A Feature-Adaptive Remote Sensing Method and System for Detecting Small Targets

Abstract

This invention discloses a feature-adaptive remote sensing method and system for detecting small targets, belonging to the field of remote sensing small target detection technology. The method includes: dividing remote sensing image data into sliding windows; using smoothing processing to treat large-scale background structures as slowly varying component maps; quantifying structural indicators by weighted combination of the maximum length of continuous pixel chains as a primary directional continuity index; performing residual energy statistics on detail residual components and weighted combination with the residual energy statistics to quantify detail complexity and, combined with structural indicators, labeling image blocks to obtain smooth background blocks and structural edge blocks. The method of this invention uses structural indicators and detail complexity indicators simultaneously within the same framework, and requires that the combination relationship between these two indicators and their respective thresholds meets specific conditions for classification into background or edge categories during block-level decision-making, thus preventing block classification from being unidirectionally dominated by a single indicator.

Description

Technical Field

[0001] This invention relates to the field of remote sensing weak target detection technology, and in particular to a remote sensing weak target detection method and system based on feature adaptation. Background Technology

[0002] Because remote sensing imaging has a long distance and a long imaging process, the observation data is often superimposed with large-scale slowly changing background, fine ground texture and multi-source noise interference. The energy distribution of weak targets in the spatial and frequency domains is highly discrete, and their significance is much lower than that of conventional targets. In order to improve the adaptability to targets at different spatial scales, multi-resolution detection strategies based on multi-scale pyramids, wavelet transforms and directional filtering are often used to extract high-frequency sub-bands or gradient abrupt responses at different scales to characterize candidate targets.

[0003] However, existing weak target detection methods based on local contrast or background prediction rarely characterize background structural differences at the block level, such as the essential differences between continuous skeleton edges, banded textures, and large flat areas. This leads to a high degree of overlap between targets and strong edges and texture clutter in statistical features. While multi-scale pyramid and wavelet domain methods provide rich scale information, they usually lack mechanisms to adaptively adjust detection strategies for different intra-block structural features. It is difficult to simultaneously ensure the visibility of targets in smooth backgrounds and complex edge regions, often resulting in residual noise in flat areas or over-suppression of targets near edges. In addition, during transform domain enhancement, most methods process sub-bands independently, ignoring the joint constraints between high-frequency components in multiple directions. They also rarely combine geometric priors such as local annular brightness contrast and banded neighborhood morphology to screen candidate responses. This makes it difficult to suppress large-scale gradually changing backgrounds and edge protrusions while providing targeted enhancement for extremely weak point targets, thereby reducing the detection accuracy of remote sensing weak targets. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a feature-adaptive remote sensing method and system for detecting weak targets. This addresses the shortcomings of existing methods based on local contrast or background prediction, which rarely characterize background structural differences at the block level and often lack mechanisms to adaptively adjust detection strategies for different intra-block structural features. These methods struggle to simultaneously ensure the visibility of targets in smooth backgrounds and complex edge regions, and rarely incorporate geometric priors such as local annular brightness contrast and banded neighborhood morphology to screen candidate responses. Consequently, they fail to effectively suppress large-scale, gradually changing backgrounds and edge protrusions while simultaneously enhancing extremely weak point-like targets, thus reducing the detection accuracy of remote sensing weak targets.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a remote sensing weak target detection method based on feature adaptation, comprising:

[0008] Based on remote sensing image data, a sliding window is divided. Through smoothing, the large-scale background structure is treated as a slowly varying component map. The maximum length of the continuous pixel chain is used as the main direction continuity index for weighted combination and quantification of structural index. The residual energy of the detail residual component is statistically analyzed and weighted combination with the residual energy statistics to quantify the detail complexity. Combined with the structural index, the image patch is calibrated to obtain smooth background patch and structural edge patch.

[0009] Two-dimensional wavelet decomposition is performed on the smooth background block to form a joint high-frequency amplitude map and a binary mask is constructed to obtain the high-frequency joint response value. The corresponding image block is decomposed into inner and outer ring pixel sets, and the brightness contrast baseline between the inner and outer rings is analyzed. Pixel-level high-frequency enhancement coefficients are constructed based on the high-frequency joint response value. Combined with the high-frequency sub-blocks of the two-dimensional wavelet decomposition, the contrast-enhanced high-frequency sub-blocks are obtained. Inverse wavelet reconstruction is performed, and spatial mapping is performed on the original remote sensing image to obtain the multi-scale response field of the smooth background block.

[0010] By setting the banded neighborhood bandwidth and spatial sampling of the structural edge blocks, a banded neighborhood set along both sides of the skeleton is obtained. The direction and normal of the high-frequency sub-blocks corresponding to the banded neighborhood set are analyzed, and protrusion detection and angle deviation analysis are performed. The high-frequency sub-blocks are structurally adjusted. Inverse wavelet reconstruction is performed on the low-frequency sub-blocks and the high-frequency sub-blocks with structural adjustments to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image.

[0011] For the multi-scale response fields of structural edge blocks and smooth background blocks, the corresponding block-level response sets are determined by the spatial mapping relationship of image blocks in the original remote sensing image. The pixel-level robust response map is calculated by the hyperbolic tangent function, and the relationship between adjacent nodes is defined to form a set of multiple weak target trajectories. The final weak target detection set is selected by the NMS global nonmaximum suppression algorithm.

[0012] As a preferred embodiment of the feature-adaptive remote sensing weak target detection method of the present invention, the method comprises: quantizing structural indices by weighting and combining the maximum length of the statistical continuous pixel chain as a main directional continuity index; performing residual energy statistics on the detail residual components and weighting and combining them with the residual energy statistics to quantify the detail complexity; and calibrating image patches by combining the structural indices to obtain smooth background patches and structural edge patches, including...

[0013] For preprocessed remote sensing image data, based on a predefined set of multi-scale block sizes and overlap step size parameters, the remote sensing image data is divided into a sliding window according to the block size and step size at each scale to obtain a set of multi-scale image blocks, where each image block corresponds to a spatial region in the remote sensing image, and is based on the starting coordinates and scale index of the image block in the remote sensing image.

[0014] For each image patch, a convolution operation is performed using a preset smoothing kernel to determine the large-scale background structure within the patch as the slowly varying component map, and the difference between the image patch and the slowly varying component map is calculated as the detail residual component.

[0015] Gradient components are calculated for the horizontal and vertical directions of the detailed residual components. Euclidean norm and arctangent operations are performed on the gradient components to obtain the local edge intensity and gradient direction distribution, respectively.

[0016] The gradient direction distribution is quantized in discrete direction intervals, and the maximum length of the continuous pixel chain along the direction in each direction interval is counted as the main direction chain length. The number of pixels of the continuous pixel chain with the maximum length is marked as the main direction continuity index.

