A deep learning-based marine remote sensing target recognition method

Abstract

This invention discloses a deep learning-based method for marine remote sensing target recognition, comprising the following steps: acquiring SAR images and generating incident angle, azimuth, sea state, land-sea mask, and range fields to form an observation field; constructing a coordinate distortion field under constraints of land-sea, coastline, and shipping channels, and completing the reparameterization of the observation field; inputting the distortion domain observation field into a dual-spectral routing neural operator network to decompose sea clutter and target residual spectra; applying differential spectral kernels to the two types of spectra and performing conditional routing fusion to obtain an operator mapping feature field; generating probability, centrality, and boundary response, and extracting center points; generating target instance regions and combining inverse mapping to output the recognition results. This invention, through sea state conditional spectrum modeling and a deep learning recognition method under coordinate constraints, achieves robust detection and fine contour reconstruction of targets in complex sea areas, possessing advantages such as strong resistance to sea clutter, high positioning accuracy, and good recognition stability.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and in particular to a deep learning-based method for identifying marine remote sensing targets. Background Technology

[0002] Synthetic Aperture Radar (SAR) images for marine remote sensing are widely used for identifying maritime targets such as ships and floating objects due to their all-weather, day-and-night imaging capabilities. Existing techniques typically use SAR images as input, employing convolutional neural networks or attention networks for feature extraction, and combining spatial transformations, pyramid multi-scale fusion, or instance segmentation heads to output target categories and region contours. Other approaches introduce sea clutter suppression and texture enhancement in the frequency or statistical domains to reduce the impact of sea surface speckle and stripe interference on detection. In engineering implementation, these methods are mostly end-to-end learning-based, relying on data-driven generalization under complex sea conditions and imaging geometric differences.

[0003] However, existing technologies still have significant shortcomings in complex sea conditions, nearshore areas, and shipping lanes: On the one hand, there is a lack of a spectral decomposition mechanism that jointly models observation conditions such as incident angle, observation azimuth, and sea conditions with spectral statistics, making it difficult to stably decouple sea clutter background and target residuals at the frequency band level, resulting in high false alarm rates and weak targets being easily submerged; on the other hand, there is a lack of coordinate reparameterization and inverse mapping closed loops constrained by land-sea mask, shoreline distance field, and shipping lane distance field, making it difficult to simultaneously satisfy nearshore boundary consistency and shipping lane interference suppression, and problems such as contour breaks, boundary drift, and inconsistent coordinate recovery often occur in the instance area.

[0004] Therefore, how to provide a deep learning-based method for marine remote sensing target recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for identifying marine remote sensing targets. This invention achieves robust detection and fine contour reconstruction of targets in complex sea areas through sea state-conditional spectrum modeling and deep learning-based identification under coordinate constraints. It has the advantages of strong resistance to sea clutter, high positioning accuracy, and good identification stability.

[0006] A deep learning-based method for marine remote sensing target identification according to an embodiment of the present invention includes the following steps: Acquire marine remote sensing synthetic aperture radar images and generate incident angle field, observation azimuth coding field, sea state level field, land-sea mask, shoreline distance field and channel distance field that are spatially aligned with them to form a SAR observation field; Based on the land-sea mask, shoreline distance field and channel distance field, a coordinate distortion field constrained by shoreline and channel is constructed. The coordinate reparameterization of the SAR observation field is performed to obtain the distorted domain SAR observation field and the corresponding inverse coordinate mapping relationship. The distorted domain SAR observation field is input into the sea state conditional bispectral routing neural operator network, and the frequency domain transformation is performed to obtain the observation spectrum. A sea clutter subspace basis is constructed, and the observation spectrum is decomposed by spectral domain projection to obtain the sea clutter component spectrum and the target residual component spectrum. Background modeling spectral kernels are applied to the spectrum of sea clutter components, and target enhancement spectral kernels are applied to the spectrum of target residual components. Frequency band segmented routing coefficients are generated based on the sea state level field, the incident angle field, and the observation azimuth coding field. Sea state conditional routing fusion is performed on the bispectral convolution results to obtain the operator mapping feature field. Based on the operator mapping feature field, a target probability response field, a centrality response field, and a boundary response field are generated, and a set of candidate target center points is formed using the centrality response field. Using the set of candidate target center points as seeds, a set of target instance regions is generated on the target probability response field. The boundary response field is used to refine the contours of the target instance region set. The results of marine remote sensing target recognition are obtained by combining the inverse coordinate mapping relationship.

[0007] Optionally, the generation of the SAR observation field specifically includes: The system acquires marine remote sensing synthetic aperture radar (SAR) images and, based on imaging geometry, observation azimuth information, sea state data corresponding to the observation time, and prior data on coastlines and waterways, generates an incident angle field, observation azimuth coding field, sea state level field, land-sea mask, coastline distance field, and waterway distance field that are spatially aligned with the marine remote sensing SAR images. These fields are then stitched together with the marine remote sensing SAR images by channel to form a SAR observation field.

[0008] Optionally, the generation of the distorted domain SAR observation field and the corresponding inverse coordinate mapping relationship specifically includes: Based on the pixel grid of the SAR observation field, the land and sea mask, shoreline distance field and channel distance field are read and uniformly cropped to the same coverage and resolution as the SAR observation field to generate a geometrically constrained input set; The distortion intensity field is generated based on the shoreline distance field and the channel distance field. Under the constraint of the land-sea mask, the distortion intensity of non-sea area pixels is suppressed to obtain the effective distortion intensity field of the sea area. Calculate the torsion direction field on the effective torsion intensity field of the sea area, and combine them to generate the coordinate displacement field; Spatial smoothing and boundary constraint processing are performed on the coordinate displacement field to generate a constrained coordinate displacement field that satisfies continuity and local monotonicity, and a coordinate distortion field is constructed. Based on the coordinate distortion field, coordinate reparameterization is performed on the SAR observation field to obtain the distorted domain SAR observation field; Perform a reversibility check on the coordinate distortion field and generate an inverse coordinate mapping relationship.

