A remote sensing target detection sample quality optimization method based on feature alignment

Abstract

The application provides a remote sensing target detection sample quality optimization method based on feature alignment, and relates to the technical field of data processing, and comprises the following steps: feature coding and feature extraction are performed on to-be-processed labeled data to obtain target intrinsic features and position embedding features; local feature analysis is performed on the target intrinsic features to determine feature extreme value regions; the target intrinsic features are guided to align with the feature extreme value regions of the same category by using category prototype features to obtain optimized intrinsic features; feature dimension remodeling and modeling are performed on the target intrinsic features to obtain corresponding state space features; whether the labeled frame is consistent with the target real features is judged based on target fusion features corresponding to the state space features, the optimized intrinsic features and the position embedding features; if yes, a classification prototype cache pool of a corresponding category is updated based on the target fusion features; and if not, the labeled frame is position-corrected and iteratively optimized to obtain target labeled data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for optimizing the sample quality of remote sensing target detection based on feature alignment. Background Technology

[0002] Remote sensing image target detection relies on high-quality labeled data. For targets such as aircraft and vehicles, labels typically represent the target's location, scale, angle, and category using oriented bounding boxes or rotated boxes. Labeling position errors can occur during manual, semi-automatic, and automatic labeling processes, as well as multi-source image registration. This error is a form of geometric noise in the training labels, manifesting as correct category labels but a misalignment of the labeled bounding box relative to the true target location. This disrupts the localization supervision signal, causing the model's regression branch to receive supervision information deviating from the true target range, leading to problems such as localization errors, inaccurate boundaries, or missed target detection. In scenarios involving small and dense targets, even slight misalignments can cause a significant decrease in the intersection-over-union ratio (IoU). Essentially, labeling position errors represent a spatial misalignment between the labeled box and the true target range. Identifying and correcting labeling position errors in remote sensing image data is a pressing issue that needs to be addressed. Summary of the Invention

[0003] In view of this, the present invention proposes a method for optimizing the sample quality of remote sensing target detection based on feature alignment.

[0004] The technical solution of this invention is implemented as follows: The first aspect of this invention provides a method for optimizing the sample quality of remote sensing target detection based on feature alignment, comprising: Feature encoding and feature extraction are performed on the labeled data to be processed to obtain the target intrinsic features and location embedding features; the target intrinsic features include the overall appearance, structure and category information of the target in the feature space, and the location embedding features include the center position, scale and orientation information of the label box; Local feature analysis is performed on the target intrinsic features to obtain the corresponding local feature distribution domain, and the positions where the feature values ​​in the local feature distribution domain reach local maxima or local minima and their neighborhoods are determined as feature extremum regions; the target intrinsic features are guided to align with the feature extremum regions of the same category using category prototype features to obtain optimized intrinsic features; the category prototype features are category-level reference templates formed by summarizing the features of labeled samples that are determined to be spatially consistent. The intrinsic features of the target are reshaped and modeled to obtain the corresponding state space features. Based on the state space features, the optimized intrinsic features, and the target fusion features corresponding to the position embedding features, it is determined whether the bounding box is consistent with the true features of the target. If so, the classification prototype cache pool of the corresponding category is updated based on the target fusion features. If not, the bounding box is corrected in position and iteratively optimized to obtain the target annotation data. The state space features include category semantic information and spatial context information. The classification prototype cache pool includes multiple instances of category prototype features.

[0005] Based on the above technical solutions, preferably, the step of performing local feature analysis on the intrinsic features of the target to obtain the corresponding local feature distribution domain includes: The irrelevant background region in the complete feature map containing the intrinsic features of the target is masked to obtain the key feature region covered by the bounding box. Local feature analysis is performed on the intrinsic features of the target within the key feature region to obtain the corresponding local feature distribution domain and the feature changes at different spatial locations within the local feature distribution domain.

