention allows different threshold directions and threshold amplitudes for layers of different scales, where coarser layers can actively reduce the threshold when protection needs prevail, thereby mitigating overall trend drift. This invention achieves "preserving the boundary first, then considering smoothness" through candidate scheme screening, fundamentally avoiding the "the smoother, the more biased" problem that occurs in traditional single-objective optimization. This invention can maintain key turning points and main boundary structures while suppressing high-frequency fluctuations, local anomalies, and bubble interference, improving the stability and interpretability of the final seabed line. Attached Figure Description

[0017] Figure 1The overall flowchart of the adaptive hierarchical threshold generation method provided by the present invention; Figure 2 This is a schematic diagram of a partial reference ridge line provided by the present invention; Figure 3 This invention provides a schematic diagram illustrating the competition between suppression and protection requirements in generating a threshold. Figure 4 A partial comparison diagram of the original seabed line, the wavelet-smoothed seabed line, and the final seabed line provided by this invention; Figure 5 The diagram shows the processing results of the adaptive layering threshold provided by this invention in a real side-scan sonar. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] Example 1; like Figure 1 The method for generating an adaptive hierarchical threshold with boundary preservation, as shown, includes the following steps; Step 1: Obtain side-scan sonar waterfall image data, construct port and starboard waterfall image matrices and enhance the matrices to generate a detection enhancement matrix; The process involves acquiring the original side-scan sonar file (XTF file) containing the port and starboard echo sampling sequences, and reading the original side-scan sonar data, including the port and starboard echo sampling sequences, echo records corresponding to continuous pings (detection cycles), and metadata such as header information, time information, and channel information. The original side-scan sonar data is parsed to obtain the port and starboard matrices, and a port and starboard waterfall plot matrix is ​​constructed. The rows of the matrix correspond to the ping sequence, equivalent to a complete acoustic wave transmission and reception sequence, while the columns correspond to the time / distance-arranged "sampling points" within a ping (one detection cycle). Based on actual needs, the port and starboard waterfall plot matrices are enhanced and optimized using methods such as logarithmic compression and row-by-row robust normalization to generate a detection enhancement matrix, which is used to improve boundary detection separability. The purpose of detection enhancement is to improve local separability near the seabed boundary, not to directly serve as the final display image. Therefore, the detection enhancement matrix is ​​used for original seabed line extraction, while the port and starboard waterfall plot matrices are still used for subsequent boundary evaluation and result display.

[0020] The waterfall image is the most intuitive display of side-scan sonar echo data. Each time the sonar transmits and receives an acoustic pulse, called a Ping (detection cycle), the system records the echo intensity sampling sequence arranged by time / distance within this cycle. By stitching together each Ping data point from the channel vertically (from top to bottom or bottom to top), a two-dimensional terrain image is formed. As the sonar moves forward, it continuously transmits, receives, and processes echo signals, converting the echo intensity information of each Ping into grayscale information, thus forming an echo image sequence. Stitching together the echo image sequences from the port and starboard sides creates the side-scan sonar waterfall image. The port and starboard waterfall image matrices are two independent two-dimensional numerical arrays. The rows of the matrix correspond to the continuous Ping sequence, and the columns correspond to the sampling points arranged by time / distance within each Ping.

[0021] The enhancement optimization generates a "detection enhancement matrix" to maximize the numerical separability of specific features (seabed boundaries) in the image from background noise. This includes at least logarithmic compression and row-wise robust normalization, and distance attenuation compensation for sampling points can be applied to maintain relative consistency in the echo intensity of the effective seabed across the entire row.

[0022] Step 2: Extract the original seabed line and perform a light pre-cleaning on the original seabed line; Ping-by-ping boundary tracing is performed on the enhanced waterfall map to obtain the original seabed line sequence: ; In the formula, This represents the original coastline sequence. This indicates that the ping sequence number is equivalent to the probe cycle number, serving as the sampling index. Indicates the first The original seabed line position values ​​corresponding to each ping sequence. This represents the total length of the seabed line, which is equivalent to the total number of ping sequences (number of rows in the matrix) in the port and starboard waterfall diagram matrix.

[0023] The original seabed line is obtained by detecting the seabed with a side-scan sonar device, followed by a series of signal processing and data conversions. It reflects the undulating features of the seabed topography. The detection result is obtained through initial seed localization, row-by-row candidate boundary search, short-term false boundary suppression, and trajectory switching protection. The extraction process of the original seabed line includes estimating the initial boundary seed based on the global echo profile, which is used for tracking the first row of the matrix. Starting from the second row, the current row search window is set according to the tracking result of the previous row, and the seabed boundary is tracked row by row, ping by ping. During row-by-row tracking, a short-term false boundary suppression mechanism and a trajectory switching protection mechanism are introduced to extract several candidate boundary positions in the current row search window. These candidates are then ranked based on local contrast, backside intensity, and relative anchor point position. The value directly corresponds to the first value in the port and starboard waterfall diagram matrix during the matrix construction process. The data is processed row by row, and the seabed boundary is tracked ping-by-ping to form a continuous spatial distribution of the seabed line. A short-term false boundary suppression mechanism is introduced to prevent burst-type false boundaries from directly taking over, and a trajectory switching protection mechanism is introduced to prevent continuous bubble bands from causing the entire track to change. Before the original seabed line is extracted, only basic enhancement and trajectory protection can be enabled, without enabling subsequent experimental modules, so as to obtain a more stable original seabed line input. Row-by-row boundary tracking (ping-by-ping boundary tracking) is a typical algorithm flow combining image processing and pattern recognition. The core logic is to traverse each row of the "detection enhancement matrix" and find the point of maximum gradient change at the intersection of water (weak echo) and seabed (strong echo).