[0017] The average gradient magnitude within the image patch is calculated for local edge intensity, and then weighted and combined with the main direction continuity index to quantify the structural index.

[0018] For the detailed residual components, the average of the squared residuals is calculated, and residual energy statistics are performed.

[0019] Based on the historical isolation threshold, the isolated extreme values ​​of the detailed residual components are detected to obtain the number of isolated strong points. The number of isolated strong points is then weighted and combined with the residual energy statistics to quantify the detail complexity.

[0020] Based on the calibrated structural index data and residual energy statistics, structural thresholds and complexity thresholds are determined respectively. The structural thresholds and complexity thresholds are adaptively determined based on the global feature statistics of the current image.

[0021] If the structural index data is less than the structural threshold and the residual energy statistics are less than the complexity threshold, the image patch is labeled as a smooth background patch. If the structural index data is greater than or equal to the structural threshold and the residual energy statistics are less than the complexity threshold, the image patch is labeled as a structural edge patch.

[0022] As a preferred embodiment of the feature-adaptive remote sensing weak target detection method of the present invention, the method involves: constructing pixel-level high-frequency enhancement coefficients based on high-frequency joint response values, combining these with high-frequency sub-blocks obtained from two-dimensional wavelet decomposition to obtain contrast-enhanced high-frequency sub-blocks, performing inverse wavelet reconstruction, and spatially mapping these sub-blocks in the original remote sensing image to obtain a smooth background block multi-scale response field, including...

[0023] For image blocks marked as smooth background blocks, two-dimensional wavelet decomposition is performed on the data within the block based on orthogonal wavelet basis functions to obtain low-frequency sub-blocks and high-frequency sub-blocks in the horizontal, vertical and diagonal directions. These are used as local high-frequency amplitudes in the corresponding directions and form a joint high-frequency amplitude map. At the same time, the joint high-frequency amplitude is defined as the infimum of the amplitudes in the three directions.

[0024] For the joint high-frequency amplitude map, pixel-level comparison and binary mask construction are performed based on the joint amplitude statistics of the corresponding image blocks and the preset high-frequency joint response threshold to obtain the high-frequency joint response value;

[0025] For remote sensing data of weak targets detected in history, the inner and outer ring radius thresholds of image blocks are defined. The image blocks marked as smooth background blocks are divided into inner ring pixel sets and outer ring pixel sets according to the inner and outer ring radius thresholds. The regional mean of the low-frequency sub-blocks of the two sets is calculated as the inner ring average brightness and the outer ring average brightness, respectively. The difference between the inner ring and outer ring average brightness is used as the brightness comparison baseline.

[0026] Nonlinear normalization is performed based on the brightness contrast baseline, the inner ring enhancement coefficient and the outer ring enhancement coefficient are calculated respectively, and the pixel-level high-frequency enhancement coefficient is constructed based on the high-frequency joint response value.

[0027] The product of the high-frequency enhancement coefficient and the high-frequency sub-block is used as the high-frequency sub-block after contrast enhancement. The frequency domain reconstruction operation is performed by wavelet inverse transform to obtain the enhanced spatial domain response map marked as the smooth background block. The map is then spatially mapped in the original remote sensing image to obtain the multi-scale response field of the smooth background block in the original remote sensing image.

[0028] As a preferred embodiment of the feature-adaptive remote sensing weak target detection method of the present invention, the method involves: analyzing the direction and normal of the high-frequency sub-blocks corresponding to the analyzed strip neighborhood set, performing protrusion detection and angle deviation analysis, adjusting the structure of the high-frequency sub-blocks, and performing inverse wavelet reconstruction on the low-frequency sub-blocks and the high-frequency sub-blocks with adjusted structure to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image, including...

[0029] For image blocks marked as structural edge blocks, the edge orientation field of the main edge direction and the set of main edge skeleton pixels are determined by gradient calculation and orientation aggregation through the slowly varying component map.

[0030] Based on the main edge skeleton pixel set and the edge direction field, the tangential vector and normal vector at the pixel are determined, and the band neighborhood bandwidth is set and spatial sampling is performed to obtain the band neighborhood set along both sides of the skeleton.

[0031] Based on the strip neighborhood set, high-frequency profiles are calculated along the skeleton direction and normal according to the corresponding high-frequency sub-blocks, and horizontal and vertical high-frequency profiles are obtained respectively.

[0032] Protrusion detection and angle deviation analysis are performed based on horizontal and vertical high-frequency profiles to determine the target candidate pixel set and the pure edge pixel set;

[0033] The structure of high-frequency sub-blocks is adjusted based on the target candidate pixel set and the pure edge pixel set;

[0034] Inverse wavelet reconstruction is performed on low-frequency sub-blocks and high-frequency sub-blocks with structural adjustments, and coordinate mapping is performed on the original remote sensing image to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image.

[0035] As a preferred embodiment of the feature-adaptive remote sensing weak target detection method of the present invention, wherein: for the multi-scale response fields of structural edge blocks and smooth background blocks, the corresponding block-level response sets are determined by the spatial mapping relationship of image blocks in the original remote sensing image, including,

[0036] The intersection of the multi-scale response fields of the structural edge blocks and the multi-scale response fields of the smooth background blocks forms the enhanced block response set;

[0037] By utilizing the spatial mapping relationship between each image block in the enhanced block response set and the original remote sensing image, the corresponding block-level response set for each pixel in the original remote sensing image at each scale is obtained;

[0038] The average response and variance within the corresponding block-level response set are calculated, pixel-level consistency is calculated, and pixel-level robust response maps are calculated using the hyperbolic tangent function.

[0039] As a preferred embodiment of the feature-adaptive remote sensing weak target detection method of the present invention, wherein: the step of calculating a pixel-level robust response map using the hyperbolic tangent function and defining the relationship between adjacent nodes to form a set of multiple weak target trajectories includes,

[0040] Based on the pixel-level robust response map, the scale is used as the third dimension coordinate for dimensional expansion, and the relationship between adjacent nodes is defined.

[0041] Based on the definition of adjacent node relationships, each node of the pixel-level robust response map is recursively connected in a direction to form a set of multiple weak target trajectories, where each trajectory is represented by the coordinates of the three-dimensional response map.

[0042] As a preferred embodiment of the feature-adaptive remote sensing target detection method of the present invention, wherein: the final target detection set is selected by the NMS global nonmaximum suppression algorithm, including,

[0043] For a set of multiple weak target trajectories, redundancy is removed using the NMS global nonmaximum suppression algorithm to obtain the final weak target detection set.

[0044] Secondly, the present invention provides a remote sensing weak target detection system based on feature adaptation, comprising,

[0045] The remote sensing image acquisition module receives preprocessed remote sensing image data and divides the sliding window according to a preset block size and step size.