[0006] Based on the above technical solutions, preferably, before using category prototype features to guide the alignment of the target intrinsic features towards the extreme value region of the same category of features to obtain optimized intrinsic features, the following steps are included: Obtain at least three sets of clean labeled data without positional errors for each target category; The feature representation obtained by feature encoding of the clean labeled data is written as the category prototype feature into the classification prototype cache pool of the corresponding category.

[0007] Based on the above technical solutions, preferably, the step of using category prototype features to guide the target intrinsic features to align with the feature extreme value region of the same category to obtain optimized intrinsic features includes: The adaptive alignment strength is determined based on the difference between the category prototype features and the target intrinsic features; The target intrinsic features are guided to align with the extreme value regions of the same category of features by using the alignment step size determined by the adaptive alignment strength, thereby obtaining optimized intrinsic features.

[0008] Based on the above technical solutions, preferably, the step of reshaping and modeling the feature dimensions of the target intrinsic features to obtain the corresponding state space features includes: A linear transformation is performed on the intrinsic features of the target to complete the dimension mapping, thereby obtaining the first intermediate feature; The local sequence associations in the first intermediate feature are captured by one-dimensional convolution, and non-linear activation is performed to enhance them, thus obtaining the second intermediate feature; The second intermediate feature is processed using a structured state-space model to model long-range dependencies and global semantics, resulting in the corresponding state-space feature.

[0009] Based on the above technical solutions, preferably, the true features of the target include the target subject, target boundary, and category prototype features; the step of determining whether the bounding box is consistent with the true features of the target based on the target fusion features corresponding to the state space features, the optimized intrinsic features, and the location embedding features includes: The state space features, the optimized intrinsic features, and the location embedding features are fused to obtain the target fusion features; Based on the target fusion features, the target subject, target boundary, and category prototype features corresponding to the annotation box are judged. Only when the target subject, target boundary, and category prototype features are all consistent, it is determined that the annotation box is consistent with the target's true features.

[0010] Based on the above technical solutions, preferably, the step of correcting and iteratively optimizing the position of the annotation box to obtain the target annotation data includes: After normalizing the target fusion features corresponding to the bounding boxes, the data is input into the C3K2 module for feature refinement, and the corrected bounding boxes are output by the Detect head. The corrected bounding boxes are used for feature extraction, feature alignment, and position correction until the preset number of iterations is reached to obtain the target labeled data.

[0011] Furthermore, a second aspect of the present invention provides a remote sensing target detection sample quality optimization system based on feature alignment, comprising: a feature extraction module, a feature alignment module, and a discrimination correction module; wherein, The feature extraction module is configured to perform feature encoding and feature extraction on the labeled data to be processed, and to obtain the target intrinsic features and the location embedding features; the target intrinsic features include the overall appearance, structure and category information of the target in the feature space, and the location embedding features include the center position, scale and orientation information of the label box; The feature alignment module is configured to perform local feature analysis on the target intrinsic features to obtain the corresponding local feature distribution domain, and to determine the positions where the feature values ​​in the local feature distribution domain reach local maxima or local minima and their neighborhoods as feature extremum regions; and to use category prototype features to guide the target intrinsic features to align with the feature extremum regions of the same category to obtain optimized intrinsic features; the category prototype features are category-level reference templates formed by summarizing the features of labeled samples that are determined to be spatially consistent. The discrimination and correction module is configured to reshape and model the feature dimensions of the target intrinsic features to obtain the corresponding state space features; based on the state space features, the optimized intrinsic features, and the target fusion features corresponding to the position embedding features, it determines whether the bounding box is consistent with the target true features. If so, it updates the classification prototype cache pool of the corresponding category based on the target fusion features; if not, it performs position correction and iterative optimization on the bounding box. The state space features include category semantic information and spatial context information; the classification prototype cache pool includes multiple instances of category prototype features.