[0009] Optionally, the generation of the sea clutter component spectrum and the target residual component spectrum specifically includes: The distorted domain SAR observation field is input into the frequency domain mapping and spectral domain representation structure of the sea state conditional bispectral routing neural operator network, and frequency domain transformation is performed to obtain the set of observation spectra. The sea state conditional bispectral routing neural operator network includes a frequency domain mapping and spectral domain representation structure, a sea clutter subspace basis generation structure, a spectral domain projection decomposition structure, a bispectral convolution processing structure, a sea state conditional frequency band segmentation routing structure, a spectral domain to spatial domain back mapping structure, and a three-response field joint prediction structure. Based on the observation spectrum set, the observation spectrum statistics are generated and spatially and frequency band aligned with the sea state level field, the incident angle field, and the observation azimuth coding field to form the input set for generating the sea clutter sub-space basis. The input set for generating the sea clutter space basis is input into the sea clutter space basis generating structure to generate the sea clutter space basis; The observed spectrum is input into the spectral domain projection decomposition structure, and the spectral domain projection is performed using the sea clutter subspace basis to obtain the sea clutter component spectrum. The target residual component spectrum is obtained by the spectral domain difference between the observed spectrum and the sea clutter component spectrum. A consistency check is performed on the spectrum of the sea clutter component and the spectrum of the target residual component to form a set of availability tags.

[0010] Optionally, the generation of the operator mapping feature field specifically includes: Read the spectrum of sea clutter component and the spectrum of target residual component, and perform masking on the frequency band positions that do not meet the consistency check according to the availability tag set. Unify the frequency band division rules and spectral domain grid representation to be consistent with the observed spectrum, and generate the input set for spectral kernel action. Background modeling spectral kernels are applied to the sea clutter component spectrum to obtain the background modeling spectral response, and target enhancement spectral kernels are applied to the target residual component spectrum to obtain the target enhancement spectral response. The background modeling spectral response and the target enhancement spectral response are input into the bispectral convolution processing structure to obtain the bispectral convolution result; The sea state level field, the incident angle field, and the observation azimuth coding field are input into the sea state conditional frequency band segmented routing structure to generate frequency band segmented routing coefficients. The frequency band segmented routing coefficients are then aligned with the bispectral convolution result to obtain a routable bispectral convolution result. Perform sea state-conditional route fusion on the routable bispectral convolution result to obtain the route fusion spectrum response; Perform cross-band consistency processing on the route fusion spectrum response to obtain a consistent spectrum response; The uniformized spectral response is input to the spectral domain and then back-mapped to the spatial domain. This back-mapping is performed to obtain the operator-mapped feature field.

[0011] Optionally, the generation of the candidate target center point set specifically includes: Read the operator mapping feature field, perform channel normalization and spatial continuity shaping, perform parallel feature aggregation in different spatial neighborhoods, form operator feature representations corresponding to different target scales and spatial context ranges, and generate a multi-scale operator feature set; The multi-scale operator feature set is input into the three-response field joint prediction structure to perform scale-wise fusion, and outputs the target probability response field, centrality response field and boundary response field that are consistent with the pixel grid of the distorted domain SAR observation field respectively. Apply sea area consistency constraints to the target probability response field to obtain the sea area-constrained target probability response field; The centrality response field is subjected to centrality enhancement processing to obtain a centrality-enhanced centrality response field; A set of candidate target center points is generated based on the centrality-enhanced centrality response field and the sea area-constrained target probabilistic response field. The boundary response field is subjected to boundary consistency processing including thinning and fracture repair to form a continuous boundary clue set, which is then spatially aligned and bound with the probability response field of the marine constrained target and the set of candidate target center points.

[0012] Optionally, the generation of the marine remote sensing target identification result specifically includes: Based on the set of candidate target center points, perform a center point-guided region growth operation on the probability response field of constrained targets in the sea area to obtain the initial set of target instances; Perform region consistency correction on the initial region set of the target instance to obtain the candidate region set of the target instance; Based on the continuous set of boundary clues, instance boundary constraints are applied to the set of candidate regions of target instances to obtain a set of target instance regions with boundary constraints. Perform instance-level overlap resolution on the set of target instance regions with boundary constraints to obtain a set of deduplicated target instance regions; Generate a set of distorted domain target instance masks based on the set of deduplicated target instance regions; Based on the inverse coordinate mapping relationship, the set of target instance masks in the distorted domain is subjected to inverse coordinate mapping to obtain the set of target instance masks in the original coordinate domain; Marine remote sensing target recognition results are generated based on the original coordinate domain target instance mask set.

[0013] The beneficial effects of this invention are: This invention explicitly constructs a SAR observation field at the synthetic aperture radar (SAR) observation level, including an incident angle field, an observation azimuth coding field, a sea state level field, and information on shoreline and channel constraints. It also introduces a coordinate distortion and inverse coordinate mapping closed loop based on a land-sea mask, ensuring that the target feature extraction process remains consistent with the actual imaging geometry and sea area spatial constraints. This fundamentally reduces the interference of complex near-shore and channel neighborhood backgrounds on target response. By combining frequency domain transformation, sea clutter subspace basis construction, and spectral domain projection decomposition, stable separation of sea clutter background and target residuals at the frequency band level is achieved, enabling weak targets to maintain a clear and continuous spectral domain representation even under complex sea conditions.