[0046] The structural detail complexity module smooths image patches to obtain gradually varying components and detail residual components, and constructs structural indices and detail complexity based on the continuous pixel chain length and residual energy.

[0047] The inner and outer ring contrast enhancement module is used to perform two-dimensional wavelet decomposition and joint high-frequency amplitude construction on the smooth background block. Based on the brightness contrast of the inner and outer rings and the joint high-frequency response, pixel-level high-frequency enhancement coefficients are generated and inverse wavelet reconstruction is completed to form a multi-scale response field of the smooth background block.

[0048] The high-frequency structure adjustment module is used to construct a strip-shaped neighborhood set based on the main edge skeleton, calculate the high-frequency profile along the skeleton direction and normal, and perform protrusion detection and angle deviation analysis. It adaptively adjusts the high-frequency sub-blocks and combines them with the low-frequency sub-blocks to inverse wavelet reconstruction to obtain the multi-scale response field of the structural edge block.

[0049] The pixel-level robust response calculation module is used to aggregate the multi-scale responses of smooth background blocks and structural edge blocks based on the mapping relationship of image blocks in the original image, construct a pixel-level block-level response set and calculate the average response and variance, and generate a pixel-level robust response map through the hyperbolic tangent function.

[0050] The weak target trajectory generation module is used to construct a three-dimensional response map by taking scale as the third dimension, defining node connectivity based on spatial proximity and scale adjacency, and performing directional recursive connection to form a set of multiple weak target trajectories.

[0051] The weak target detection result output module is used to perform a global nonmaximum suppression algorithm on the weak target trajectory set to filter out redundant trajectories and output the final weak target detection result set.

[0052] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the feature-adaptive remote sensing weak target detection method as described in the first aspect of the present invention.

[0053] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the feature-adaptive remote sensing weak target detection method as described in the first aspect of the present invention.

[0054] The beneficial effects of this invention are as follows: By using structural indicators and detail complexity indicators simultaneously within the same framework, and requiring that the combination relationship between these two indicators and their respective thresholds meets specific conditions in block-level decision-making before being classified into background or edge categories, block classification is no longer unidirectionally dominated by a single indicator. By constructing pixel-level high-frequency enhancement factors by combining the nonlinear normalized enhancement coefficients of the inner and outer loops with the high-frequency joint response, and then applying them to high-frequency sub-blocks in each direction for inverse wavelet reconstruction, the brightness enhancement in the spatial domain is strictly constrained by two sources: ring contrast and multi-sub-band consistency. By adaptively determining the adjustment coefficients for candidate targets and pure edges based on the filter response statistics and coefficient of variation, and then scaling the high-frequency components in different directions point by point, a set of reweighting factors consistent with skeleton geometry, local high-frequency morphology, and global statistics is formed, so that the high-frequency structure of the same position in each directional sub-band is collaboratively reconstructed. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the feature-adaptive remote sensing target detection method in Example 1.

[0057] Figure 2 This is a schematic diagram of the remote sensing weak target detection system based on feature adaptation in Example 1. Detailed Implementation

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0060] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0061] Example 1, referring to Figures 1 to 2 This is the first embodiment of the present invention, which provides a remote sensing weak target detection method based on feature adaptation, including the following steps:

[0062] S1. Based on remote sensing image data, a sliding window is divided. Through smoothing, the large-scale background structure is taken as a slowly varying component map. The maximum length of the continuous pixel chain is used as the main direction continuity index for weighted combination and quantification of structural index. The residual energy of the detail residual component is statistically analyzed and weighted combination with the residual energy statistics to quantify the detail complexity and combine with the structural index to calibrate the image block, thus obtaining the smooth background block and the structural edge block.

[0063] Preferably, the maximum length of the continuous pixel chain is used as a weighted combination of structural indicators to quantize the main direction continuity index. Residual energy statistics are performed on the detail residual components, and then weighted and combined with the residual energy statistics to quantify the detail complexity. Combined with the structural indicators, image blocks are calibrated to obtain smooth background blocks and structural edge blocks, including...

[0064] For preprocessed remote sensing image data, based on a predefined set of multi-scale block sizes and overlap step size parameters, the remote sensing image data is divided into a sliding window according to the block size and step size at each scale to obtain a set of multi-scale image blocks, where each image block corresponds to a spatial region in the remote sensing image, and is based on the starting coordinates and scale index of the image block in the remote sensing image.

[0065] For each image patch, a convolution operation is performed using a preset smoothing kernel to extract the large-scale background structure within the patch as a slowly varying component map. The difference between the image patch and the slowly varying component map is then calculated as the detail residual component, represented as follows:

[0066]

[0067]

[0068] in, Represents the gradually varying component plot. This indicates the support window for the smoothing kernel. This represents the weight of the smoothing kernel at (p, q). This represents the image patch with index i at coordinates (u, v) at scale s. Indicates the detailed residual components;

[0069] The gradient vector of each pixel within the block is calculated using the spatial gradient operator. Gradient components are calculated in both the horizontal and vertical directions for the detail residual components. Euclidean norm and arctangent operations are then performed on these gradient components to obtain the local edge intensity, expressed in grayscale level differences, and the gradient direction distribution, expressed in angles. These are represented as follows:

[0070]

[0071]

[0072] in, Indicates local edge intensity. and These represent the horizontal and vertical gradient components of the detail residual components, respectively. Indicates the gradient direction distribution;

[0073] The gradient direction distribution is quantized in discrete direction intervals, and the maximum length of the continuous pixel chain along the direction in each direction interval is counted as the main direction chain length. The number of pixels of the continuous pixel chain with the maximum length is marked as the main direction continuity index.

[0074] The average gradient magnitude within the image patch is calculated for local edge intensity, and then weighted and combined with the main direction continuity index to quantify the structural index. Through historical structural index data, the weighting coefficients of the average gradient magnitude and the main direction continuity are experimentally calibrated.

[0075] For the detailed residual components, the average of the squared residuals is calculated, and residual energy statistics are performed. The dimension is the square of the gray level, reflecting the overall intensity of high-frequency content within the block, expressed as:

[0076]

[0077] in, Represents the set of all pixel coordinates of an image patch. Represents residual energy statistics;

[0078] Based on historical isolation thresholds, the isolated extreme values ​​of the detail residual components are detected to obtain the number of isolated strong points. The number of isolated strong points is then weighted and combined with the residual energy statistics to quantify the detail complexity. Through historical data on the quantification of detail complexity, the weighting coefficients of the number of isolated strong points and the residual energy statistics are experimentally calibrated.