[0012] More preferably, a third aspect of the present invention provides an electronic device, including a processor and a memory; the memory has a computer program stored thereon, wherein the computer program, when executed by the processor, implements the feature alignment-based remote sensing target detection sample quality optimization method described in the first aspect.

[0013] More preferably, a fourth aspect of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the feature alignment-based remote sensing target detection sample quality optimization method described in the first aspect.

[0014] The remote sensing target detection sample quality optimization method based on feature alignment of the present invention has the following advantages over the prior art: 1. By extracting extreme feature regions through local feature analysis, the matching degree between the coverage area of ​​the bounding box and the core area of ​​the real target is directly compared to quantify the degree of spatial misalignment and accurately identify annotation errors. Based on this, category prototype features formed by aggregating similar categories are used as a general template to guide the offset intrinsic features of the target to align with stable extreme feature regions of the same category. This allows features distorted by annotation misalignment to revert to the standard feature distribution of that category of target, fundamentally resolving feature distortion caused by spatial misalignment of the bounding box. Combining local details, global context, and location information for joint judgment avoids misjudgment based on a single dimension, ensuring the accuracy of error identification and correction results.

[0015] 2. By introducing a classification prototype cache pool, prototype-guided feature extremum region alignment, and state-space feature branching, automatic correction of annotation position errors was achieved. Throughout multiple iterations, the stability of the category prototype, the effectiveness of feature alignment, and the stability of later corrections were maintained, improving the geometric quality of bounding box annotations in the remote sensing target detection dataset. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0017] Figure 1 A flowchart illustrating a method for optimizing the sample quality of remote sensing target detection based on feature alignment, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the changes in overall IoU and pure IoU with the number of calibration rounds provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the structure of a remote sensing target detection sample quality optimization system based on feature alignment, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In some embodiments, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for optimizing the quality of remote sensing target detection samples based on feature alignment, provided by an embodiment of the present invention. The method includes: S110, perform feature encoding and feature extraction on the labeled data to be processed to obtain the intrinsic features of the target and the location embedding features; the intrinsic features of the target include the overall appearance, structure and category information of the target in the feature space, and the location embedding features include the center position, scale and orientation information of the label box.

[0020] S120, perform local feature analysis on the intrinsic features of the target to obtain the corresponding local feature distribution domain, and determine the position where the feature value in the local feature distribution domain reaches a local maximum or local minimum and its neighborhood as the feature extreme value region; use the category prototype feature to guide the intrinsic features of the target to align with the feature extreme value region of the same category to obtain the optimized intrinsic features; the category prototype feature is a category-level reference template formed by summarizing the features of the labeled samples that are judged to be spatially consistent.

[0021] S130: Reshape and model the intrinsic features of the target to obtain the corresponding state space features; Based on the target fusion features corresponding to the state space features, optimized intrinsic features, and position embedding features, determine whether the bounding box is consistent with the target's true features. If so, update the classification prototype cache pool of the corresponding category based on the target fusion features. If not, perform position correction and iterative optimization on the bounding box to obtain the target annotation data. The state space features include category semantic information and spatial context information. The classification prototype cache pool includes multiple category prototype features.

[0022] In this embodiment, YOLOv11-obb is used as a feature encoder to extract features from the labeled data to be processed, obtaining the intrinsic features of the target instance to be judged. Simultaneously, the center position, scale, and orientation information of the bounding box are encoded as position embedding features, providing spatial constraints for subsequent position error discrimination and target box correction. The labeled data to be processed here includes remote sensing images and their target box annotations.

[0023] Local feature analysis of the target's intrinsic features reveals the extreme feature regions at the target's true subject, boundaries, and foreground-background boundaries. These extreme regions represent the core effective areas of the true target and can serve as objective criteria for determining whether the annotation has shifted. Using the category prototype features condensed from multiple instances of the same category as a standard template, the intrinsic features of the currently shifted target are constrained to converge towards the stable extreme feature regions of the same category of targets. This eliminates feature distortion caused by misalignment in the annotation space, resulting in optimized intrinsic features that better reflect the true target.