[0014] Furthermore, this invention employs sea state-conditional bispectral routing and multi-scale operator mapping feature modeling to explicitly incorporate the influence of sea state, incident angle, and observation azimuth on different frequency bands and spatial scales into the routing fusion process, forming a physically consistent operator mapping feature field in the spatial domain. Based on this, a joint prediction and constrained instance generation mechanism using the target probability response, centrality response, and boundary response three-response field is adopted to achieve a collaborative closed loop of target center localization, region growing, and boundary refinement. This significantly reduces false detection and false negative rates, improves the integrity and positioning stability of the target contour, and ensures the spatial consistency and engineering usability of the final identification results in the original coordinate domain. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based marine remote sensing target recognition method proposed in this invention; Figure 2 This is a schematic diagram of the spectral domain projection decomposition and frequency band segmentation routing fusion process of a deep learning-based marine remote sensing target recognition method proposed in this invention. Figure 3 This is a schematic diagram of the joint prediction of three response fields and target instance generation process of a deep learning-based marine remote sensing target recognition method proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 A deep learning-based method for marine remote sensing target identification includes the following steps: Acquire marine remote sensing synthetic aperture radar images and generate incident angle field, observation azimuth coding field, sea state level field, land-sea mask, shoreline distance field and channel distance field that are spatially aligned with them to form a SAR observation field; Based on the land-sea mask, shoreline distance field and channel distance field, a coordinate distortion field constrained by shoreline and channel is constructed. The coordinate reparameterization of the SAR observation field is performed to obtain the distorted domain SAR observation field and the corresponding inverse coordinate mapping relationship. The distorted domain SAR observation field is input into the sea state conditional bispectral routing neural operator network, and the frequency domain transformation is performed to obtain the observation spectrum. A sea clutter subspace basis is constructed, and the observation spectrum is decomposed by spectral domain projection to obtain the sea clutter component spectrum and the target residual component spectrum. Background modeling spectral kernels are applied to the spectrum of sea clutter components, and target enhancement spectral kernels are applied to the spectrum of target residual components. Frequency band segmented routing coefficients are generated based on the sea state level field, the incident angle field, and the observation azimuth coding field. Sea state conditional routing fusion is performed on the bispectral convolution results to obtain the operator mapping feature field. Based on the operator mapping feature field, a target probability response field, a centrality response field, and a boundary response field are generated, and a set of candidate target center points is formed using the centrality response field. Using the set of candidate target center points as seeds, a set of target instance regions is generated on the target probability response field. The boundary response field is used to refine the contours of the target instance region set. The results of marine remote sensing target recognition are obtained by combining the inverse coordinate mapping relationship.

[0018] In this embodiment, the generation of the SAR observation field specifically includes: Acquire marine remote sensing synthetic aperture radar (SAR) images, and based on imaging geometry information, observation azimuth information, sea state data corresponding to the observation time, and prior data of coastline and waterway, generate an incident angle field, observation azimuth coding field, sea state level field, land-sea mask, coastline distance field, and waterway distance field that are spatially aligned with the marine remote sensing SAR images. These fields are then stitched together with the marine remote sensing SAR images by channel to form a SAR observation field. Specifically, based on imaging geometric information, the angle between the radar line of sight and the surface normal direction is calculated pixel by pixel on the image pixel grid to generate an incident angle field. Based on the observation azimuth information, the observation azimuth angle is generated within the image coverage area and converted into an observation azimuth coding field composed of sine and cosine components. The sea state data corresponding to the observation time of the marine remote sensing synthetic aperture radar image is obtained, mapped and aligned to the image geographic grid, and then discretized and labeled according to the preset level classification rules to generate a sea state level field. Based on the coastline prior data, a land-sea mask is generated in the image geographic coordinate system, and the shortest distance from each pixel position to the coastline is calculated under the constraint of the land-sea mask to generate a shoreline distance field. Based on the channel prior data, the channel pixel set is determined in the image geographic coordinate system, and the shortest distance from each pixel position to the channel is calculated to generate a channel distance field.

[0019] In this embodiment, the generation of the distorted domain SAR observation field and the corresponding inverse coordinate mapping relationship specifically includes: Based on the pixel grid of the SAR observation field, the land and sea mask, shoreline distance field and channel distance field are read and uniformly cropped to the same coverage and resolution as the SAR observation field to generate a geometrically constrained input set; The distortion intensity field is generated based on the shoreline distance field and the channel distance field. Under the constraint of the land-sea mask, the distortion intensity of non-sea area pixels is suppressed to obtain the effective distortion intensity field of the sea area. The value of the effective torsion intensity field of the sea area at each pixel position is obtained by weighting and superimposing the sea-land mask value, the exponential decay term of the shoreline distance field, and the exponential decay term of the channel distance field according to a preset weight. The exponential decay term uses the shoreline distance field value and the channel distance field value as independent variables and controls the decay rate through a preset decay control term. Calculate the torsion direction field on the effective torsion intensity field of the sea area, and combine them to generate the coordinate displacement field; The direction vector of the torsion direction field at each pixel position is determined by the spatial gradient of the effective torsion intensity field of the sea area at that pixel position. The displacement vector of the coordinate displacement field at each pixel position is obtained by multiplying the preset displacement scale term with the normalized result of the direction vector and multiplying it with the value of the effective torsion intensity field of the sea area at that pixel position. The normalization is determined by the L2 norm of the direction vector and the preset stability term to ensure the stability of the displacement calculation. Spatial smoothing and boundary constraint processing are performed on the coordinate displacement field to generate a constrained coordinate displacement field that satisfies continuity and local monotonicity, and a coordinate distortion field is constructed. The spatial smoothing is obtained by weighted aggregation of the neighborhood displacements of the coordinate displacement field using a preset spatial smoothing operator, and the boundary constraint is achieved by restricting the consistency of displacement amplitude and displacement direction near the boundary of the land-sea mask. The distorted coordinates of the coordinate distortion field at each pixel position are obtained by adding the original coordinates of the pixel position to the displacement vector of the constrained coordinate displacement field at that pixel position. Based on the coordinate distortion field, coordinate reparameterization is performed on the SAR observation field to obtain the distorted domain SAR observation field; The coordinate reparameterization is achieved by sampling the SAR observation field according to the distorted coordinates specified by the coordinate distortion field and obtaining the channel values ​​of each pixel position using a preset interpolation rule; Perform an invertibility check on the coordinate distortion field and generate an inverse coordinate mapping relationship; The reversibility verification is achieved by checking the mapping consistency of the distorted coordinates of the coordinate distortion field and performing backtracking correction on conflicting mapping positions. The inverse coordinate mapping relationship is obtained by solving the original coordinate position corresponding to the pixel position of each distortion domain based on the coordinate distortion field.