[0024] To reduce the impact of extreme jump points on multi-scale decomposition, the original seabed line was modified. Lightweight pre-cleaning is performed, including identifying local anomalies and bridging anomalies in the original seabed line and generating a pre-cleaned seabed line through interpolation repair. This reduces the interference of extreme anomalies on subsequent multi-scale decomposition while maintaining the overall boundary structure from being disrupted by strong smoothing. Anomalies in the original seabed line sequence caused by factors such as bubble interference, echo mutations, and local mistracking typically exhibit two characteristics: a significant deviation from the local seabed line trend and a sudden jump relative to adjacent ping sequences. Therefore, two criteria are introduced: trend residual and adjacent jump. Only when both exceed their corresponding thresholds is the location marked as a pre-cleaned anomaly, avoiding misjudging real terrain transitions or abrupt boundary changes as anomalies.

[0025] Let the first The local median trend of the ping sequence is In the Calculate the local median trend within a neighborhood window of each ping sequence: ; In the formula, Let be the window radius. Preferably... Then the length of the neighborhood window is ; The trend residual is defined as: ; In the formula, Indicates the first Trend residuals of each ping sequence; Adjacent transitions are defined as: ; In the formula, Indicates the first The adjacent jumps of a ping sequence relative to the previous ping sequence characterize the degree of drastic change in seabed position between adjacent ping sequences. At that time, it can be defined Its value is zero or it is not included in the judgment.

[0026] When the absolute value of the trend residual exceeds the preset trend threshold, an abnormal trend deviation is determined to have occurred at the current position (at the current ping sequence). At that time, the judgment of the first (If an abnormal trend deviation occurs at a ping position), when the absolute value of an adjacent jump exceeds a preset jump threshold, an abnormal jump is determined to have occurred at the current position (at the current ping sequence). At that time, the judgment of the first When an abnormal jump occurs at a ping location, and a location simultaneously exhibits both abnormal trend deviation and abnormal jump, that location is marked as a pre-cleaning anomaly point, with a preset trend threshold. and jump preset threshold It is automatically calculated based on the current original submarine line sequence.

[0027] Trend preset threshold for: ; ; ; In the formula, Represents the robustness scale of the trend residual series. Represents the trend residual sequence. This represents the total number of ping sequences that exhibit abnormal trend deviations. Sensitivity coefficient calculated for the trend; To improve the ability to identify isolated jump points, adjacent jumps are defined using bilateral adjacent jumps: ; when At that time, the judgment of the first If an abnormal jump occurs at a ping position, calculate the adjacent jump sequence: ; Calculate the robustness scale of adjacent jump sequences : ; Calculate the preset threshold for the jump : ; In the formula, This is the jump sensitivity coefficient.

[0028] and This is the sensitivity coefficient, used to control the stringency of outlier detection. The value of the sensitivity coefficient is based on the robust three-times scaling criterion, and is adjusted to broaden its applicability. and It can be set in the range of [2.5, 4.0], preferably. =3.0, =3.0, this setting can prevent a small number of normal terrain transitions from being misjudged as outliers. An abnormal trend deviation or abnormal jump is only identified when the trend residual or adjacent jump exceeds the median level by about three times the robust scale.

[0029] If only the condition is satisfied Not satisfied If the condition is met, it indicates that the point may belong to continuous terrain undulations or a true boundary transition, and is not directly identified as an outlier; if only the condition is met... Not satisfied If the condition is met, it indicates that the point may only be a rapidly changing local boundary, and it is not directly classified as an outlier. Only when both conditions are met... and Only when the location deviates from the local trend and exhibits a sudden jump is it considered a pre-cleaning anomaly point. , The two preset thresholds are automatically calculated based on the median and median absolute deviation of the current seabed line sequence, and are equivalent to the final... and The data is automatically determined from the input data, which can both eliminate extreme jump points and reduce the risk of mistakenly deleting real seabed boundary transitions. Here, 1.4826 is the consistency correction coefficient for the median absolute deviation (MAD). For approximately normally distributed data, MAD≈0.6745. , Indicates the robustness scale. ≈MAD / 0.6745≈1.4826MAD. Using this coefficient, MAD obtained based on the median can be converted into a robust noise estimate scaled with the standard deviation.

[0030] when , , When, jump to the preset threshold and trend preset threshold for: ; ; Finally, the set of pre-cleaned anomalies. Defined as: ; ; when At that time, the first The locations of the seabed lines corresponding to each ping sequence are marked as pre-cleaned outliers. These outliers are then repaired using linear interpolation, spline interpolation, or neighborhood valid point interpolation to obtain the pre-cleaned seabed line sequence. ; In the formula, This indicates a pre-cleaned submarine line sequence. Indicates the first time after cleaning The original seabed line position values ​​corresponding to each ping sequence, and the anomaly points. The values ​​obtained after repair through interpolation (such as linear interpolation or spline interpolation) are non-outlier values. Keep the original value, which is equivalent to a non-outlier. The original seabed line was lightly pre-cleaned to eliminate a very small number of abnormal points that would interfere with subsequent decomposition, but the original seabed line was not smoothed as a whole.

