Decoupled dual spatial directional diffusion method and system for sar image remote sensing target detection

Abstract

This invention provides a decoupled dual-space directional diffusion method and system for remote sensing target detection in SAR images. The method includes a model training process and a model inference process. The model training process includes: extracting multi-scale features from labeled SAR images; determining a candidate bounding box set using the multi-scale features; determining candidate labels based on the candidate bounding box set; obtaining the initial state of the label branch based on the candidate labels; splitting each bounding box in the candidate bounding box set into a center coordinate branch and a size branch; jointly obtaining the initial diffusion state by combining the initial states of the center coordinate branch, the size branch, and the label branch; obtaining a denoising prediction result based on the initial diffusion state; and calculating new model parameters for the loss function using the denoising prediction result. The method provided by this invention effectively improves the accuracy of target detection.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, remote sensing image processing and artificial intelligence target perception technology, and in particular to a decoupled dual-space directional diffusion method and system for remote sensing target detection in SAR images. Background Technology

[0002] SAR imagery possesses all-weather, all-day imaging capabilities, making it invaluable for applications such as maritime vessel monitoring, airport aircraft identification, coastal surveillance, disaster assessment, and national remote sensing. However, the imaging mechanism of SAR differs significantly from that of visible light images. SAR images commonly exhibit problems such as speckle noise, strong background clutter, unstable target shape and texture, significant differences in scattering patterns among similar targets, and similar appearances among dissimilar targets. These issues make traditional target detection methods prone to missed detections, false detections, and category confusion in complex scenarios.

[0003] Existing SAR target detection technologies generally include two-stage detection techniques based on candidate regions, one-stage detection techniques based on dense prediction, end-to-end ensemble prediction techniques based on queries, and diffusion-based detection techniques that have gradually emerged in recent years. Diffusion-based detection technology transforms bounding box prediction into a reverse denoising process of "gradually restoring a clean target box from a noisy state," which can naturally support iterative hypothesis correction and uncertainty modeling, and has become an important development direction for target detection.

[0004] However, existing diffusion-based detection methods still suffer from low detection accuracy when targeting SAR targets. Summary of the Invention

[0005] This invention provides a decoupled dual-space directional diffusion method and system for remote sensing target detection in SAR images, in order to solve the problem of low accuracy in existing diffusion-based detection methods.

[0006] A decoupled dual-space directional diffusion method for remote sensing target detection in SAR images includes: a model training process and a model inference process; The model training process includes: Step 1: Obtain labeled SAR image training samples; the training samples include: a set of SAR images, and a set of real target bounding boxes corresponding to each SAR image. Real category tag set ; Step 2: Perform multi-scale feature extraction on the labeled SAR image to obtain multi-scale features F; Step 3: Generate conditional priors for the multi-scale features F to obtain conditional prior boxes. ; Step 4: Based on the aforementioned conditional prior box Construct a candidate box set B; and based on the true category label set... Assign a category label to each candidate box to obtain the candidate labels. ; Step 5: Split each bounding box in the candidate box set B into a center coordinate branch and a size branch. The initial states of the center coordinate branch and the size branch are represented as follows:

[0007]

[0008] The candidate labels Convert to a smooth one-hot vector and map to a continuous semantic embedding to obtain the initial state of the label branch:

[0009] in, This indicates a smooth one-hot encoding operation for the label. Represents the semantic embedding mapping of tags; Step 6: Based on the initial states of the center coordinate branch, the size branch, and the label branch, obtain the initial diffusion state. ; Step 7: For the initial diffusion state The diffusion time steps are randomly sampled and noise is injected to obtain the noisy diffusion state. ; Step 8: Based on the multi-scale feature F and the noisy diffusion state Semantic-geometric joint denoising prediction is performed to obtain the denoising prediction result; Step 9: Utilize the denoised prediction results and the set of true target bounding boxes Real category tag set Calculate the loss function to update the model parameters, and iterate the training until convergence.

[0010] Further, in the method described above, step 3 includes: The multi-scale features Input condition prior network to predict target center heatmap and target size drawing , is represented as:

[0011] in, This indicates a conditional prior network. Represents the heat map of the target center. Represents the target dimensions; From the heat map of the target center Select the top scorers Each response location is used as a candidate target center, and in the size map The width and height information are read from the corresponding positions to form a set of conditional prior boxes. :

[0012] in, This indicates the operation of generating candidate boxes based on the response location and size map information from the heatmap. This indicates the number of candidate centers selected.

[0013] Further, in the method described above, step 6 includes: From diffusion time step set Random sampling time step :

[0014] Gaussian noise was sampled for the center coordinate branch, the size branch, and the label semantic branch, respectively.