[0024] The original intrinsic features of the target are transformed in dimension and structured modeled to mine the global semantics and spatial context associations to obtain state space features. Then, the state space features, intrinsic features and position embedding features are fused and the spatial range of the annotation box is comprehensively judged to determine whether it matches the real target and complete the annotation error identification. If the judgment is consistent, the category prototype cache pool is updated with high-quality features to enrich the template library. If the judgment is inconsistent, there is a spatial misalignment. The coordinates, scale and angle of the annotation box are corrected and iteratively optimized until the requirements are met, and the final qualified annotation data is output.

[0025] In some embodiments, local feature analysis is performed on the intrinsic features of the target to obtain the corresponding local feature distribution domain, including: Masking is performed on irrelevant background regions in the complete feature map containing the intrinsic features of the target to obtain the key feature regions covered by the bounding boxes; Local feature analysis is performed on the intrinsic features of the target within the key feature region to obtain the corresponding local feature distribution domain and the feature changes at different spatial locations within the local feature distribution domain.

[0026] In this embodiment, the current target instance is separated from the complete feature map, and irrelevant background regions are masked, so that the alignment process mainly focuses on the geometric inconsistencies within the target region. For example, a binary mask matrix is ​​generated based on information such as the center, scale, orientation, and rotation angle of the bounding box. The binary mask matrix is ​​then multiplied position-by-position with the complete feature map. The original feature values ​​are retained for regions inside the bounding box, while the features of irrelevant background regions outside the box are set to zero. Finally, only the region covered by the bounding box is retained, yielding the key feature region. By focusing local feature analysis only on the true target, misclassification of background regions as target feature regions can be effectively avoided, and the computational load is significantly reduced.

[0027] In some embodiments, before guiding the target intrinsic features to align with the extreme value regions of features of the same category using category prototype features to obtain optimized intrinsic features, the process includes: Obtain at least three sets of clean labeled data without positional errors for each target category; The feature representation obtained by feature encoding of clean labeled data is written as the category prototype feature into the classification prototype cache pool of the corresponding category.

[0028] In this embodiment, at least three clean annotations, verified manually and free of positional errors, are selected for each target category. These clean annotations are processed by a feature encoder to obtain the corresponding intrinsic features of the target, which are then written into the corresponding category's classification prototype cache pool as category prototype features. In subsequent iterations, annotation samples determined to be spatially consistent can continue to be used to update the cache pool; samples determined to have positional errors do not directly participate in category prototype updates to avoid contaminating the category-level reference.

[0029] In some embodiments, the target intrinsic features are guided to align with the extreme value regions of features of the same category using category prototype features to obtain optimized intrinsic features, including: The adaptive alignment strength is determined based on the difference between the category prototype features and the target intrinsic features; By using the alignment step size determined by the adaptive alignment strength, the intrinsic features of the target are guided to align with the extreme value regions of the same category of features, thus obtaining optimized intrinsic features.

[0030] In this embodiment, the target intrinsic feature X and the category prototype feature are used. Calculate the adaptive alignment strength based on the difference between them Then, guided by the category prototype, the intrinsic features of the target are aligned, so that the offset intrinsic features of the target gradually move closer to the more stable feature extremum region of the same category target. After feature alignment, the aligned optimized intrinsic features are obtained, and the calculation process is as follows:

[0031] ; Where X is the original target feature representation of the current target instance. These are optimized intrinsic features after alignment of the feature extremum regions guided by the prototype. For the corresponding category prototype features in the current iteration, For adaptive alignment strength, s is a scaling factor. For numerical stability, The number of spatial locations in the feature map. and Let these represent the category prototype feature and the target intrinsic feature at the i-th spatial location, respectively. This represents the local mean of the two values.