[0031] Step 3: Based on the original seabed line sequence, extract the reference boundary ridge line in the neighborhood of the original seabed line, and construct constraint quantities for the reference boundary ridge line. Based on the original seabed line sequence, a reference boundary ridge is extracted in the neighborhood of the original seabed line. Constraints such as boundary existence degree, boundary locking degree, and key turning point importance are constructed around the reference ridge to characterize whether there is a stable boundary response at the current position, whether it is close to the main boundary ridge, and whether it belongs to a key turning point region.

[0032] Using the original seabed line location, local boundary response, and turning point importance as inputs, this algorithm fuses the three inputs within the neighborhood of the original seabed line using existing weighted averaging algorithms (such as linear weighted average and geometric weighted average). Based on the local boundary response and turning point importance, the original seabed line is optimized to construct a reference boundary ridge. ; In the formula, Indicates the reference boundary ridge line. Indicates the ridge line corresponding to the first Sampling points of a ping sequence.

[0033] The original seabed line is a curve initially extracted using simple methods (such as amplitude thresholding and envelope peaking). Due to noise, reverberation, or drastic topographic changes, it may not be smooth enough or contain biases. Based on the original seabed line sequence, a reference boundary ridge is extracted within the neighborhood of the original seabed line. That is, for each ping sequence (a sampling point along the survey line direction), a search band (neighborhood) is defined within a certain range above and below (or vertically) the original seabed line position value. The width of this search band is typically from tens of centimeters to several meters, depending on water depth, sonar resolution, and expected seabed undulations. Within the neighborhood of the original seabed line, based on the original seabed line position, local boundary response, and turning point importance of each ping sequence, an optimal ridge point is selected or calculated for each ping sequence. The local boundary response includes boundary strength, response concentration, and boundary locking constraint. Boundary strength is extracted from the original seabed line sequence using an edge detection operator. Response concentration is calculated using directional adaptive density clustering, dynamically adjusting the shape and orientation (narrow band) of the candidate point's search neighborhood, and counting the number of edge feature points falling within this area; a higher number indicates greater clustering. Boundary locking constraint is used to evaluate the boundary's locking state, calculated by assessing the symmetry between the candidate point and the track line, the asymmetry of changes between adjacent points, and the consistency of trends among adjacent points along the track direction. Turning point importance includes curvature and priority protection weights, obtained from the local bending and curvature information of the original seabed line. This includes calculating the curvature of the curve for each ping sequence on the original seabed line, equivalent to the curvature value of each ping sequence, and mapping the curvature value to protection weights as algorithm input. A larger curvature indicates a more dramatic topographic turning point at that point.

[0034] Simultaneously, three constraint vectors are constructed around the reference boundary ridge. Boundary constraint vectors exist for: ; in, for The corresponding number in the middle The components of the ping sequence are used to measure the ping sequence. Does a stable boundary response exist near a given location? A larger value indicates a more stable boundary. Boundary locking constraint vector for: ; In the formula, for The corresponding number in the middle The components of the ping sequence are used to describe the ping sequence. The degree of fit between a location and the main ridge of the boundary; the larger the value, the closer the location is to the main ridge. The transition importance constraint vector is: ; In the formula, for The corresponding number in the middle The components of the ping sequence are used to describe the ping sequence. Whether a position belongs to a critical turning point, the larger the value, the more important the geometric turning point.

[0035] The aforementioned constraints can be automatically generated based on boundary response strength, local concentration, geometrical inflection degree, and local alignment degree, and normalized to a uniform numerical range. The reference ridge and its constraints are subsequently used both to construct protection requirements and to screen the boundary feasibility of candidate schemes.

[0036] Step 4: Perform multi-scale discrete wavelet decomposition on the pre-cleaned seabed line to obtain approximate components and detail components at different scales. For each detail component, calculate the basic threshold of that detail component, which is used as the initial amount for subsequent adaptive threshold correction.

[0037] Pre-cleaned submarine line Performing multi-scale discrete wavelet decomposition yields: ; In the formula, This is the coarsest approximate component. Indicates the total number of scale layers. For the first Layer detail components, For scale layer indexing. This represents the multi-scale discrete wavelet transform.

[0038] Coarsest layer approximate component Used to obtain Driven by the need for improved protection, the data output process is as follows: ; in, For reconstruction The obtained overall trend sequence, Indicates the first The consistency of trend between the detailed reconstruction components at each layer and the overall trend sequence. It is a coarse-layer a priori protection factor.

[0039] For each layer of detail components First, calculate the robust noise scale: ; in, Indicates the first The noise standard deviation of the layer detail component, This indicates the calculation of the median absolute deviation.

[0040] The base threshold for this layer is redefined as follows: ; in, Indicates the first The length of the layer detail component, This represents a logarithmic function with the digital constant e as the base. The base threshold for each detail component represents only the initial threshold of that layer at a normal noise scale, not the final threshold.