[0015] Set separate cumulative noise scheduling functions for the center coordinate branch and the dimension branch:

[0016] in, Indicates the first Each branch at time step The cumulative noise figure, This represents the cosine scheduling offset parameter for the corresponding branch. Indicates the total number of diffusion steps; for each label semantic branch, sets the corresponding label noise scheduling coefficient. ; Forward diffusion noise is applied to each of the three branches to obtain the training time. Noisy joint diffusion state:

[0017]

[0018] Same diffusion time step After the center coordinate branch, size branch, and label semantic branch each complete the positive noise addition, the three are combined into the training time. The noisy joint diffusion state.

[0019] Further, in the method described above, step 7 includes: The time step is obtained by decoding the noisy center coordinate branch and the noisy size branch. Noisy candidate boxes:

[0020] embedding noisy tag semantics By fusing with visual features, fused features are obtained. :

[0021] in, Indicates multi-scale features Visual features extracted from candidate bounding box regions. Represents the characteristic projection matrix, Indicates feature concatenation operation; The diffusion decoder is in the current noisy joint diffusion state. Multi-scale features and time step As input, and combined with by The current candidate box obtained by decoding and fusion characteristics Clean state estimation results for the center coordinate branch, size branch, and label semantic branch:

[0022] in, The parameter is A semantic-geometric joint denoising network.

[0023] Furthermore, in the method described above, the loss function in step 8 is as follows:

[0024] in, Describing set prediction loss This represents the semantic denoising loss of the labels. Indicates prior supervised loss. and These represent the weighting coefficients of the corresponding loss terms; Among them, ensemble prediction loss Based on the predicted bounding boxes and category results, compare them with the true target boxes. Authentic Labels The set matching calculation yielded the following:

[0025]

[0026]

[0027]

[0028] in, Represents the category cross-entropy loss. This represents the bounding box coordinate regression loss. This indicates the generalized intersection and comparison loss. , , These represent the weight coefficients of the corresponding loss terms; M represents the number of real targets participating in the set matching. Indicates the first Category labels for real targets; Indicates the relationship with the first The probability distribution of the predicted candidate target categories that match the real target; This indicates that after set matching, it matches the first... A true target bounding box Matching predicted bounding boxes; Indicates the relationship with the first A predicted bounding box that matches a real target; This represents the x-coordinate of the center point of the predicted bounding box; Indicates the first The x-coordinate of the center point of the actual target bounding box; This represents the ordinate of the center point of the predicted bounding box; Indicates the first The ordinate of the center point of the actual target bounding box; Indicates the width of the prediction bounding box; Indicates the first The width of a real target bounding box; Indicates the height of the predicted bounding box; Indicates the first The height of the actual target bounding box; Label semantic denoising loss as follows:

[0029] Where N represents the number of candidate targets participating in the label semantic denoising training; Indicates the first The candidate target in the first Smooth label values ​​across categories; The first term obtained from the semantic branch prediction of the label The candidate target belongs to the first The probability of each category; Conditional prior monitoring loss Based on the conditional prior box With the true target bounding box The comparison yielded the following results: .

[0030] Furthermore, according to the method described above, the model inference process includes: Acquire the SAR image to be detected; Multi-scale feature extraction is performed on the SAR image to be detected to obtain multi-scale features. ; Based on the multi-scale features Generate conditional prior boxes ; According to the conditional prior box An initial set of candidate boxes is obtained by combining random candidate boxes. ; The initial candidate box set Decoupling coding for center coordinate branch and size branches Initialize the label semantic branch to a Gaussian noise state. ; According to the central coordinate branch and size branches and Gaussian noise state Determine the initial joint diffusion state ; Based on the multi-scale features and initial joint diffusion state Semantic-geometric joint denoising prediction is performed, and the final denoised state is obtained by progressively updating. ; Using the final denoising state The target bounding box is obtained by decoding based on the center coordinate branch and the size branch. Class probability and confidence level; For the target bounding box The results are filtered using category probability and confidence level to obtain the final detection result; the final detection result includes: the target's location bounding box, the target category, and the target's confidence level.

[0031] Furthermore, according to the method described above, the final denoising state The acquisition includes: From the diffusion termination time step Begin, according to Inverse denoising is performed sequentially at each time step. According to the current noisy center branch and size branches Decode to obtain the current candidate box :

[0032] Current candidate box Corresponding visual features With the current noisy tag semantics Embedding is used for fusion to obtain fused features. :

[0033] The diffusion decoder is in the current noisy joint diffusion state. Multi-scale features of the image to be detected and time step As input, and combined within the network by The current candidate box obtained by decoding and its fusion characteristics Predict the clean state estimate for the current time step:

[0034] in, Indicating the first step in the reverse diffusion process The current noisy joint diffusion state at each time step. Indicates the current backdiffusion time step number. , and These represent the clean state estimation results predicted by the network at the current time step; Based on the clean state estimation results, noise estimates for the center coordinate branch, size branch, and label semantic branch are obtained respectively:

[0035] in, Indicates the branch type, including: central branch ,size and tag semantic branches ; For noise scheduling parameters, This indicates the proportion of clean states retained in the noisy states within the current branch. This indicates the proportion of noise components in the current branch; Perform backdiffusion updates on the center coordinate branch, size branch, and label semantic branch respectively to finally obtain the denoised state. :

[0036] in, This represents the update operation of a single branch at a back-diffusion time step. By repeatedly executing the above back-diffusion update, the initial noisy state is gradually restored to a clean detection state. This indicates a splicing operation, used to recombine the update results of the three branches into a complete joint diffusion state. , and These represent the current time step. The state of the center coordinate branch, the size branch, and the label semantic branch.

[0037] Furthermore, in the method described above, the utilization of the final denoised state... The target bounding box is obtained by decoding based on the center coordinate branch and the size branch. Class probabilities and confidence levels include: Using the final denoising state Decode the target bounding box based on the center coordinate branch and the size branch:

[0038] in, This represents the final set of predicted target bounding boxes. Each predicted target box can be represented as:

[0039] in, Indicates the first One predicted target bounding box, This indicates the coordinates of the center point of the predicted target bounding box. These represent the width and height of the predicted target bounding box, respectively. Predicting category probabilities based on label semantic branches:

[0040] in, Indicates the first The probability distribution of each category for each candidate target, the th... The predicted categories and corresponding confidence levels of each candidate target are represented as follows:

[0041]

[0042] in, Indicates the prediction category, Indicates the prediction confidence level. Indicates the first The candidate target belongs to the first The probability of a class.

[0043] Furthermore, in the method described above, the target bounding box... The results are filtered using category probability and confidence level to obtain the final detection results, which include: All candidate detection results are filtered using a confidence threshold to remove candidate targets with confidence scores below a preset threshold; duplicate detection boxes are removed using non-maximum suppression or ensemble prediction filtering strategies to obtain the final detection result set.

[0044] in, This represents the final set of detection results. Indicates the final target number to retain. Indicates the first The bounding box of the target's location. Indicates the first The categories of targets, Indicates the first The confidence level of each objective.

[0045] This application also provides a decoupled dual-space directional diffusion system for remote sensing target detection in SAR images, including: a model training module and a model inference module; The model training module includes: The acquisition unit is used to acquire labeled SAR image training samples; the training samples include: a set of SAR images and a set of real target bounding boxes corresponding to each SAR image. Real category tag set ; The extraction unit is used to extract multi-scale features from the labeled SAR image to obtain multi-scale features F; The generation unit is used to generate conditional priors from the multi-scale features F to obtain conditional prior boxes. ; Construction unit, used to construct based on the conditional prior box Construct a candidate box set B; The allocation unit is used to allocate according to the set of real category labels. Assign a category label to each candidate box to obtain the candidate labels. ; A splitting unit is used to split each bounding box in the candidate box set B into a center coordinate branch and a size branch. The initial states of the center coordinate branch and the size branch are respectively represented as follows:

[0046]

[0047] Transformation mapping unit, used to transform the candidate tags Convert to a smooth one-hot vector and map to a continuous semantic embedding to obtain the initial state of the label branch:

[0048] in, This indicates a smooth one-hot encoding operation for the label. Represents the semantic embedding mapping of tags; The joint unit is used to obtain the initial diffusion state based on the initial states of the center coordinate branch, the size branch, and the label branch. ; Diffusion unit, used for the initial diffusion state The diffusion time steps are randomly sampled and noise is injected to obtain the noisy diffusion state. ; A denoising unit is configured to, based on the multi-scale feature F and the noisy diffusion state... Semantic-geometric joint denoising prediction is performed to obtain the denoising prediction result; A computing unit is used to utilize the denoised prediction results and the set of true target bounding boxes. Real category tag set Calculate the loss function to update the model parameters, and iterate the training until convergence.