[0032] In some embodiments, the intrinsic features of the target are reshaped and modeled to obtain the corresponding state space features, including: A linear transformation is performed on the intrinsic features of the target to complete the dimension mapping, resulting in the first intermediate feature; One-dimensional convolution is used to capture local sequence associations in the first intermediate feature, and non-linear activation is applied to enhance them to obtain the second intermediate feature; The second intermediate feature is processed using a structured state-space model to model long-range dependencies and global semantics, thereby obtaining the corresponding state-space features.

[0033] In this embodiment, the intrinsic features of the current target are taken as input and processed sequentially through linear transformation, one-dimensional convolution, nonlinear activation and structured state space model to obtain state space features that take into account both local details and global correlations. This preserves the category semantic information and spatial context information of the current target instance, providing a basis for subsequent discrimination and correction.

[0034] In some embodiments, the true features of the target include the target body, target boundary, and category prototype features; determining whether the bounding box is consistent with the true features of the target based on the target fusion features corresponding to state space features, optimized intrinsic features, and location embedding features includes: The target fusion feature is obtained by fusing state space features, optimized intrinsic features, and location embedding features; Based on the target fusion features, the target subject, target boundary and category prototype features corresponding to the bounding box are judged. Only when the target subject, target boundary and category prototype features are consistent are the bounding box determined to be consistent with the target's true features.

[0035] In this embodiment, state space features, optimized intrinsic features, and location embedding features are fused to determine whether the current bounding box is spatially consistent with the target's true location. If the current bounding box is consistent with the target body, target boundary, and category prototype features, it is determined to be a spatially consistent bounding box, and its features can be used to update the corresponding category's classification prototype cache pool. If the current bounding box has a significant spatial deviation from the target body or target boundary, and the fused features show that there is a feature mismatch between it and the category prototype, it is determined to be a bounding box with a location error. Bounding boxes determined to have location errors enter the subsequent target bounding box correction process, and their features are not directly written to the classification prototype cache pool to avoid polluting the category-level reference.

[0036] In some embodiments, the position of the annotation box is corrected and iteratively optimized to obtain target annotation data, including: After normalizing the target fusion features corresponding to the bounding boxes, the data is input into the C3K2 module for feature refinement, and the corrected bounding boxes are output by the Detect head. The corrected bounding boxes are used for feature extraction, feature alignment, and position correction until the preset number of iterations is reached to obtain the target labeled data.

[0037] In this embodiment, normalization maps all feature values ​​to a unified numerical range, eliminating the influence of dimensions, smoothing feature distribution, and preventing features with excessively large or small values ​​from dominating subsequent calculations. This also improves the stability of network training and inference, allowing the C3K2 module to learn effective information across all dimensions in a balanced way. The C3K2 module is a lightweight feature enhancement module based on an improved C3 structure. It integrates multi-branch convolution, gradient splitting, and dense feature fusion design, enabling deep semantic purification, detail enhancement, and feature denoising of the normalized target fused features. On one hand, it strengthens the response of key features such as the target subject and boundaries, weakening residual noise; on the other hand, it further associates features with the position and pose information of the bounding boxes, mining the mapping relationship between features and spatial positions, providing better feature basis for coordinate correction. The Detect head is a dedicated output head for detection tasks. Based on the refined features from the front end, it completes target coordinate regression and pose correction. Combined with the remote sensing rotating bounding box scenario, based on the current features, it learns the offset between the bounding box and the real target, and reversely corrects the center coordinates, scale, and rotation angle of the bounding box, transforming the deviation at the feature level into specific geometric parameter adjustments, and then directly outputting new bounding box parameters.