[0041] Step 5: Calculate the suppression correlation and protection correlation of detail components at each scale layer, normalize and weight the suppression correlation and protection correlation respectively, and construct suppression demand-driven and protection demand-driven models. Constructing the suppression demand driver involves calculating suppression-related quantities for each detailed component of the layer, including noise dominance, cross-scale persistence, and local volatility, to characterize whether the layer is more biased towards noise or efficient structure; then normalizing and weighting the suppression-related quantities to form the suppression demand driver for that layer. The construction of protection demand drivers involves reconstructing the detailed components at each scale to the original seabed line length, obtaining the reconstructed components at each scale, coupling each detailed reconstructed component with the relevant constraint quantities of the reference ridge line, and calculating protection-related quantities, including boundary existence degree, boundary locking degree, turning point protection degree, and trend risk, which are used to characterize the contribution of this layer to the real boundary main ridge, key turning points, and overall trend; then normalizing and weighting the protection-related quantities to form the protection demand drivers for this layer, and combining the coarse-layer a priori protection factors to provide additional protection for the low-frequency layer.

[0042] The noise dominance is:

[0043] ; in, Indicates the first Noise dominance of layer detail components, Indicates the first Noise variance of layer detail components, This is the regularization constant, usually taken as... The magnitude is used to avoid the denominator being zero and to ensure numerical stability. For the first Energy of layer detail components, Indicates the position index of the detail component. Indicates the first The total number of sampling points for the layer detail components. Indicating the first detail component The detail coefficients for each sampling point. The larger the value, the more dominant the noise energy is relative to the signal energy in that component, meaning the worse the signal quality. If... A larger value indicates that the layer is more likely to be dominated by noise.

[0044] To calculate cross-scale persistence, we first assume the coarse-scale amplitude, aligned with the current layer length, is... , It can be obtained by interpolating the amplitudes of adjacent coarse-scale detail components, or by resampling the amplitudes of coarser-scale detail components to the length. It was obtained later.

[0045] The basic cross-scale persistence can then be defined as: ; In the formula, Indicates the first The fundamental cross-scale persistence of layer detail components. express The absolute value; if If it is larger, then the first The fact that amplitude variations in the layer detail components can be consistently reflected in adjacent coarse-scale layers suggests that the layer is more likely to contain effective structure than isolated noise. It is the function that takes the minimum value.

[0046] Furthermore, by introducing locked weights, a locked-weighted cross-scale persistence is constructed: ; The final cross-scale persistence can be written as: ; In the formula, Indicates the first Cross-scale persistence of layer detail components, Indicates the first Weighted cross-scale persistence of layer detail components, Indicates the first The first layer of detail component Locking weights for each sampling point are used to enhance or suppress contributions at specific locations (e.g., based on phase consistency, local energy, or fluctuations). This is a weighted mixing parameter, ranging from 0 to 1, used to balance the basic persistence and the weighted persistence. When the value of is 0, only weighted cross-scale persistence is used; when the value of is 1, only basic cross-scale persistence is used. The locking weights are mapped from the reference boundary locking constraints to the ... The layer detail component length is obtained afterward and is used to improve the contribution of the boundary-locked region in persistent computation.

[0047] Local volatility is: ; ; In the formula, Indicates the first Local fluctuations in layer detail components; express The first difference is used to measure the first... Local rate of change of layer detail components.

[0048] First , and Normalize each component separately, then automatically calculate weights based on entropy weighting or normalized contribution to construct a demand-suppressing driver: ; ; In the formula, Indicates the first The need to suppress layer detail components, Indicates the first The non-discontinuity of layer detail components, , , They represent the normalized values ​​respectively. , , , They represent , , The weight. The larger the value, the more the suppression strength of that layer should be increased in subsequent threshold generation.

[0049] The first The layer detail components are reconstructed to the original seabed line length to obtain the first layer. The detailed reconstruction components of the layer are denoted as .

[0050] ; ; In the formula, Represents the inverse discrete wavelet transform. Indicates the first Layer detail reconstruction components in the first layer The value of the position of the ping sequence.

[0051] In order to measure the The contribution of the layer detail reconstruction components at each ping location is defined by the normalized magnitude distribution: ; In the formula, Indicates the first Layer detail reconstruction component corresponds to the first The normalized magnitude of each ping sequence.

[0052] The normalized amplitude distribution is coupled with the relevant constraint quantities of the reference boundary ridge to calculate the protection-related quantities, which include the boundary existence degree, boundary locking degree, turning point protection degree, and trend risk defined according to the normalized amplitude distribution. The existence degree of the boundary is: ; Boundary locking degree is: ; The transition protection degree is: ; The trend risk is: ; in, Let be the trend risk function, used to characterize the . Will excessive suppression of layer detail components disrupt the overall trend of the seabed line?

[0053] To enhance executability, the trend risk function can be defined as: ; ; In the formula, Indicates the first The scale weight of the layer, when At that time, it represents the highest frequency detail layer. When representing the coarsest level of detail, The larger the value, the closer the layer is to a coarse-scale layer, and the stronger its influence on the overall trend.

[0054] Normalizing and weighting the above quantities, we obtain the basic protection demand drivers: ; In the formula, Indicates the first Basic protection requirements for layer detail components , , , They represent the normalized values ​​respectively. , , , , Indicates the first The existence degree of the boundary of the layer detail component. Indicates the first Boundary locking degree of layer detail components, Indicates the first The degree of protection of the transition of layer detail components. Indicates the first Trend risks of layer detail components.

[0055] Then introduce a coarse-layer prior protection factor This leads to a revised protection requirement-driven approach: ; In the formula, Indicates the first The need for protection of layer detail components, To protect the requirements, a correction function is used.