[0020] In this embodiment, the generation of the sea clutter component spectrum and the target residual component spectrum specifically includes: The distorted domain SAR observation field is input into the frequency domain mapping and spectral domain representation structure of the sea state conditional bispectral routing neural operator network, and frequency domain transformation is performed to obtain the set of observation spectra. The frequency domain transformation is achieved by performing a discrete frequency domain transformation on the distorted domain SAR observation field in the spatial dimension to obtain a complex representation of the spectral domain, and then further extracting the spectral domain amplitude and spectral domain phase. The sea state conditional bispectral routing neural operator network includes a frequency domain mapping and spectral domain representation structure, a sea clutter subspace basis generation structure, a spectral domain projection decomposition structure, a bispectral convolution processing structure, a sea state conditional frequency band segmentation routing structure, a spectral domain to spatial domain back mapping structure, and a three-response field joint prediction structure. Based on the observation spectrum set, the observation spectrum statistics are generated and spatially and frequency band aligned with the sea state level field, the incident angle field, and the observation azimuth coding field to form the input set for generating the sea clutter sub-space basis. The observed spectrum statistics include spectral power statistics, spectral kurtosis statistics, and spectral sparsity statistics obtained by frequency band aggregation. The frequency band alignment is obtained by segmenting the observed spectrum according to a preset frequency band division rule and performing statistical aggregation within each frequency band. Specifically, the spectral power statistics are obtained by aggregation of the squared amplitude of the spectral domain within the corresponding frequency band, the spectral kurtosis statistics are obtained by calculating the kurtosis of the amplitude distribution of the spectral domain within the corresponding frequency band, and the spectral sparsity statistics are obtained by aggregation of the sparsity measure of the amplitude distribution of the spectral domain within the corresponding frequency band. The input set for generating the sea clutter space basis is input into the sea clutter space basis generating structure to generate the sea clutter space basis; Specifically, the input set for generating the sea clutter subspace basis is aligned according to a preset frequency band division rule. Within each preset frequency band, a set of frequency band-level basis vectors is generated based on the sea state conditions corresponding to the sea state level field, the geometric conditions corresponding to the incident angle field, the observation direction conditions corresponding to the observation azimuth coding field, and the spectral statistics conditions corresponding to the observation spectral statistics. Normalization is performed to unify the energy scale of each basis vector, and orthogonalization is performed to eliminate the correlation between basis vectors, thus obtaining the sea clutter subspace basis for spectral domain projection decomposition. The observed spectrum is input into the spectral domain projection decomposition structure, and the spectral domain projection is performed using the sea clutter subspace basis to obtain the sea clutter component spectrum. The target residual component spectrum is obtained by the spectral domain difference between the observed spectrum and the sea clutter component spectrum. A consistency check is performed between the sea clutter component spectrum and the target residual component spectrum to form an availability tag set; The consistency verification includes spectral energy conservation verification and frequency band consistency verification. Spectral energy conservation verification is determined by performing convergence on the squared amplitudes of the spectral domain within the same frequency band and comparing whether the differences in the convergence results fall within a preset consistency threshold range. Frequency band consistency verification is determined by comparing the joint matching degree of the converged projection coefficient values ​​with the spectral power statistics and spectral kurtosis statistics within the same frequency band.