[0038] In one example, see Figure 2 , Figure 2This diagram illustrates the changes in overall IoU and pure IoU with each correction round, as provided in this embodiment of the invention. Overall IoU reflects the average overlap between all target bounding boxes and the actual target area, and is influenced by factors such as annotation error, target distribution, and scene complexity. Pure IoU refers to the overlap between the bounding boxes of the target body region and the actual target after removing background interference, thus better reflecting the accuracy of the correction algorithm in aligning the target body. In a position error simulation scenario constructed using the UCAS_AOD dataset, the noise ratio was set to 50%, and the displacement amplitude was set to 50%. Experimental results show that the overall IoU at the noise starting point in round 0 was 0.54, and the pure IoU was 0.11. After five rounds of correction, the overall IoU increased to 0.80, and the pure IoU increased to 0.60. The significant improvement in pure IoU indicates that this scheme can effectively improve the geometric position of target boxes with offset errors. The results show that the corrected target boxes for aircraft and vehicles are closer to the reference boxes than the initial noise boxes, the spatial offset between the target boxes and the actual target areas is significantly reduced, and the target center coverage is more reasonable.

[0039] In some embodiments, please refer to Figure 3 , Figure 3 This is a schematic diagram of a remote sensing target detection sample quality optimization system based on feature alignment, provided in an embodiment of the present invention. The present invention provides a remote sensing target detection sample quality optimization system 300 based on feature alignment, comprising: a feature extraction module 310, a feature alignment module 320, and a discrimination correction module 330.

[0040] The feature extraction module 310 is configured to perform feature encoding and feature extraction on the labeled data to be processed, and to obtain the intrinsic features of the target and the location embedding features. The intrinsic features of the target include the overall appearance, structure and category information of the target in the feature space, and the location embedding features include the center position, scale and orientation information of the label box.

[0041] The feature alignment module 320 is configured to perform local feature analysis on the intrinsic features of the target to obtain the corresponding local feature distribution domain, and to determine the position where the feature value in the local feature distribution domain reaches a local maximum or local minimum and its neighborhood as the feature extreme value region; the target intrinsic features are guided to align with the feature extreme value region of the same category using the category prototype features to obtain optimized intrinsic features; the category prototype features are category-level reference templates formed by summarizing the features of labeled samples that are determined to be spatially consistent.

[0042] The discrimination and correction module 330 is configured to reshape and model the feature dimensions of the target intrinsic features to obtain the corresponding state space features; based on the target fusion features corresponding to the state space features, optimized intrinsic features and position embedding features, it judges whether the bounding box is consistent with the target true features. If so, it updates the classification prototype cache pool of the corresponding category based on the target fusion features. If not, it performs position correction and iterative optimization on the bounding box. The state space features include category semantic information and spatial context information. The classification prototype cache pool includes multiple class prototype features.

[0043] In some embodiments, the feature alignment module 320 is specifically configured as follows: Masking is performed on irrelevant background regions in the complete feature map containing the intrinsic features of the target to obtain the key feature regions covered by the bounding boxes; Local feature analysis is performed on the intrinsic features of the target within the key feature region to obtain the corresponding local feature distribution domain and the feature changes at different spatial locations within the local feature distribution domain.

[0044] In some embodiments, the remote sensing target detection sample quality optimization system based on feature alignment further includes a cache pool creation module; the cache pool creation module is specifically configured as follows: Obtain at least three sets of clean labeled data without positional errors for each target category; The feature representation obtained by feature encoding of clean labeled data is written as the category prototype feature into the classification prototype cache pool of the corresponding category.

[0045] In some embodiments, the feature alignment module 320 is specifically configured as follows: The adaptive alignment strength is determined based on the difference between the category prototype features and the target intrinsic features; By using the alignment step size determined by the adaptive alignment strength, the intrinsic features of the target are guided to align with the extreme value regions of the same category of features, thus obtaining optimized intrinsic features.

[0046] In some embodiments, the discrimination correction module 330 is specifically configured as follows: A linear transformation is performed on the intrinsic features of the target to complete the dimension mapping, resulting in the first intermediate feature; One-dimensional convolution is used to capture local sequence associations in the first intermediate feature, and non-linear activation is applied to enhance them to obtain the second intermediate feature; The second intermediate feature is processed using a structured state-space model to model long-range dependencies and global semantics, thereby obtaining the corresponding state-space features.