[0056] Among them, the coarse-layer a priori protection factor is used to reflect the overall trend of the submarine line under the low-frequency layer, so it should be more conservative under the condition of strong boundary support.

[0057] The coarse-layer a priori protection factor is calculated from the coarsest-layer approximation component. To use the coarsest-layer approximation component to perform trend correction on protection requirements, the coarsest-layer approximation component obtained in step 4 is... Reconstructing to the original seabed line length yields the overall trend sequence: ; ; In the formula, This represents the approximate components of the coarsest layer. The overall trend sequence obtained from the reconstruction Indicates the first The overall trend value at each ping location.

[0058] The first difference of the overall trend sequence is defined as follows: ; Definition of the first The first-order difference of the layer detail reconstruction components is divided into: ; Then the first Layer detail reconstruction components With the overall trend sequence The consistency of trends between them is as follows: ; In the formula, Indicates the first The consistency of trend between the layer-level detailed reconstruction components and the overall trend sequence. A larger value indicates greater consistency. Layer detail reconstruction components and The more consistent the overall trend represented, the more likely the layer used to represent the corresponding scale is to contain structural components that contribute to the actual trend of the seabed.

[0059] Constructing a coarse-level prior protection factor based on trend consistency and scale weighting: ; In the formula, This represents the strength coefficient of the coarse layer protection.

[0060] Instead of being directly used as the final threshold or final seabed line output, it serves as a priori for the overall trend, acting as a coarse-level prior protection factor. Modify basic protection requirements drive This allows effective structures that align with the overall trend and are located in coarse-scale layers to receive higher protection requirements.

[0061] Step 6 determines the suppression or protection tendency based on the balance score of each layer's suppression and protection drive, generates multiple sets of candidate threshold offsets, and performs inter-layer consistency correction on each set of candidate threshold offsets. Based on the corrected threshold offsets, the candidate threshold weights and layer thresholds are calculated. After inter-layer consistency correction (except for coarse layer protection dominance), the thresholds of multi-scale detail components are reconstructed by threshold contraction. After light continuity correction, the candidate final seabed line is obtained. Then, candidates with obvious deviations are eliminated according to the boundary feasibility index, and the final seabed line is selected by smoothness quality sorting.

[0062] like Figure 3 As shown, a balance score is constructed based on the suppression drive and protection drive of each layer, and the layer is determined to be more suppressed or more protected based on the balance score; multiple sets of threshold offsets are generated under the candidate configuration set, and allowable offset ranges are set in the protection direction and the suppression direction, thereby obtaining multiple sets of candidate threshold weights and candidate layer thresholds.

[0063] Define the balance fraction: ; In the formula Indicates the first The balance fraction of the layer is equivalent to and The competition score, when When, it indicates that suppressed demand is dominant; when At that time, it indicates that the need for protection is dominant.

[0064] Further Implement robust centralization: ; In the formula, Indicates the first The balance value of layer detail components, for The central value is equivalent to the median of the sequence, Indicates the range of values ​​for the median; To avoid hardcoding the final threshold into a single closed result, a candidate configuration set is used to generate multiple candidate threshold schemes. The candidate configuration set is as follows: ; In the formula, Indicates the first Group candidate configuration, This represents the total number of candidate configurations. Candidate configurations must include at least the score pruning scale. Suppression direction offset coefficient Protection direction offset coefficient Interlayer smoothing coefficient Maximum offset difference with adjacent layers .

[0065] For any candidate configuration First, construct the clipping response: ; In the formula, Indicates the first The response value of the group candidate configuration. for The fractional cropping scale, This is a numerical clipping function.

[0066] To reflect the design concept of "asymmetry between protection and suppression directions", allowable offsets for the protection direction and suppression direction are defined for each candidate configuration. The allowable offset for the protection direction is: ; The allowable offset of the suppression direction is: ; in, Indicates the first Consistency between layers Indicates the first Layer threshold adjustment degrees of freedom Indicates the allowable offset in the protection direction. The table shows the allowable offset amount for the suppression direction.

[0067] Definition of the first The degrees of freedom for adjusting the layer threshold are: ; In the formula, The larger the value, the greater the threshold adjustment allowed for that layer; the closer the layer is to a coarse scale and the stronger the protection requirement, the better. The smaller the value, the less likely the coarse-scale trend will be to be excessively altered. This is a truncation function.

[0068] Definition of the first Inter-layer consistency: ; in, For consistency scale parameters, Indicates the first Layer balance fraction, Indicates the first Layer balance fraction, Indicates the first The balance fraction of the layer; for the boundary layer, the calculation can be performed using adjacent layers on one side.

[0069] The allowable offset of the suppression direction is: ; The allowable offset for the protection direction is: ; In the formula, This represents the suppression direction offset coefficient. This represents the protection direction offset coefficient.

[0070] when When this occurs, it indicates that the demand for suppression is dominant, and the threshold should be shifted towards increasing; when... When this occurs, it indicates that protection needs are dominant, and the threshold should be shifted towards decreasing values. This yields the original candidate threshold offset: ; When a layer is coarse and the protection requirement is clearly dominant, protective reverse compensation can be added: ; In the formula, The protective reverse compensation strength coefficient is used to further shift the layer towards the threshold reduction direction by adding protective reverse compensation, in order to avoid over-thresholding of coarse-scale trends when protection needs prevail.