[0079] Based on the calibrated structural index data and residual energy statistics, a structural threshold and a complexity threshold are determined respectively. These thresholds are adaptively determined based on the global feature statistics of the current image, and are expressed as follows:

[0080]

[0081]

[0082]

[0083] in, and These represent the structural threshold and the complexity threshold, respectively. and Let represent the mean and standard deviation of the structural indices for all image patches, respectively. and Let represent the mean and standard deviation of the detail complexity of all image patches, respectively. and These represent the preset coefficients obtained through optimization using historical datasets. Furthermore, the optimal parameter combination can be searched using a Bayesian optimization algorithm with a representative training dataset as the objective function, and the classification F1-score as the objective function. express and The optimal parameter combination is used to achieve adaptive global feature statistics for the current image;

[0084] If the structural index data is less than the structural threshold and the residual energy statistics are less than the complexity threshold, the image patch is labeled as a smooth background patch. If the structural index data is greater than or equal to the structural threshold and the residual energy statistics are less than the complexity threshold, the image patch is labeled as a structural edge patch.

[0085] By combining multi-scale overlapping block division with intra-block gradual background separation and detail residual construction, structural measures at all scales are established in a unified detail space after removing large-scale trends. Combined with a unified global threshold system for classification, large and small blocks are comparable in the same discriminative coordinate system, avoiding the problem of misalignment of structural indicators between different scales when relying solely on multi-scale or single-scale methods, thus improving the overall cross-scale consistent discrimination capability.

[0086] By simultaneously introducing gradient direction chain length statistics and residual energy statistics into the detailed residuals, and combining them with isolated strong point detection, linear continuous structures and isolated scattered point structures are jointly modeled in the same block. This allows the same complex region to represent linear objects such as boundaries and roads based on the length of the main direction chain, and to represent small strong reflection targets based on the number of isolated strong points, thus taking into account both types of typical targets under the same discrimination framework.

[0087] By weighting structural indices with the mean gradient intensity and directional continuity, and then statistically analyzing these two composite indices at the global level, the classification boundary is no longer directly controlled by the fluctuation of a single physical quantity. Instead, it is constrained by the combination of two complementary statistics and their global distribution. By using structural and detail complexity indices simultaneously within the same framework, and requiring that the combination of these two indices with their respective thresholds meets specific conditions for block-level decision-making, block classification is no longer unidirectionally dominated by a single index, but is jointly determined by the dual conditions of structural organization degree and overall detail energy. Thus, when large-area textured backgrounds are mixed with real boundaries, it effectively avoids high-energy textures being mistaken for structural edges or low-energy but highly continuous weak edges being missed, achieving a more refined distinction between background blocks and structural blocks.

[0088] S2, perform two-dimensional wavelet decomposition on the smooth background block to form a joint high-frequency amplitude map and construct a binary mask to obtain the high-frequency joint response value. Decompose the corresponding image block into inner and outer ring pixel sets and analyze the brightness contrast baseline between the inner and outer rings. Construct pixel-level high-frequency enhancement coefficients based on the high-frequency joint response value. Combine the high-frequency sub-blocks of the two-dimensional wavelet decomposition to obtain the contrast-enhanced high-frequency sub-blocks. Perform inverse wavelet reconstruction and spatial mapping on the original remote sensing image to obtain the multi-scale response field of the smooth background block.

[0089] Preferably, pixel-level high-frequency enhancement coefficients are constructed based on the high-frequency joint response values. Combined with high-frequency sub-blocks obtained from two-dimensional wavelet decomposition, contrast-enhanced high-frequency sub-blocks are obtained. Inverse wavelet reconstruction is then performed, and spatial mapping is applied to the original remote sensing image to obtain a multi-scale response field for the smooth background block, including...

[0090] For image blocks marked as smooth background blocks, two-dimensional wavelet decomposition is performed on the data within the block based on orthogonal wavelet basis functions to obtain low-frequency sub-blocks and high-frequency sub-blocks in the horizontal, vertical and diagonal directions. These are used as local high-frequency amplitudes in the corresponding directions, with the dimension being the absolute value of the high-frequency coefficients. A joint high-frequency amplitude map is then formed, and the joint high-frequency amplitude is defined as the infimum of the amplitudes in the three directions.

[0091] For the joint high-frequency amplitude map, pixel-level comparison and binary mask construction are performed based on the joint amplitude statistics of the corresponding image blocks and a preset high-frequency joint response threshold to obtain the high-frequency joint response value, expressed as:

[0092]

[0093] in, Indicates the high-frequency joint response value. This represents the lower bound where the amplitudes in the three directions reach their minimum values. Indicates the preset high-frequency joint response threshold;

[0094] For remote sensing data of weak targets detected in history, the inner and outer ring radius thresholds of image blocks are defined. The image blocks marked as smooth background blocks are divided into inner ring pixel sets and outer ring pixel sets according to the inner and outer ring radius thresholds. The regional mean of the low-frequency sub-blocks of the two sets is calculated as the inner ring average brightness and the outer ring average brightness, respectively. The difference between the inner ring and outer ring average brightness is used as the brightness comparison baseline.

[0095] Nonlinear normalization is performed based on the brightness contrast baseline, and the inner ring enhancement coefficient and outer ring enhancement coefficient are calculated respectively. Then, the pixel-level high-frequency enhancement coefficient is constructed based on the high-frequency joint response value, which is expressed as follows:

[0096]

[0097]

[0098]

[0099] in, Indicates the inner ring reinforcement coefficient. Indicates the outer ring suppression coefficient. This represents the inner ring radius, defined by the inner and outer ring radius thresholds. This represents the outer ring radius, defined by the inner and outer ring radius thresholds. Indicates the brightness contrast baseline. To represent a very small number to prevent the denominator from being 0, and These represent the sets of pixel coordinates for the inner and outer rings, respectively. Indicates the high-frequency enhancement factor;

[0100] Based on the product of the high-frequency enhancement coefficient and the high-frequency sub-block, the high-frequency sub-block is calculated as the contrast-enhanced high-frequency sub-block. Frequency domain reconstruction is then performed using inverse wavelet transform to obtain the enhanced spatial domain response map labeled as a smooth background block. This map is then spatially mapped onto the original remote sensing image to obtain the multi-scale response field of the smooth background block in the original remote sensing image, expressed as:

[0101]

[0102] in, Represents the multi-scale response field of a smooth background block. Indicates the total number of scales. This represents the total number of image patches at scale s. The preset importance weight of scale s can be determined based on historical experience. The preset importance weight of scale s is inversely proportional to its scale value; that is, the smaller the scale number, the higher the image resolution, and the larger its weight value. Specifically, it is calculated using an inverse proportional function to ensure that higher resolution detail information occupies a more important position in the fusion process. This represents the block coverage indicator function, which takes the value 1 if the original image (x, y) falls within the block coverage area of ​​the image, and 0 otherwise. This represents the strong spatial domain response diagram.