[0047] In some embodiments, the true features of the target include the target subject, the target boundary, and category prototype features; the discrimination correction module 330 is specifically configured as follows: The target fusion feature is obtained by fusing state space features, optimized intrinsic features, and location embedding features; Based on the target fusion features, the target subject, target boundary and category prototype features corresponding to the bounding box are judged. Only when the target subject, target boundary and category prototype features are consistent are the bounding box determined to be consistent with the target's true features.

[0048] In some embodiments, the discrimination correction module 330 is specifically configured as follows: After normalizing the target fusion features corresponding to the bounding boxes, the data is input into the C3K2 module for feature refinement, and the corrected bounding boxes are output by the Detect head. The corrected bounding boxes are used for feature extraction, feature alignment, and position correction until the preset number of iterations is reached to obtain the target labeled data.

[0049] It should be noted that the remote sensing target detection sample quality optimization system based on feature alignment provided in this application embodiment and the remote sensing target detection sample quality optimization method based on feature alignment provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned remote sensing target detection sample quality optimization method based on feature alignment, and the repeated parts will not be described again.

[0050] In some embodiments, please refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 400 provided in this embodiment includes a processor 410 and a memory 420; the memory 420 stores a computer program, wherein the computer program, when executed by the processor, implements the aforementioned method for optimizing the sample quality of remote sensing target detection based on feature alignment.

[0051] Specifically, processor 410 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. Processor 410 may also include onboard memory for caching purposes. Processor 410 may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0052] Memory 420 may be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory 420 may include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory 420 include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and may also be random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0053] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this program implements the aforementioned method for optimizing the sample quality of remote sensing target detection based on feature alignment. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0054] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0055] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by the equivalents of the appended claims. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A method for optimizing the sample quality of remote sensing target detection based on feature alignment, characterized in that, include: Perform feature encoding and feature extraction on the labeled data to be processed to obtain the intrinsic features of the target and the location embedding features; The intrinsic features of the target include the overall appearance, structure and category information of the target in the feature space, and the location embedding features include the center position, scale and orientation information of the bounding box; Local feature analysis is performed on the intrinsic features of the target to obtain the corresponding local feature distribution domain, and the positions where the feature values ​​in the local feature distribution domain reach local maxima or local minima and their neighborhoods are determined as feature extremum regions. The adaptive alignment strength is determined based on the difference between the category prototype feature and the target intrinsic feature; the alignment step determined by the adaptive alignment strength is used to guide the target intrinsic feature to align to the feature extreme value region of the same category, thereby obtaining the optimized intrinsic feature; the category prototype feature is a category-level reference template formed by summarizing the features of labeled samples that are determined to be spatially consistent. The intrinsic features of the target are reshaped and modeled to obtain the corresponding state space features; Based on the target fusion features corresponding to the state space features, the optimized intrinsic features, and the position embedding features, it is determined whether the bounding box is consistent with the target true features. If so, the classification prototype cache pool of the corresponding category is updated based on the target fusion features. If not, the bounding box is corrected in position and iteratively optimized to obtain the target annotation data. The state space features include category semantic information and spatial context information; the classification prototype cache pool includes multiple category prototype features.

2. The method for optimizing the sample quality of remote sensing target detection based on feature alignment as described in claim 1, characterized in that, The step of performing local feature analysis on the intrinsic features of the target to obtain the corresponding local feature distribution domain includes: The irrelevant background region in the complete feature map containing the intrinsic features of the target is masked to obtain the key feature region covered by the bounding box. Local feature analysis is performed on the intrinsic features of the target within the key feature region to obtain the corresponding local feature distribution domain and the feature changes at different spatial locations within the local feature distribution domain.