[0071] Perform inter-layer consistency correction on the original candidate threshold offset: ; In the formula, This is the consistency correction function. Interlayer consistency correction is a technique used in multi-scale, hierarchical processing (such as multi-scale image analysis, signal processing, wavelet transform, etc.) to ensure that data or features between different layers remain coordinated and consistent. It includes at least moderate smoothing of adjacent layer offsets, limiting the offset difference between adjacent layers, and providing an exception mechanism for coarse layers when coarse layer protection is dominant, so as to prevent the intermediate frequency layer from flattening the coarse layer protection again.

[0072] To correct the original candidate threshold offset for inter-layer consistency, first perform neighboring layer smoothing: ; In the formula, Indicates the first Interlayer smoothing coefficient under group candidate configuration Indicates the first The smoothed candidate threshold offset under the group candidate configuration can be calculated using the adjacent layers on one side for the boundary layer. Then, the offset difference between adjacent layers is limited to obtain the threshold offset after consistency correction: ; In the formula, Indicates the first Maximum offset difference between adjacent layers under the group candidate configuration.

[0073] Definition of the first The lower and upper bound constraints for the layer threshold weights are as follows: ; ; In the formula, As a lower bound constraint, As an upper limit constraint, and It is a constraint generating function used to dynamically calculate the upper and lower bounds of the threshold weights based on the input parameters, thereby obtaining the first... The layer in the first Threshold weights under group candidate configuration: ; The final stratification threshold is: ; In the formula, Indicates the first Group candidate configuration Layer detail component layering threshold, This indicates the result obtained in step 4. Layer base threshold.

[0074] Example 2; A candidate boundary hierarchical control method based on hierarchical thresholds, which, in addition to generating hierarchical thresholds and determining their magnitude, can also perform hierarchical control on threshold patterns, is based on an adaptive hierarchical threshold generation method oriented towards boundary preservation.

[0075] Let the mode selection function be: ; ; Here, "soft" indicates coarse layer preservation, and "garrote" can be used in the mid-frequency layer to improve the local disturbance suppression effect when the boundary locking and transition protection are strong.

[0076] In the Under the group candidate configuration, the detail components of each layer are arranged according to... and Threshold shrinkage and reconstruction are performed to obtain candidate wavelet-smoothed seabed lines. After a light continuity correction, candidate final seabed lines are obtained. Then define the candidate boundary evaluation index. Boundary feasibility screening is conducted.

[0077] The multi-scale detail components are thresholded and reconstructed using the candidate hierarchical thresholds of each group to obtain candidate wavelet smoothed seabed lines. Then, a light continuity correction is performed to obtain the candidate final seabed lines. Subsequently, the boundary feasibility index is calculated based on the reference ridge line to eliminate candidates with obvious boundary deviations. Then, the remaining candidates are sorted according to smoothing quality, and the final seabed line result is selected.

[0078] Define candidate boundary evaluation: ; In the formula, Indicates the candidate boundary evaluation index, , , , These represent the average reference ridge offset, the proportion of misalignment in the strong boundary area, the turning offset, and the precise edge retention rate, respectively.

[0079] Inter-layer consistency correction is applied to the candidate threshold offset to maintain continuous threshold changes between adjacent scale layers. An exception mechanism is set up for cases where coarse-layer protection is dominant to prevent coarse-layer protection from being smoothed out under the constraints of mid-to-high frequency layers. The multi-scale detail components are thresholded and reconstructed using the candidate layer thresholds of each group to obtain candidate wavelet-smoothed seabed lines. A light continuity correction is then performed to obtain the final candidate seabed lines. Subsequently, the boundary feasibility index is calculated based on the reference ridge line to eliminate candidates with significant boundary deviations. The remaining candidates are then sorted by smoothing quality, and the final seabed line result is selected.

[0080] For candidates that meet the boundary feasibility requirements, then proceed according to the smoothing quality. Sort the candidates and select the final seabed line. If no strictly feasible candidate exists, select the one with the smallest overall default value from the near-feasible candidates. The candidate selection principle is to first protect the boundary, and then consider smoothing.

[0081] After initial processing (such as threshold shrinkage and reconstruction to obtain intermediate results like candidate seabed lines), various factors (such as noise interference and algorithm limitations) may lead to discontinuities in the results, such as broken lines, isolated points, or segments. "Lightweight continuity correction" aims to fine-tune these discontinuous parts in a relatively simple and efficient way, making the results smoother, more coherent, and more consistent with the continuity characteristics of real-world scenarios, without introducing excessive complex calculations or altering the main features of the results.

[0082] Figure 2 This is a schematic diagram of a local reference boundary ridge, showing a magnified view of the seabed boundary in real side-scan sonar data. The yellow curve represents the original seabed line (original seabed line detection result), reflecting the tortuous changes of the actual seabed topography. The red curve represents the reference boundary ridge (main boundary line) extracted based on local boundary responses within the neighborhood of the original seabed line. This ridge is obtained through a constraint degree algorithm for smoothing and reflects the main boundary features. The blue dashed line and light blue area represent candidate boundary response bands and their ranges, used to show the possible range of boundary responses. The width covers approximately 20-30 pixels on both sides of the original line. Figure 2 As shown, the seabed boundary in side-scan sonar images often corresponds to a response area of ​​a certain width, rather than a strictly single-pixel boundary. Compared to the original seabed line, the reference boundary ridge can more stably characterize the main response position of the boundary, while the candidate response band further reflects the uncertainty range and acceptable fluctuation range of the boundary location.