[0021] In this embodiment, the generation of the operator mapping feature field specifically includes: Read the spectrum of sea clutter component and the spectrum of target residual component, and perform masking on the frequency band positions that do not meet the consistency check according to the availability tag set. Unify the frequency band division rules and spectral domain grid representation to be consistent with the observed spectrum, and generate the input set for spectral kernel action. Background modeling spectral kernels are applied to the sea clutter component spectrum to obtain the background modeling spectral response, and target enhancement spectral kernels are applied to the target residual component spectrum to obtain the target enhancement spectral response. The background modeling spectral kernel forms a stable sea clutter background spectral shape by applying smoothing constraints to the frequency band energy distribution and spectral kurtosis distribution of the sea clutter component spectrum and suppressing narrowband peak responses. The target enhancement spectral kernel forms a significant target spectral shape by applying enhancement constraints to the local high-frequency texture response and azimuth strip response of the target residual component spectrum and suppressing low-frequency drift. The background modeling spectral response and the target enhancement spectral response are input into the bispectral convolution processing structure to obtain the bispectral convolution result; The bispectral convolution processing constructs independent sets of spectral domain convolution operators for the background modeling spectral response and the target enhancement spectral response in the frequency band dimension. The spectral domain convolution operators perform local weighted aggregation along the spectral domain amplitude dimension and the spectral domain phase dimension in each frequency band, and share the convolution kernel structure but not the convolution weights between different frequency bands. The spectral domain convolution performed on the background modeling spectral response is constrained by maintaining low-frequency continuity and smoothing energy within the frequency band, while the spectral domain convolution performed on the target enhancement spectral response is constrained by maintaining high-frequency texture and enhancing azimuth spectral peaks. This generates sea clutter suppression convolution results and target enhancement convolution results, respectively. The sea clutter suppression convolution results and target enhancement convolution results are then stacked after being aligned by frequency band to form a dual-spectral convolution result. The sea state level field, the incident angle field, and the observation azimuth coding field are input into the sea state conditional frequency band segmented routing structure to generate frequency band segmented routing coefficients. The frequency band segmented routing coefficients are then aligned with the bispectral convolution result to obtain a routable bispectral convolution result. The frequency band segmentation routing coefficients are generated separately into three frequency band groups according to a preset frequency band division rule. The first frequency band group extracts the values ​​of the sea state level field and the incident angle field at the corresponding spatial positions of the frequency band group, performs normalization and joint mapping processing to generate a frequency band weight descriptor characterizing the combined influence of sea state roughness and incident geometry as the frequency band segmentation routing coefficient of the first frequency band group. The second frequency band group extracts the values ​​of the incident angle field and the observation azimuth coding field at the corresponding spatial positions of the frequency band group, performs direction consistency coding and angle coupling mapping processing to generate a frequency band weight descriptor characterizing the coupling relationship between imaging direction anisotropy and incident geometry as the frequency band segmentation routing coefficient of the second frequency band group. The third frequency band group extracts the values ​​of the sea state level field and the observation azimuth coding field at the corresponding spatial positions of the frequency band group, performs joint mapping processing of sea state condition modulation and observation direction modulation to generate a frequency band weight descriptor characterizing the directional modulation effect of sea clutter as the frequency band segmentation routing coefficient of the third frequency band group. Perform sea state-conditional route fusion on the routable bispectral convolution result to obtain the route fusion spectrum response; The sea state conditional route fusion performs coefficient weighting, segmented gating, and intra-group normalization fusion on the sea clutter suppression convolution result and the target enhancement convolution result within the frequency band group of the bispectral convolution result, outputting the route fusion spectrum response while maintaining the fusion independence between each frequency band group; Perform cross-band consistency processing on the route fusion spectrum response to obtain a consistent spectrum response; The cross-band uniformity processing includes applying continuity constraints to the fusion boundary of adjacent frequency band groups, applying spectral energy upper limit constraints to frequency band groups corresponding to high sea state levels, and applying spectral texture stability constraints to frequency band groups corresponding to high incident angles, outputting a uniform spectral response that satisfies the continuity of frequency band boundaries and controllable amplitude. The uniformized spectral response is input from the spectral domain to the spatial domain back mapping structure, and back mapping is performed to obtain the operator-mapped feature field. The back-mapping process restores the complete spectral domain representation by performing frequency band-level reconstruction on the uniform spectral response according to a preset frequency band division rule. While maintaining the consistency between the amplitude distribution and phase distribution of the spectral domain, spectral domain reconstruction processing is performed, and continuity constraints are applied at the frequency band boundaries. Spatial domain back-mapping operation is then performed on the reconstructed spectral domain representation. Spatial consistency constraints are introduced to suppress local discontinuous responses caused by spectral domain processing, and the edge and texture responses of the target region are kept stable. This results in a spatial domain response distribution consistent with the spatial resolution of the distorted domain SAR observation field, which is then integrated to form an operator-mapped feature field.

[0022] In this embodiment, the generation of the candidate target center point set specifically includes: Read the operator mapping feature field, perform channel normalization and spatial continuity shaping, perform parallel feature aggregation in different spatial neighborhoods, form operator feature representations corresponding to different target scales and spatial context ranges, and generate a multi-scale operator feature set; The multi-scale operator feature set is input into the three-response field joint prediction structure to perform scale-wise fusion, and outputs the target probability response field, centrality response field and boundary response field that are consistent with the pixel grid of the distorted domain SAR observation field respectively. The joint prediction structure of the three response fields includes a shared feature aggregation backbone, a target probability response field prediction branch, a centrality response field prediction branch, and a boundary response field prediction branch. The scale-wise fusion process involves sequentially inputting multi-scale operator feature sets into a shared feature aggregation backbone in order of scale from coarse to fine. At each scale, channel alignment and spatial alignment are first performed on the operator features of that scale to obtain scale-aligned features. These features are then pixel-wise concatenated with the aggregation state features of the previous scale and updated by convolutional aggregation to obtain the aggregation state features of the current scale. The convolutional aggregation update at each scale includes both neighborhood feature aggregation and residual update to maintain information transfer between scales. Scale-gated weighting is applied to feature contributions from different scales to suppress inconsistent scale noise. After iterative aggregation at all scales, the final aggregation state features are output as a shared representation of the joint prediction structure of the three response fields. These features are then fed into the target probability response field prediction branch, the centrality response field prediction branch, and the boundary response field prediction branch to generate three types of response fields consistent with the pixel grid of the distorted domain SAR observation field. Among them, the target probability response field prediction branch generates a target probability response field reflecting the confidence of the target region by performing target existence discrimination mapping on the feature mapping results; the centrality response field prediction branch generates a centrality response field reflecting the salience of the target center position by performing centrality focusing mapping on the feature mapping results; and the boundary response field prediction branch generates a boundary response field reflecting the continuity of the target contour position by performing edge sensitive mapping on the feature mapping results. Apply sea area consistency constraints to the target probability response field to obtain the sea area-constrained target probability response field; The sea area consistency constraint processing suppresses the target probability response field values ​​of non-sea area pixel locations based on the sea-land mask, and performs boundary smoothing on the target probability response field values ​​of the sea-land boundary neighborhood. The centrality response field is subjected to centrality enhancement processing to obtain a centrality-enhanced centrality response field; The centrality enhancement process includes performing local extremum enhancement and neighborhood competition suppression on the centrality response field, and performing joint unification on the centrality response field in the high response region of the target probability response field. A set of candidate target center points is generated based on the centrality-enhanced centrality response field and the sea area-constrained target probabilistic response field. The candidate target center point set is obtained by performing a local extremum search on the centrality enhancement centrality response field, the center candidate set is then subjected to minimum spacing constraints and duplicate points are merged to obtain a deduplicated center candidate set, and the deduplicated center candidate set is then subjected to joint screening based on the sea area constraint target probability response field to obtain the candidate target center point set. The boundary response field is subjected to boundary consistency processing including thinning and fracture repair to form a continuous boundary clue set, which is then spatially aligned and bound with the probability response field of the marine constrained target and the set of candidate target center points.