[0103] By performing wavelet multi-subband expansion only on blocks that have been determined to be smooth backgrounds and constructing joint amplitudes with the infimum of high-frequency coefficients in three directions, these regions are transformed into a set of small-scale candidate points that respond simultaneously in all directions in the frequency domain. This weakens the single-direction dominant structure such as linear edges at the source and highlights the weak and isolated targets that are easily submerged by the large background and long boundaries.

[0104] By constructing a pixel-level binary response by combining intra-block statistics with the joint high-frequency amplitude, the retained pixels simultaneously satisfy the two conditions of smooth background judgment and multi-directional high-frequency consistent enhancement, thus distinguishing the noise coefficient that is only occasionally large in a single sub-band from small targets with truly stable spatial structures, forming a candidate point screening mechanism that combines directional consistency and statistical anomaly.

[0105] By constructing an inner and outer ring brightness contrast baseline using low-frequency components within the same smooth block and binding it to the joint high-frequency response, a large enhancement coefficient can only be obtained when the central region exhibits both multi-directional high-frequency consistent enhancement and a significant brightness difference relative to the outer ring. This allows the frequency domain candidate points to synchronously satisfy the central surrounding brightness structure of typical small targets, thus constraining the small target determination from the joint spatial domain and frequency domain dual-mode constraints.

[0106] By combining the nonlinear normalized enhancement coefficients of the inner and outer rings with the high-frequency joint response to form a pixel-level high-frequency enhancement factor, and then applying it to the high-frequency sub-blocks in each direction for inverse wavelet reconstruction, the brightness enhancement in the spatial domain is strictly controlled by two sources: ring contrast and multi-sub-band consistency. This avoids the phenomenon of overall texture coarsening caused by simply enhancing the high frequency, and makes the enhancement mainly concentrated on weak targets with isolated center and ring background features.

[0107] By superimposing the reconstruction results at each scale according to the scale importance weight and using the block coverage indicator function to precisely control the pixel participation range, the response of the same small target at different scales is fused according to the principle of high resolution priority. At the same time, the wide and gentle structure that only appears at the coarse scale is suppressed, so that the final smooth background region multi-scale response field is more sensitive to point and small patch targets without over-amplifying large-scale gradient components.

[0108] S3, set the banded neighborhood bandwidth and spatial sampling for the structural edge blocks to obtain the banded neighborhood set along both sides of the skeleton, analyze the direction and normal of the high-frequency sub-blocks corresponding to the banded neighborhood set, perform protrusion detection and angle deviation analysis, adjust the structure of the high-frequency sub-blocks, and perform inverse wavelet reconstruction for the low-frequency sub-blocks and the high-frequency sub-blocks with adjusted structure to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image.

[0109] Preferably, the orientation and normal of the high-frequency sub-blocks corresponding to the strip neighborhood set are analyzed, protrusion detection and angle deviation analysis are performed, the structure of the high-frequency sub-blocks is adjusted, and inverse wavelet reconstruction is performed on the low-frequency sub-blocks and the structured high-frequency sub-blocks to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image, including...

[0110] For image blocks marked as structural edge blocks, the edge direction field and the set of main edge skeleton pixels are determined by gradient calculation and direction aggregation through the slowly varying component map. The edge direction field is identified by non-maximum suppression and connected component tracking along the direction to identify one or more main edge curves, and the set of pixel coordinates and their local direction vectors on these curves are recorded.

[0111] Based on the main edge skeleton pixel set and the edge direction field, the tangent vector and normal vector at each pixel are determined. Bandwidth setting and spatial sampling are then performed to obtain a banded neighborhood set along both sides of the skeleton, represented as:

[0112]

[0113] in, Represents a strip-shaped neighborhood set. Represents neighborhood coordinates, This represents the displacement parameters along the skeleton direction. Represents the tangential vector. Represents the normal vector. Displacement parameters representing the direction perpendicular to the skeleton. Indicates the predicted pixel bandwidth;

[0114] Based on the strip neighborhood set, high-frequency profiles are calculated along the skeleton direction and normal according to the corresponding high-frequency sub-blocks to obtain horizontal high-frequency profiles and vertical high-frequency profiles respectively. These profiles are used to identify local protrusions that smoothly change along the edge and deviate from the skeleton, providing a basis for subsequent attachment target identification.

[0115] Protrusion detection and angle deviation analysis are performed based on horizontal and vertical high-frequency profiles to determine the target candidate pixel set and the pure edge pixel set. The Frangi blood vessel enhancement filter based on the eigenvalue of the Hessian matrix can be used for protrusion detection algorithm. If the high-frequency amplitude on the vertical profile is a local maximum and its direction differs greatly from the tangential direction of the skeleton, then the position is judged as an attachment target candidate. If a position is continuous and gentle along the skeleton profile, then it is judged as a pure edge.

[0116] The high-frequency sub-blocks are structurally adjusted based on the target candidate pixel set and the pure edge pixel set, as follows:

[0117]

[0118] in, High-frequency sub-blocks representing structural adjustments This represents a high-frequency sub-block, where d represents the direction index of the high-frequency sub-block, including horizontal, vertical, and diagonal directions. and These represent the target candidate pixel set and the pure edge pixel set, respectively. and Let represent the adjustment coefficients for the target candidate pixel set and the pure edge pixel set, respectively. These coefficients are adaptively determined based on the blood vessel similarity metric output by the Frangi filter. The ratio of the mean Frangi response of pixels within the target candidate pixels to the mean global Frangi response is used, and this ratio is mapped to a preset range of enhancement coefficients via a Sigmoid function. For a set of pixels with pure edges, the high-frequency response uniformity along the edge direction is analyzed. The degree of suppression is determined based on the response variation coefficient; the more uniform the response, the higher the degree of suppression. ;

[0119] Inverse wavelet reconstruction is performed on low-frequency sub-blocks and high-frequency sub-blocks with structural adjustments, and coordinate mapping is performed on the original remote sensing image to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image.

[0120] By extracting the main edge direction and skeleton from the gradual transformation component, and using this as a reference to construct strip-shaped sampling areas on both sides, all subsequent high-frequency analyses are closely aligned with the real geometric contour. Then, by combining the directional information of the wavelet subband for profile calculation, the high-frequency textures and noise near the edge are no longer averaged out, but are confined to local bands strongly correlated with the skeleton. This establishes a dual constraint of space and direction to distinguish the attached target from the simple boundary.

[0121] By simultaneously calculating high-frequency profiles along the skeleton direction and perpendicular to the skeleton direction within the strip region, and unifying them with the direction field derived from the gradient component into the same reference system, local high-frequency protrusions are not only considered as intensity anomalies, but must also show a significant deviation from the main body orientation on the normal profile. This intensity-direction coupling criterion makes weak targets close to the edge statistically distanced from continuous and smooth boundary responses, thus achieving the decomposition of the target and background geometry.