[0083] Figure 4 The local comparison diagrams of the original seabed line, the wavelet-smoothed seabed line, and the final seabed line provided by this invention demonstrate the processing results of the adaptive layered threshold in a real side-scan sonar local region. Figure 4The mid-grayscale background image represents the actual side-scan sonar echo image. The yellow curve, cyan dashed line, and red curve represent the original seabed line, the wavelet-smoothed seabed line, and the final seabed line, respectively. It can be seen that the original seabed line is affected by noise interference, texture undulations, and local anomaly responses, exhibiting significant high-frequency jitter and local shifts. After wavelet smoothing, the overall trend of the curve becomes significantly more stable, and random disturbances are effectively suppressed. Further combining continuity constraints and anomaly correction, the final seabed line maintains the overall shape of the seabed boundary while exhibiting better smoothness, continuity, and stability, more accurately representing the dominant changing trend of the real seabed boundary. During the smoothing process, a layered threshold strategy is used to differentiate the wavelet detail components at different scales. This strategy comprehensively considers the noise intensity, cross-scale persistence, local fluctuation characteristics, and their supporting role for the boundary structure of each scale component. Stronger suppression is applied to noise-dominated and less stable scale components, while effective scale components that reflect the main shape and transition characteristics of the boundary are preserved as much as possible. Therefore, the wavelet smoothing result can effectively preserve boundary structure information while reducing random jitter, providing a reliable basis for the stable output of the final seabed line.

[0084] Figure 5 The figure shows the processing results of the adaptive layered threshold provided by this invention in real side-scan sonar, presenting the final seabed line extraction result of the central core region after stitching the port and starboard sides. The grayscale background image is the real side-scan sonar echo image, and the red curve represents the final output seabed line. To highlight the main distribution characteristics of the seabed boundary in the entire data segment, redundant areas on both sides have been removed from the complete stitched result, retaining only the stitched core region. Figure 5 As can be seen, the final seabed line maintains good continuity and overall stability across both port and starboard sides, clearly reflecting the spatial variation trend of the seabed boundary throughout the data. The seabed boundary response in the starboard region is relatively clear, with a more pronounced echo structure, thus the final seabed line closely matches the actual boundary position, exhibiting high boundary consistency. In contrast, the image quality in the port region is relatively poor, exhibiting weak boundary contrast, insufficient local texture information, and a less prominent effective echo structure. These conditions typically increase the difficulty of seabed line detection and stable reconstruction. Figure 5As shown, even with low-quality data on the port side, a continuous, smooth seabed line with a reasonable geometric shape can still be recovered, without obvious breaks, abrupt jumps, or large-scale drift. This result indicates that boundary line extraction based on adaptive layered thresholds not only maintains good extraction accuracy in areas with strong boundary information and clear structural representation, but also demonstrates good recovery capability and robustness for data with poor echo quality and weak boundary information, such as that from the port side. Furthermore, boundary line extraction based on adaptive layered thresholds balances the stability of boundary tracking with the structural rationality of the results under complex acoustic imaging conditions, providing reliable results for seabed boundary extraction in full-frame side-scan sonar data.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A boundary-preserving adaptive hierarchical threshold generation method, characterized in that, include: S1. Read and parse the raw side-scan sonar data containing port and starboard echo sampling sequences, construct port and starboard waterfall diagram matrices, enhance and optimize the matrices, and generate a detection enhancement matrix; S2. Perform row-by-row boundary tracking on the detection enhancement matrix, extract the original seabed line sequence, and perform light pre-cleaning on the original seabed line sequence through interpolation repair to obtain the pre-cleaned seabed line sequence. S3. Based on the original seabed line sequence, extract reference boundary ridges in the neighborhood of the original seabed line, and construct constraint quantities for the reference boundary ridges. The constraint quantities include boundary existence degree, boundary locking degree, and key turning point importance. S4. Perform multi-scale discrete wavelet decomposition on the pre-cleaned seabed line sequence to obtain approximate components and detail components at different scales, and calculate the basic threshold of the detail components at each scale. S5. Calculate the suppression correlation and protection correlation of detail components at each scale layer, normalize and weight the suppression correlation and protection correlation respectively, construct the suppression demand driver and basic protection demand driver at each scale layer, correct the basic protection demand driver at each scale layer through approximation components, and construct the protection demand driver at each scale layer. S6. Construct a balanced score based on the suppression demand drive and the protection demand drive, generate multiple sets of candidate threshold offsets, and perform inter-layer consistency correction on each set of candidate threshold offsets. Calculate the candidate threshold weights based on the corrected threshold offsets, and process the basic thresholds of S4 through the candidate threshold weights to obtain the layered thresholds. Original submarine line sequence for: ； In the formula, Indicates the sequence number of the ping sequence. Indicates the first The original seabed line position values ​​corresponding to each ping sequence. Indicates the total length of the submarine line; Anomaly queries were performed on the original seabed line sequence. Ping sequences exhibiting both abnormal trend deviations and anomalous jumps were marked as pre-cleaned anomalies. The original seabed line position values ​​of these pre-cleaned anomalies were then repaired using interpolation to obtain the pre-cleaned seabed line sequence. : ； In the formula, Indicates the first time after cleaning The original seabed line position values ​​corresponding to each ping sequence; The reference boundary ridge line is: ； In the formula, Indicates the reference boundary ridge line. Indicates the ridge line corresponding to the first Sampling points of a ping sequence; The existence degree of the boundary is: ； In the formula, This indicates that there is a constraint vector at the boundary. for The corresponding number in the middle The components of a ping sequence; Boundary locking degree is: ； In the formula, Represents the boundary locking constraint vector. for The corresponding number in the middle The components of a ping sequence; The importance of key turning points is as follows: ； In the formula, This represents the vector of importance constraints for transitions. for The corresponding number in the middle The components of a ping sequence; The basic threshold for the detail components at each scale layer is: ； ； In the formula, Indicates the first The base threshold for the detail components of the i-th scale layer is equivalent to the i-th scale layer. The base threshold for layer detail components, Indicates the first The length of the layer detail factor, Let e ​​represent the logarithmic function with the numerical constant e as the base. Indicates the first The noise standard deviation of the detail components at each scale layer. This indicates the calculation of the median absolute deviation. Indicates the first Detail components of each scale layer; Suppression-related quantities include noise dominance, cross-scale persistence, and local volatility; protection-related quantities include boundary presence, boundary locking, turning point protection, and trend risk. Normalize and weight the inhibition-related quantities to construct the inhibition demand driver: ； ； In the formula, Indicates the scale layer index. Indicates the total number of scale layers. Indicates the first The need to suppress layer detail components, , , They represent the normalized values ​​respectively. , Indicates the first Noise dominance of layer detail components, Indicates the first The non-discontinuity of layer detail components, Indicates the first Cross-scale persistence of layer detail components, Indicates the first Local fluctuations in layer detail components, , , They represent , , The weights; Normalize and weight the protection-related quantities to calculate the basic protection demand drivers: ； In the formula, Indicates the first Basic protection requirements for layer detail components. , , , They represent the normalized values ​​respectively. , , , , Indicates the first The existence degree of the boundary of the layer detail component. Indicates the first Boundary locking degree of layer detail components, Indicates the first Protection of transitions in layer detail components. Indicates the first The trend risk of layer detail components, , , , They represent , , , The weights; Introducing coarse-layer prior protection factors right Make corrections to construct protection requirements-driven mechanisms: ； In the formula, Indicates the first The need for protection of layer detail components, To protect the requirements, a correction function is used.