[0023] In this embodiment, the generation of the marine remote sensing target identification result specifically includes: Based on the set of candidate target center points, perform a center point-guided region growth operation on the probability response field of constrained targets in the sea area to obtain the initial set of target instances; The initial set of target instances is formed by using the center point of each candidate target as a growth seed and performing pixel-level expansion in the sea-constrained target probability response field along the direction of monotonically decreasing probability value. During the expansion process, only pixel positions with a value not lower than a preset growth threshold in the target probability response field are allowed to enter. Crossing the sea-land boundary is prohibited under the sea-land mask constraint. The output is the initial set of target instances that corresponds one-to-one with the center point of each candidate target. Perform region consistency correction on the initial region set of the target instance to obtain the candidate region set of the target instance; The region consistency correction includes applying internal target probability consistency constraints and region connectivity constraints to each instance region in the initial set of target instance regions. The internal target probability consistency constraints filter out regions by comparing the consistency between the target probability response field value distribution of pixels inside the instance region and the value corresponding to the center point of the region. The region connectivity constraints ensure the spatial continuity of the instance regions by removing isolated sub-regions and fragmented connected regions, thus obtaining a set of candidate regions for target instances. Based on the continuous set of boundary clues, instance boundary constraints are applied to the set of candidate regions of target instances to obtain a set of target instance regions with boundary constraints. The instance boundary constraint introduces highly consistent boundary pixels from a continuous boundary cues set as constraint cues at the outer edge and internal void positions of the target instance candidate region set. It performs shrinkage correction on the edges of regions that do not continuously match the continuous boundary cues set, and performs filling expansion on regions that are completely surrounded by the continuous boundary cues set but are not covered by region growth, thus forming a boundary constraint target instance region set with continuous contours and closed boundaries. Perform instance-level overlap resolution on the set of target instance regions with boundary constraints to obtain a set of deduplicated target instance regions; The instance-level overlap resolution calculates the overlap ratio of target instance regions that overlap in space and sorts them by combining the values ​​of the corresponding candidate target center points in the centrality enhancement centrality response field. When the overlap ratio exceeds the preset resolution threshold, the target instance region with the higher centrality value is retained and the other target instance region is pruned or removed, and the deduplicated target instance region set is output. Generate a set of distorted domain target instance masks based on the set of deduplicated target instance regions; The set of target instance masks in the distorted domain is obtained by mapping each instance region in the set of deduplicated target instance regions to a binary mask representation consistent with the pixel grid of the SAR observation field in the distorted domain. Pixels within the instance region are marked as valid instance pixels, and pixels outside the instance region are marked as background pixels. Based on the inverse coordinate mapping relationship, the set of target instance masks in the distorted domain is subjected to inverse coordinate mapping to obtain the set of target instance masks in the original coordinate domain; The coordinate inverse mapping determines the original coordinate position of each pixel position in the distorted domain target instance mask set according to the inverse coordinate mapping relationship, and resamples on the original coordinate domain pixel grid to generate the original coordinate domain target instance mask set. During the resampling process, consistency merging is performed on many-to-one mapping positions, and neighborhood interpolation is performed on hole mapping positions to ensure the continuity of the mask space. Marine remote sensing target recognition results are generated based on the original coordinate domain target instance mask set; The marine remote sensing target recognition results include the spatial location, instance outline, instance area, and instance center point coordinates of each target instance. The instance outline is extracted from the boundary pixels of the original coordinate domain target instance mask set, and the instance center point coordinates are obtained by transforming the corresponding candidate target center point through the inverse coordinate mapping relationship, forming a marine remote sensing target recognition result set consistent with the spatial coordinate system of the original marine remote sensing synthetic aperture radar image.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a marine remote sensing target identification scenario in a complex nearshore waterway with numerous islands and reefs. This sea area has long been characterized by frequent sea state changes, complex coastline morphology, dense shipping channels, and strong background clutter. In synthetic aperture radar (SAR) images, targets such as small vessels and work platforms often overlap with sea clutter and shoreline echoes, leading to issues such as missed target detection, false alarms concentrated near the shoreline, and fragmented or incomplete target outlines in existing methods. These problems make it difficult to meet the needs of refined marine monitoring and identification.

[0025] In this application scenario, firstly, marine remote sensing synthetic aperture radar (SAR) images covering the sea area are acquired, and simultaneously, imaging geometry, observation azimuth, sea state information, and prior data on the coastline and shipping channels are incorporated to construct a multi-source observation field consistent with the radar image space. Through geometric constraints formed by the land-sea mask, the coastline range field, and the shipping channel range field, the original observation field is coordinately distorted, effectively expanding the coastal and densely packed shipping channel areas within the distorted domain, thereby alleviating the spatial compression and overlap problems between the target and the background. Subsequently, the distorted domain observation field is input into a sea state-conditional bispectral routing neural operator network, decomposing the observation signal in the frequency domain. Sea clutter subspace modeling is used to separate the sea clutter component from the target residual component, enabling adaptive characterization of background characteristics under different sea states and observation geometric conditions.

[0026] In the spectrum processing stage, differentiated spectral kernels are applied to the sea clutter component and the target residual component, respectively. Frequency band segmentation routing relationships are generated by combining sea state, incident angle, and observation azimuth, enabling the orderly fusion of spectral features across different frequency bands according to sea state conditions. This fusion is ultimately mapped back to the spatial domain to form a stable operator-mapped feature field. Based on this, target probability response, centrality response, and boundary response are generated through joint prediction structure, allowing the target location, center distribution, and contour cues to be co-expressed in the same spatial coordinate system. The centrality response guides the growth process of target instances, and the boundary response constrains and corrects the instance contours, effectively suppressing the ineffective expansion of background interference regions and preventing mutual erosion between target instances.