[0122] By inputting the above profile analysis results into a filter based on Hessian features, and dividing candidate target pixels and pure edge pixels on this basis, and then writing it back to the wavelet high-frequency sub-band for structural adjustment, the enhancement or suppression is no longer directed towards the entire sub-band, but is limited to the classification results within the strip area. In this way, candidate targets adjacent to the main edge are synergistically amplified in multiple sub-bands, while the contour itself is synchronously weakened in multiple directions, achieving differentiated shaping of edge and attached point targets from the frequency domain structure.

[0123] By adaptively determining the adjustment coefficients for candidate targets and pure edges based on the filter response statistics and coefficient of variation, and then scaling the high-frequency components in different directions point by point, a set of reweighting factors consistent with the skeleton geometry, local high-frequency morphology and global statistics is formed, so that the high-frequency structure of the same position in each directional subband is collaboratively reconstructed.

[0124] S4. For the multi-scale response fields of structural edge blocks and smooth background blocks, the corresponding block-level response sets are determined by the spatial mapping relationship of image blocks in the original remote sensing image. The pixel-level robust response map is calculated by the hyperbolic tangent function, and the relationship between adjacent nodes is defined to form a set of multiple weak target trajectories. The final weak target detection set is selected by the NMS global nonmaximum suppression algorithm.

[0125] Preferably, for the multi-scale response fields of structural edge blocks and smooth background blocks, the corresponding block-level response sets are determined by the spatial mapping relationship of image blocks in the original remote sensing image, including:

[0126] The intersection of the multi-scale response fields of the structural edge blocks and the multi-scale response fields of the smooth background blocks forms the enhanced block response set;

[0127] By utilizing the spatial mapping relationship between each image block in the enhanced block response set and the original remote sensing image, the corresponding block-level response set for each pixel in the original remote sensing image at each scale is obtained;

[0128] The average response and variance within the corresponding block-level response set are calculated, pixel-level consistency is calculated, and a pixel-level robust response map is calculated using the hyperbolic tangent function, as shown below:

[0129]

[0130]

[0131] in, Indicates pixel-level consistency. This represents the average response of pixels at scale s. The variance response of pixels at scale s. This represents a pixel-level robust response map.

[0132] By simultaneously calculating the mean and variance of the response set of the same pixel across multiple scales and multiple overlapping blocks, and then using the two to form a consistency metric and mapping it through the hyperbolic tangent function, whether a pixel is considered a reliable candidate depends not only on the strength of the single-scale response, but also on the stability across scales and blocks. This compresses the originally independent multi-scale block responses into a robust response map that contains both strength and stability, providing a cleaner starting point for subsequent trajectory construction.

[0133] Furthermore, pixel-level robust response maps are calculated using the hyperbolic tangent function, and the relationships between adjacent nodes are defined to form a set of multiple weak target trajectories, including...

[0134] Based on the pixel-level robust response map, the scale is used as the third dimension for dimensional expansion, and the relationship between adjacent nodes is defined (two nodes are considered adjacent if and only if their scales are adjacent and their Euclidean distance in the image space is not greater than a preset spatial proximity threshold), expressed as:

[0135]

[0136]

[0137] in, A three-dimensional response plot representing dimensional expansion. Indicates the binary adjacency indicator value. This represents a preset spatial proximity threshold, measured in pixels. and The coordinates of the three-dimensional response plot are indices 1 and 2;

[0138] Based on the definition of adjacent node relationships, each node in the pixel-level robust response map is recursively connected in direction, forming a set of multiple weak target trajectories. Each trajectory is represented by three-dimensional response map coordinates, as follows:

[0139]

[0140] in, This represents a set of trajectories for multiple weak targets. This represents the number of nodes in the j-th trajectory, where k is the trajectory index.

[0141] By introducing scale as a third coordinate on the basis of robust response, the response of each pixel at different scales is regarded as a three-dimensional node. At the same time, dual constraints of scale adjacency and spatial proximity are applied, allowing node connection only under these strict conditions. This allows the continuous traces of weak targets on the scale axis and spatial axis to be automatically connected into a trajectory, while even if the response of occasional noise points is large at individual scales, it is difficult for them to form long chains in the three-dimensional map, thus completing a round of structure screening at the trajectory level.

[0142] By combining the directional recursive connection of 3D nodes with the robust response map obtained from the previous set of intersection blocks, each trajectory is accumulated from multi-scale, multi-block, and highly consistent nodes. Even if a weak target leaves only a local faint trace at a single scale, as long as it remains connected at several adjacent scales and spatial locations, it can achieve significant accumulation on the trajectory. Meanwhile, isolated noise and local textures are difficult to extend continuously in the 3D structure, thus achieving a leap from single-point response enhancement to multi-scale chain evidence.

[0143] Furthermore, the final weak target detection set is selected using the NMS global nonmaximum suppression algorithm, including:

[0144] For a set of multiple weak target trajectories, redundancy is removed using the NMS global nonmaximum suppression algorithm to obtain the final weak target detection set.

[0145] By applying global nonmaximum suppression at the trajectory level, instead of making local comparisons of single pixels or single scales, the competition is based on the overall performance of the entire trajectory in three-dimensional space. Combined with the aforementioned screening and consistency compression, redundant trajectories that are spatially close and highly overlapping in scale are automatically merged or deleted. The final set of weak targets combines precise spatial positioning, multi-scale evidence support, and low redundancy. This joint design, from block-level intersection to pixel consistency to three-dimensional trajectory plus global suppression, forms a detection chain that adds evidence and removes noise for weak targets step by step. This effect cannot be directly achieved by any single step or combination of any parts.

[0146] This embodiment also provides a remote sensing weak target detection system based on feature adaptation, including,

[0147] The remote sensing image acquisition module receives preprocessed remote sensing image data and divides the sliding window according to a preset block size and step size.

[0148] The structural detail complexity module smooths image patches to obtain gradually varying components and detail residual components, and constructs structural indices and detail complexity based on the continuous pixel chain length and residual energy.

[0149] The inner and outer ring contrast enhancement module is used to perform two-dimensional wavelet decomposition and joint high-frequency amplitude construction on the smooth background block. Based on the brightness contrast of the inner and outer rings and the joint high-frequency response, pixel-level high-frequency enhancement coefficients are generated and inverse wavelet reconstruction is completed to form a multi-scale response field of the smooth background block.

[0150] The high-frequency structure adjustment module is used to construct a strip-shaped neighborhood set based on the main edge skeleton, calculate the high-frequency profile along the skeleton direction and normal, and perform protrusion detection and angle deviation analysis. It adaptively adjusts the high-frequency sub-blocks and combines them with the low-frequency sub-blocks to inverse wavelet reconstruction to obtain the multi-scale response field of the structural edge block.