3. The method for optimizing the sample quality of remote sensing target detection based on feature alignment as described in claim 1, characterized in that, Before obtaining optimized intrinsic features, the process involves determining an adaptive alignment strength based on the difference between the category prototype features and the target intrinsic features, and using the alignment step size determined by the adaptive alignment strength to guide the target intrinsic features to align towards the feature extremum region of the same category. Obtain at least three sets of clean labeled data without positional errors for each target category; The feature representation obtained by feature encoding of the clean labeled data is written as the category prototype feature into the classification prototype cache pool of the corresponding category.

4. The method for optimizing the sample quality of remote sensing target detection based on feature alignment as described in claim 1, characterized in that, The step of reshaping and modeling the intrinsic features of the target to obtain the corresponding state space features includes: A linear transformation is performed on the intrinsic features of the target to complete the dimension mapping, thereby obtaining the first intermediate feature; The local sequence associations in the first intermediate feature are captured by one-dimensional convolution, and non-linear activation is performed to enhance them, thus obtaining the second intermediate feature; The second intermediate feature is processed using a structured state-space model to model long-range dependencies and global semantics, resulting in the corresponding state-space feature.

5. The method for optimizing the sample quality of remote sensing target detection based on feature alignment as described in claim 1, characterized in that, The true features of the target include the target subject, target boundary, and category prototype features; the step of determining whether the bounding box is consistent with the true features of the target based on the target fusion features corresponding to the state space features, the optimized intrinsic features, and the location embedding features includes: The state space features, the optimized intrinsic features, and the location embedding features are fused to obtain the target fusion features; Based on the target fusion features, the target subject, target boundary, and category prototype features corresponding to the annotation box are judged. Only when the target subject, the target boundary, and the category prototype features are all consistent, it is determined that the annotation box is consistent with the target's true features.

6. The method for optimizing the sample quality of remote sensing target detection based on feature alignment as described in claim 1, characterized in that, The step of correcting and iteratively optimizing the position of the annotation box to obtain the target annotation data includes: After normalizing the target fusion features corresponding to the labeled boxes, the data is input into the C3K2 module for feature refinement, and the detection head outputs the corrected labeled boxes. The corrected bounding boxes are used for feature extraction, feature alignment, and position correction until the preset number of iterations is reached to obtain the target labeled data.

7. A remote sensing target detection sample quality optimization system based on feature alignment, characterized in that, include: The module consists of a feature extraction module, a feature alignment module, and a discrimination correction module; among which, The feature extraction module is configured to perform feature encoding and feature extraction on the labeled data to be processed, and to obtain the target intrinsic features and the location embedding features; the target intrinsic features include the overall appearance, structure and category information of the target in the feature space, and the location embedding features include the center position, scale and orientation information of the label box; The feature alignment module is configured to perform local feature analysis on the target intrinsic features to obtain the corresponding local feature distribution domain, and to determine the positions where the feature values ​​in the local feature distribution domain reach local maxima or local minima and their neighborhoods as feature extremum regions; to determine the adaptive alignment strength based on the difference between the category prototype features and the target intrinsic features; and to guide the target intrinsic features to align to the feature extremum regions of the same category using the alignment step size determined by the adaptive alignment strength, thereby obtaining optimized intrinsic features; the category prototype features are category-level reference templates formed by summarizing the features of labeled samples that are determined to be spatially consistent. The discrimination and correction module is configured to reshape and model the feature dimensions of the target intrinsic features to obtain the corresponding state space features; based on the state space features, the optimized intrinsic features, and the target fusion features corresponding to the position embedding features, it determines whether the bounding box is consistent with the target's true features. If so, it updates the classification prototype cache pool of the corresponding category based on the target fusion features; if not, it performs position correction and iterative optimization on the bounding box to obtain target annotation data. The state space features include category semantic information and spatial context information; the classification prototype cache pool includes multiple instances of category prototype features.

8. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the remote sensing target detection sample quality optimization method based on feature alignment as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the feature alignment-based remote sensing target detection sample quality optimization method as described in any one of claims 1 to 6.

Images