[0027] In practical applications, multiple radar images acquired at different times in this sea area were continuously processed. Results show that this invention can stably extract target instances while maintaining the continuity and spatial consistency of target outlines, even under conditions of significant sea state variations and substantial shoreline and channel echo interference. Compared to traditional methods that directly detect targets on raw images, this invention significantly reduces false detections near the shoreline in complex sea clutter backgrounds, while improving the ability to distinguish between weak and dense targets. This makes the identification results more reliable in terms of spatial positioning and instance integrity, providing a more robust foundation for subsequent maritime situation analysis and target management.

[0028] Table 1 Comparison of Marine Remote Sensing Target Recognition Performance in Complex Nearshore Scenarios

[0029] In terms of target detection rate, traditional methods are significantly affected by sea clutter and strong scattering from the shoreline in complex nearshore environments, with a detection rate of only 82.6%. Introducing depth features improves the detection rate of end-to-end methods to 88.9%, but still suffers from insufficient response to small targets and targets close to the shore. The method of this invention achieves a target detection rate of 94.3% under the same data conditions, demonstrating stable target response even in high-clutter backgrounds. This improvement directly stems from spectral domain projection decomposition effectively distinguishing sea clutter components from target residual components, preventing target energy from being masked by the background spectral structure.

[0030] Regarding the false alarm rate, traditional methods achieve a false alarm rate of 14.8%, with false alarms highly concentrated in the shoreline and channel neighborhoods. While end-to-end methods show a slight decrease, the rate remains at 10.6%. The method of this invention controls the false alarm rate to 6.9% and simultaneously reduces the proportion of false detections in the shoreline neighborhood to 15.8%. This result is directly related to the coordinate distortion mechanism and sea state-conditional frequency band routing. The shoreline and channel areas are expanded in the distortion domain, and the superimposed frequency band segmentation suppression mechanism makes it less likely for background responses to be misidentified as targets.

[0031] Regarding instance contour integrity, traditional methods rely on thresholding and morphological operations, resulting in a contour integrity of only 0.71. End-to-end methods improve this to 0.78, but discontinuities still exist at the break points of the target boundary. The method of this invention achieves a contour integrity of 0.87, indicating that through the boundary response field and instance boundary constraint mechanism, the region morphology can be continuously corrected during the instance generation stage, ensuring contour closure and boundary continuity.

[0032] The target center localization error further reflects the difference in localization accuracy. The traditional method has an average error of 5.4 pixels, the end-to-end method has an error of 4.1 pixels, while the method of this invention reduces the error to 2.6 pixels. This improvement comes from the explicit modeling of the centrality response field on the saliency of the target center, so that the generation of the center point no longer depends on the region mean or posterior regression, but is directly guided by the spatial centrality constraint.

[0033] Regarding the success rate of multi-target separation, traditional methods are prone to sticking together when targets are densely packed or adjacent, resulting in a success rate of only 68.5%. End-to-end methods offer some improvement, but instance overlap still exists in high sea state environments. The method of this invention achieves a multi-target separation success rate of 88.4%, thanks to the center-point guided region growth and subsequent instance-level overlap resolution mechanism, which gives instance generation clear growth boundaries and priority constraints.

[0034] Based on the above data, it can be seen that the present invention achieves stable improvements in multiple key indicators such as detection capability, false detection suppression, positioning accuracy, and instance completeness in complex nearshore environments. Its performance advantage comes from the synergistic effect of joint modeling of the spectral and spatial domains, sea state conditional operator routing, and instance-level constraint generation strategy, rather than from single structure or parameter tuning. It has high engineering reproducibility and promotion value.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for marine remote sensing target identification, characterized in that, Includes the following steps: Acquire marine remote sensing synthetic aperture radar images and generate incident angle field, observation azimuth coding field, sea state level field, land-sea mask, shoreline distance field and channel distance field that are spatially aligned with them to form a SAR observation field; Based on the land-sea mask, shoreline distance field and channel distance field, a coordinate distortion field constrained by shoreline and channel is constructed. The coordinate reparameterization of the SAR observation field is performed to obtain the distorted domain SAR observation field and the corresponding inverse coordinate mapping relationship. The distorted domain SAR observation field is input into the sea state conditional bispectral routing neural operator network, and the frequency domain transformation is performed to obtain the observation spectrum. A sea clutter subspace basis is constructed, and the observation spectrum is decomposed by spectral domain projection to obtain the sea clutter component spectrum and the target residual component spectrum. Background modeling spectral kernels are applied to the spectrum of sea clutter components, and target enhancement spectral kernels are applied to the spectrum of target residual components. Frequency band segmented routing coefficients are generated based on the sea state level field, the incident angle field, and the observation azimuth coding field. Sea state conditional routing fusion is performed on the bispectral convolution results to obtain the operator mapping feature field. Based on the operator mapping feature field, a target probability response field, a centrality response field, and a boundary response field are generated, and a set of candidate target center points is formed using the centrality response field. Using the set of candidate target center points as seeds, a set of target instance regions is generated on the target probability response field. The boundary response field is used to refine the contours of the target instance region set. The results of marine remote sensing target recognition are obtained by combining the inverse coordinate mapping relationship.

2. The deep learning-based marine remote sensing target recognition method according to claim 1, characterized in that, The generation of the SAR observation field specifically includes: The system acquires marine remote sensing synthetic aperture radar (SAR) images and, based on imaging geometry, observation azimuth information, sea state data corresponding to the observation time, and prior data on coastlines and waterways, generates an incident angle field, observation azimuth coding field, sea state level field, land-sea mask, coastline distance field, and waterway distance field that are spatially aligned with the marine remote sensing SAR images. These fields are then stitched together with the marine remote sensing SAR images by channel to form a SAR observation field.