[0151] The pixel-level robust response calculation module is used to aggregate the multi-scale responses of smooth background blocks and structural edge blocks based on the mapping relationship of image blocks in the original image, construct a pixel-level block-level response set and calculate the average response and variance, and generate a pixel-level robust response map through the hyperbolic tangent function.

[0152] The weak target trajectory generation module is used to construct a three-dimensional response map by taking scale as the third dimension, defining node connectivity based on spatial proximity and scale adjacency, and performing directional recursive connection to form a set of multiple weak target trajectories.

[0153] The weak target detection result output module is used to perform a global nonmaximum suppression algorithm on the weak target trajectory set to filter out redundant trajectories and output the final weak target detection result set.

[0154] This embodiment also provides a computer device applicable to the feature-adaptive remote sensing weak target detection method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the feature-adaptive remote sensing weak target detection method proposed in the above embodiment.

[0155] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0156] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the feature-adaptive remote sensing weak target detection method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0157] In summary, this invention uses both structural and detail complexity indicators within the same framework, and requires that the combination of these two indicators with their respective thresholds meet specific conditions before a block can be classified as background or edge. This prevents block classification from being unidirectionally dominated by a single indicator. By combining the nonlinear normalized enhancement coefficients of the inner and outer loops with the high-frequency joint response to form a pixel-level high-frequency enhancement factor, and then applying it to high-frequency sub-blocks in each direction for inverse wavelet reconstruction, the brightness enhancement in the spatial domain is strictly constrained by two sources: ring contrast and multi-sub-band consistency. By adaptively determining the adjustment coefficients for candidate targets and pure edges based on the filter response statistics and coefficient of variation, and then scaling the high-frequency components in different directions point by point, a set of reweighting factors consistent with skeleton geometry, local high-frequency morphology, and global statistics is formed, enabling the high-frequency structure of the same location in each directional sub-band to be collaboratively reconstructed.

[0158] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A remote sensing method for detecting small targets based on feature adaptation, characterized in that, include: Based on remote sensing image data, a sliding window is divided. Through smoothing, the large-scale background structure is treated as a slowly varying component map. The maximum length of the continuous pixel chain is used as the main direction continuity index for weighted combination and quantification of structural index. The residual energy of the detail residual component is statistically analyzed and weighted combination with the residual energy statistics to quantify the detail complexity. Combined with the structural index, the image patch is calibrated to obtain smooth background patch and structural edge patch. Two-dimensional wavelet decomposition is performed on the smooth background block to form a joint high-frequency amplitude map and a binary mask is constructed to obtain the high-frequency joint response value. The corresponding image block is decomposed into inner and outer ring pixel sets, and the brightness contrast baseline between the inner and outer rings is analyzed. Pixel-level high-frequency enhancement coefficients are constructed based on the high-frequency joint response value. Combined with the high-frequency sub-blocks of the two-dimensional wavelet decomposition, the contrast-enhanced high-frequency sub-blocks are obtained. Inverse wavelet reconstruction is performed, and spatial mapping is performed on the original remote sensing image to obtain the multi-scale response field of the smooth background block. By setting the banded neighborhood bandwidth and spatial sampling of the structural edge blocks, a banded neighborhood set along both sides of the skeleton is obtained. The direction and normal of the high-frequency sub-blocks corresponding to the banded neighborhood set are analyzed, and protrusion detection and angle deviation analysis are performed. The high-frequency sub-blocks are structurally adjusted. Inverse wavelet reconstruction is performed on the low-frequency sub-blocks and the high-frequency sub-blocks with structural adjustments to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image. For the multi-scale response fields of structural edge blocks and smooth background blocks, the corresponding block-level response sets are determined by the spatial mapping relationship of image blocks in the original remote sensing image. The pixel-level robust response map is calculated by the hyperbolic tangent function, and the relationship between adjacent nodes is defined to form a set of multiple weak target trajectories. The final weak target detection set is selected by the NMS global nonmaximum suppression algorithm.

2. The remote sensing weak target detection method based on feature adaptation as described in claim 1, characterized in that: The maximum length of the statistical continuous pixel chain is used as a weighted combination of structural indices based on the main direction continuity index. Residual energy statistics are performed on the detail residual components, and then weighted and combined with these residual energy statistics to quantify the detail complexity. Combined with the structural indices, image blocks are calibrated to obtain smooth background blocks and structural edge blocks, including... For preprocessed remote sensing image data, based on a predefined set of multi-scale block sizes and overlap step size parameters, the remote sensing image data is divided into a sliding window according to the block size and step size at each scale to obtain a set of multi-scale image blocks, where each image block corresponds to a spatial region in the remote sensing image, and is based on the starting coordinates and scale index of the image block in the remote sensing image. For each image patch, a convolution operation is performed using a preset smoothing kernel to determine the large-scale background structure within the patch as the slowly varying component map, and the difference between the image patch and the slowly varying component map is calculated as the detail residual component. Gradient components are calculated for the horizontal and vertical directions of the detailed residual components. Euclidean norm and arctangent operations are performed on the gradient components to obtain the local edge intensity and gradient direction distribution, respectively. The gradient direction distribution is quantized in discrete direction intervals, and the maximum length of the continuous pixel chain along the direction in each direction interval is counted as the main direction chain length. The number of pixels of the continuous pixel chain with the maximum length is marked as the main direction continuity index. The average gradient magnitude within the image patch is calculated for local edge intensity, and then weighted and combined with the main direction continuity index to quantify the structural index. For the detailed residual components, the average of the squared residuals is calculated, and residual energy statistics are performed. Based on the historical isolation threshold, the isolated extreme values ​​of the detailed residual components are detected to obtain the number of isolated strong points. The number of isolated strong points is then weighted and combined with the residual energy statistics to quantify the detail complexity. Based on the calibrated structural index data and residual energy statistics, structural thresholds and complexity thresholds are determined respectively. The structural thresholds and complexity thresholds are adaptively determined based on the global feature statistics of the current image. If the structural index data is less than the structural threshold and the residual energy statistics are less than the complexity threshold, the image patch is labeled as a smooth background patch. If the structural index data is greater than or equal to the structural threshold and the residual energy statistics are less than the complexity threshold, the image patch is labeled as a structural edge patch.