2. The boundary-preserving adaptive hierarchical threshold generation method according to claim 1, characterized in that, The noise dominance is: ； ； In the formula, Indicates the first Noise variance of layer detail components, Here is the regularization constant. For the first Energy of layer detail components, Indicates the position index of the detail component. Indicates the first The total number of sampling points for the layer detail components. Indicating the first detail component Detail coefficients for each sampling point; Cross-scale persistence is: ； ； ； In the formula, For continuously weighted mixing parameters, No. The fundamental cross-scale persistence of layer detail components. Indicates the first Weighted cross-scale persistence of layer detail components, Indicates the coarse-scale amplitude. Indicates the first The first layer of detail component The locking weight of each sampling point; Local volatility is: ； In the formula, express The first-order difference.

3. The adaptive hierarchical threshold generation method for boundary preservation according to claim 2, characterized in that, The detail components of each scale layer are reconstructed to the original seabed line length, the detail reconstruction components of each scale layer are obtained, and a normalized amplitude distribution is constructed: ； In the formula, Indicates the first Layer detail reconstruction component corresponds to the first The normalized magnitude of a ping sequence. Indicates the first Layer detail reconstruction component corresponds to the first The reconstruction coefficients of each ping sequence, based on Calculate protection-related quantities; The existence degree of the boundary is: ； Boundary locking degree is: ； The transition protection degree is: ； The trend risk is: ； In the formula, For trend risk function, Indicates the first Layer detail reconstruction components.

4. The boundary-preserving adaptive hierarchical threshold generation method according to claim 3, characterized in that, The balance fraction is: ； In the formula, Indicates the first The balance value of layer detail components; right Implement robust centralization: ； In the formula, for The central value, This indicates the value of the median.

5. The boundary-preserving adaptive hierarchical threshold generation method according to claim 4, characterized in that, Multiple candidate threshold schemes are generated using a candidate configuration set, which is: : ； In the formula, Indicates the first Group candidate configuration; right Constructing a clipping response: ； In the formula, Indicates the first The response value of the group candidate configuration. for The fractional cropping scale, This is a numerical clipping function; based on Generate candidate threshold offsets: ； In the formula, This indicates the allowable offset in the threshold direction. Indicates the allowable offset in the protection direction; Perform inter-layer consistency correction on the candidate threshold offset to obtain the corrected threshold offset: ； In the formula, Indicates the corrected number Threshold offset for group candidate configurations. Indicates degrees of freedom. Indicates consistency. This is a consistency correction function.

6. The boundary-preserving adaptive hierarchical threshold generation method according to claim 5, characterized in that, definition Upper and lower bound constraints are used to calculate candidate threshold weights: ； ； ； In the formula, Indicates the first Candidate threshold weights for group candidate configurations. This indicates the lower limit of the candidate threshold. Indicates the upper limit of the candidate threshold. Let be the lower bound constraint function. For upper limit constraint functions; The stratification threshold is: ； In the formula, Indicates the first Layer detail components in the first Hierarchical threshold under group candidate configuration.

Images