3. The deep learning-based marine remote sensing target identification method according to claim 1, characterized in that, The generation of the distorted domain SAR observation field and the corresponding inverse coordinate mapping relationship specifically includes: Based on the pixel grid of the SAR observation field, the land and sea mask, shoreline distance field and channel distance field are read and uniformly cropped to the same coverage and resolution as the SAR observation field to generate a geometrically constrained input set; The distortion intensity field is generated based on the shoreline distance field and the channel distance field. Under the constraint of the land-sea mask, the distortion intensity of non-sea area pixels is suppressed to obtain the effective distortion intensity field of the sea area. Calculate the torsion direction field on the effective torsion intensity field of the sea area, and combine them to generate the coordinate displacement field; Spatial smoothing and boundary constraint processing are performed on the coordinate displacement field to generate a constrained coordinate displacement field that satisfies continuity and local monotonicity, and a coordinate distortion field is constructed. Based on the coordinate distortion field, coordinate reparameterization is performed on the SAR observation field to obtain the distorted domain SAR observation field; Perform a reversibility check on the coordinate distortion field and generate an inverse coordinate mapping relationship.

4. The deep learning-based marine remote sensing target identification method according to claim 1, characterized in that, The generation of the sea clutter component spectrum and the target residual component spectrum specifically includes: The distorted domain SAR observation field is input into the frequency domain mapping and spectral domain representation structure of the sea state conditional bispectral routing neural operator network, and frequency domain transformation is performed to obtain the set of observation spectra. The sea state conditional bispectral routing neural operator network includes a frequency domain mapping and spectral domain representation structure, a sea clutter subspace basis generation structure, a spectral domain projection decomposition structure, a bispectral convolution processing structure, a sea state conditional frequency band segmentation routing structure, a spectral domain to spatial domain back mapping structure, and a three-response field joint prediction structure. Based on the observation spectrum set, the observation spectrum statistics are generated and spatially and frequency band aligned with the sea state level field, the incident angle field, and the observation azimuth coding field to form the input set for generating the sea clutter sub-space basis. The input set for generating the sea clutter space basis is input into the sea clutter space basis generating structure to generate the sea clutter space basis; The observed spectrum is input into the spectral domain projection decomposition structure, and the spectral domain projection is performed using the sea clutter subspace basis to obtain the sea clutter component spectrum. The target residual component spectrum is obtained by the spectral domain difference between the observed spectrum and the sea clutter component spectrum. A consistency check is performed on the spectrum of the sea clutter component and the spectrum of the target residual component to form a set of availability tags.

5. The deep learning-based marine remote sensing target identification method according to claim 1, characterized in that, The generation of the operator mapping feature field specifically includes: Read the spectrum of sea clutter component and the spectrum of target residual component, and perform masking on the frequency band positions that do not meet the consistency check according to the availability tag set. Unify the frequency band division rules and spectral domain grid representation to be consistent with the observed spectrum, and generate the input set for spectral kernel action. Background modeling spectral kernels are applied to the sea clutter component spectrum to obtain the background modeling spectral response, and target enhancement spectral kernels are applied to the target residual component spectrum to obtain the target enhancement spectral response. The background modeling spectral response and the target enhancement spectral response are input into the bispectral convolution processing structure to obtain the bispectral convolution result; The sea state level field, the incident angle field, and the observation azimuth coding field are input into the sea state conditional frequency band segmented routing structure to generate frequency band segmented routing coefficients. The frequency band segmented routing coefficients are then aligned with the bispectral convolution result to obtain a routable bispectral convolution result. Perform sea state-conditional route fusion on the routable bispectral convolution result to obtain the route fusion spectrum response; Perform cross-band consistency processing on the route fusion spectrum response to obtain a consistent spectrum response; The uniformized spectral response is input to the spectral domain and then back-mapped to the spatial domain. This back-mapping is performed to obtain the operator-mapped feature field.

6. The deep learning-based marine remote sensing target identification method according to claim 1, characterized in that, The generation of the candidate target center point set specifically includes: Read the operator mapping feature field, perform channel normalization and spatial continuity shaping, perform parallel feature aggregation in different spatial neighborhoods, form operator feature representations corresponding to different target scales and spatial context ranges, and generate a multi-scale operator feature set; The multi-scale operator feature set is input into the three-response field joint prediction structure to perform scale-wise fusion, and outputs the target probability response field, centrality response field and boundary response field that are consistent with the pixel grid of the distorted domain SAR observation field respectively. Apply sea area consistency constraints to the target probability response field to obtain the sea area-constrained target probability response field; The centrality response field is subjected to centrality enhancement processing to obtain a centrality-enhanced centrality response field; A set of candidate target center points is generated based on the centrality-enhanced centrality response field and the sea area-constrained target probabilistic response field. The boundary response field is subjected to boundary consistency processing including thinning and fracture repair to form a continuous boundary clue set, which is then spatially aligned and bound with the probability response field of the marine constrained target and the set of candidate target center points.

7. The deep learning-based marine remote sensing target identification method according to claim 1, characterized in that, The generation of the marine remote sensing target identification results specifically includes: Based on the set of candidate target center points, perform a center point-guided region growth operation on the probability response field of constrained targets in the sea area to obtain the initial set of target instances; Perform region consistency correction on the initial region set of the target instance to obtain the candidate region set of the target instance; Based on the continuous set of boundary clues, instance boundary constraints are applied to the set of candidate regions of target instances to obtain a set of target instance regions with boundary constraints. Perform instance-level overlap resolution on the set of target instance regions with boundary constraints to obtain a set of deduplicated target instance regions; Generate a set of distorted domain target instance masks based on the set of deduplicated target instance regions; Based on the inverse coordinate mapping relationship, the set of target instance masks in the distorted domain is subjected to inverse coordinate mapping to obtain the set of target instance masks in the original coordinate domain; Marine remote sensing target recognition results are generated based on the original coordinate domain target instance mask set.

Images