3. The remote sensing weak target detection method based on feature adaptation as described in claim 2, characterized in that: The process involves constructing pixel-level high-frequency enhancement coefficients based on high-frequency joint response values, combining these with high-frequency sub-blocks from two-dimensional wavelet decomposition to obtain contrast-enhanced high-frequency sub-blocks, performing inverse wavelet reconstruction, and spatially mapping them onto the original remote sensing image to obtain a multi-scale response field for a smooth background block. For image blocks marked as smooth background blocks, two-dimensional wavelet decomposition is performed on the data within the block based on orthogonal wavelet basis functions to obtain low-frequency sub-blocks and high-frequency sub-blocks in the horizontal, vertical and diagonal directions. These are used as local high-frequency amplitudes in the corresponding directions and form a joint high-frequency amplitude map. At the same time, the joint high-frequency amplitude is defined as the infimum of the amplitudes in the three directions. For the joint high-frequency amplitude map, pixel-level comparison and binary mask construction are performed based on the joint amplitude statistics of the corresponding image blocks and the preset high-frequency joint response threshold to obtain the high-frequency joint response value; For remote sensing data of weak targets detected in history, the inner and outer ring radius thresholds of image blocks are defined. The image blocks marked as smooth background blocks are divided into inner ring pixel sets and outer ring pixel sets according to the inner and outer ring radius thresholds. The regional mean of the low-frequency sub-blocks of the two sets is calculated as the inner ring average brightness and the outer ring average brightness, respectively. The difference between the inner ring and outer ring average brightness is used as the brightness comparison baseline. Nonlinear normalization is performed based on the brightness contrast baseline, the inner ring enhancement coefficient and the outer ring enhancement coefficient are calculated respectively, and the pixel-level high-frequency enhancement coefficient is constructed based on the high-frequency joint response value. The product of the high-frequency enhancement coefficient and the high-frequency sub-block is used as the high-frequency sub-block after contrast enhancement. The frequency domain reconstruction operation is performed by wavelet inverse transform to obtain the enhanced spatial domain response map marked as the smooth background block. The map is then spatially mapped in the original remote sensing image to obtain the multi-scale response field of the smooth background block in the original remote sensing image.

4. The remote sensing weak target detection method based on feature adaptation as described in claim 3, characterized in that: The analysis of the direction and normal of the high-frequency sub-blocks corresponding to the strip neighborhood set is performed, and protrusion detection and angle deviation analysis are conducted. Structural adjustments are then made to the high-frequency sub-blocks. Inverse wavelet reconstruction is performed on the low-frequency sub-blocks and the structurally adjusted high-frequency sub-blocks to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image, including... For image blocks marked as structural edge blocks, the edge orientation field of the main edge direction and the set of main edge skeleton pixels are determined by gradient calculation and orientation aggregation through the slowly varying component map. Based on the main edge skeleton pixel set and the edge direction field, the tangential vector and normal vector at the pixel are determined, and the band neighborhood bandwidth is set and spatial sampling is performed to obtain the band neighborhood set along both sides of the skeleton. Based on the strip neighborhood set, high-frequency profiles are calculated along the skeleton direction and normal according to the corresponding high-frequency sub-blocks, and horizontal and vertical high-frequency profiles are obtained respectively. Protrusion detection and angle deviation analysis are performed based on horizontal and vertical high-frequency profiles to determine the target candidate pixel set and the pure edge pixel set; The structure of high-frequency sub-blocks is adjusted based on the target candidate pixel set and the pure edge pixel set; Inverse wavelet reconstruction is performed on low-frequency sub-blocks and high-frequency sub-blocks with structural adjustments, and coordinate mapping is performed on the original remote sensing image to obtain the multi-scale response field of the structural edge blocks of the original remote sensing image.

5. The remote sensing weak target detection method based on feature adaptation as described in claim 4, characterized in that: The multi-scale response fields for structural edge blocks and smooth background blocks are determined by using the spatial mapping relationship between image blocks in the original remote sensing image to determine the corresponding block-level response set, including: The intersection of the multi-scale response fields of the structural edge blocks and the multi-scale response fields of the smooth background blocks forms the enhanced block response set; By utilizing the spatial mapping relationship between each image block in the enhanced block response set and the original remote sensing image, the corresponding block-level response set for each pixel in the original remote sensing image at each scale is obtained; The average response and variance within the corresponding block-level response set are calculated, pixel-level consistency is calculated, and pixel-level robust response maps are calculated using the hyperbolic tangent function.

6. The remote sensing weak target detection method based on feature adaptation as described in claim 5, characterized in that: The process involves calculating a pixel-level robust response map using the hyperbolic tangent function and defining the relationships between adjacent nodes to form a set of multiple weak target trajectories, including... Based on the pixel-level robust response map, the scale is used as the third dimension coordinate for dimensional expansion, and the relationship between adjacent nodes is defined. Based on the definition of adjacent node relationships, each node of the pixel-level robust response map is recursively connected in a direction to form a set of multiple weak target trajectories, where each trajectory is represented by the coordinates of the three-dimensional response map.

7. The remote sensing weak target detection method based on feature adaptation as described in claim 6, characterized in that: The final weak target detection set, selected using the NMS global nonmaximum suppression algorithm, includes: For a set of multiple weak target trajectories, redundancy is removed using the NMS global nonmaximum suppression algorithm to obtain the final weak target detection set.

8. A feature-adaptive remote sensing target detection system, based on the feature-adaptive remote sensing target detection method according to any one of claims 1 to 7, characterized in that: include, The remote sensing image acquisition module receives preprocessed remote sensing image data and divides the sliding window according to a preset block size and step size. The structural detail complexity module smooths image patches to obtain gradually varying components and detail residual components, and constructs structural indices and detail complexity based on the continuous pixel chain length and residual energy. The inner and outer ring contrast enhancement module is used to perform two-dimensional wavelet decomposition and joint high-frequency amplitude construction on the smooth background block. Based on the brightness contrast of the inner and outer rings and the joint high-frequency response, pixel-level high-frequency enhancement coefficients are generated and inverse wavelet reconstruction is completed to form a multi-scale response field of the smooth background block. The high-frequency structure adjustment module is used to construct a strip-shaped neighborhood set based on the main edge skeleton, calculate the high-frequency profile along the skeleton direction and normal, and perform protrusion detection and angle deviation analysis. It adaptively adjusts the high-frequency sub-blocks and combines them with the low-frequency sub-blocks to inverse wavelet reconstruction to obtain the multi-scale response field of the structural edge block. The pixel-level robust response calculation module is used to aggregate the multi-scale responses of smooth background blocks and structural edge blocks based on the mapping relationship of image blocks in the original image, construct a pixel-level block-level response set and calculate the average response and variance, and generate a pixel-level robust response map through the hyperbolic tangent function. The weak target trajectory generation module is used to construct a three-dimensional response map by taking scale as the third dimension, defining node connectivity based on spatial proximity and scale adjacency, and performing directional recursive connection to form a set of multiple weak target trajectories. The weak target detection result output module is used to perform a global nonmaximum suppression algorithm on the weak target trajectory set to filter out redundant trajectories and output the final weak target detection result set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the feature-adaptive remote sensing weak target detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the feature-adaptive remote sensing weak target detection method according to any one of claims 1 to 7.

Images