[0049] Beneficial effects: This invention provides a decoupled dual-space directional diffusion method and system for remote sensing target detection in SAR images. Addressing issues such as strong speckle noise, complex background clutter, and inconsistent target geometric parameter distribution in SAR images, the method splits the bounding box into a center coordinate branch and a size branch, setting independent noise scheduling functions for each branch. It combines conditional prior generation, embedding noisy label semantics with visual features, and a diffusion decoder simultaneously predicts the clean state of the center coordinates, size, and category label, as well as constructing an initial state using a mixture of conditional prior boxes and random boxes. It then performs iterative back-diffusion denoising, ensuring clear step boundaries and collaborative relationships between image input, feature extraction, conditional prior generation, decoupled noise addition and denoising, parameter updates, and result output. The introduction of joint constraints from category semantics and conditional prior boxes during denoising improves the accuracy and inference stability of SAR target detection in complex scenarios, reducing the risk of missed and false detections. Furthermore, the entire method is easily deployed on servers, edge computing devices, or dedicated detection terminals, enhancing the completeness, feasibility, and engineering application value of the technical solution. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the decoupled dual-space directional diffusion method for remote sensing target detection in SAR images provided in this application; Figure 2 This is a schematic diagram illustrating the detection results of the method provided in this application for different input images. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0053] Currently, existing diffusion-based detection methods have the following shortcomings when it comes to SAR target detection: First, there is significant heterogeneity in the statistical distribution of the bounding box center coordinates and size parameters. The center coordinates are usually more concentrated, while the width and height parameters show a stronger long-tail distribution. If a uniform noise schedule is applied to all parameters, some parameters will be excessively disturbed and some parameters will be insufficiently disturbed, which will affect the denoising quality and result in significant changes in the predicted target scale.

[0054] Second, existing methods often treat semantic classification as an auxiliary branch independent of the diffusion state, making it difficult to fully utilize category information to constrain the localization process during high-noise stages, resulting in insufficient coupling between category recognition and localization.

[0055] Third, if inference starts entirely from an unconditional random box, the search space in the early stages of backdiffusion is too large, which can easily lead to high sampling variance, slow convergence speed, and early candidate regions deviating from the true target, resulting in strong speckle noise and complex background clutter.

[0056] To address this, this application provides a decoupled dual-space directional diffusion method and system for SAR image remote sensing target detection, in order to solve the accuracy and stability problems of existing SAR target detection technologies under conditions of strong speckle noise, complex background clutter, significant target scale variations, and insufficient coupling between category recognition and positioning.

[0057] Figure 1 The decoupled dual-space directional diffusion method for SAR image remote sensing target detection provided in this application comprises two parts: a model training process and a model inference process. The model training process enables the network to learn to recover the true target bounding box and target category from noisy detection states; the model inference process inputs unlabeled SAR images into the trained model and outputs the target location, target category, and corresponding confidence score.

[0058] The model training process specifically includes the following steps: Step 1: Obtain labeled SAR image training samples.

[0059] Specifically, obtain labeled SAR images from publicly available datasets. True target bounding box set and the set of real category labels Each ground truth target bounding box is represented as:

[0060] in, and Indicates the coordinates of the center point of the target bounding box. and These represent the width and height of the target bounding box, respectively; each ground truth category label indicates the category to which the corresponding target belongs.

[0061] Step 2: Based on the labeled SAR image, perform multi-scale feature extraction to obtain multi-scale features F.

[0062] Specifically, SAR images Input the backbone network and the feature pyramid network to extract multi-scale feature representations with different spatial resolutions and semantic levels. , is represented as:

[0063] in, Indicates the input SAR image, This represents the multi-scale features of the output. This represents the feature extraction map, which consists of a backbone network and a feature pyramid network. This feature is used to provide information about texture, scattering structure, target regions, and background in the detected image.

[0064] Step 3: Generate conditional priors for the multi-scale features F to obtain conditional prior boxes. And calculate the prior supervised loss. .

[0065] Multiscale features Input condition prior network to predict target center heatmap and target size drawing , is represented as:

[0066] in, This indicates a conditional prior network. Represents the heat map of the target center. This represents the target dimension drawing.

[0067] Subsequently, from the target center heat map Select the top scorers Each response location is used as a candidate target center, and in the size map The width and height information are read from the corresponding positions to form a set of conditional prior boxes. :

[0068] in, This indicates the operation of generating candidate boxes based on the response location and size map information from the heatmap. This indicates the number of candidate centers selected.

[0069] Conditional Prior Box With the true target bounding box Compare and calculate the prior supervised loss under the given conditions:

[0070] The method provided in this application generates conditional priors by using multi-scale features F, enabling the conditional prior network to learn the approximate center location and size information of targets in SAR images, thereby providing more reliable initial candidate regions for the subsequent diffusion process.

[0071] Step 4: Based on the aforementioned conditional prior box Construct a training candidate box set B, and based on the true category label set... Assign category labels to candidate boxes to obtain candidate labels. .

[0072] Specifically, the conditional prior boxes are first mixed with random candidate boxes to form a candidate box set for training the diffusion detector:

[0073] in, This represents a set of randomly generated candidate boxes. This indicates the total number of candidate boxes used during the training phase.

[0074] Then, based on the candidate boxes With the true target bounding box The matching relationship between them is used to assign a corresponding training category label to each candidate box:

[0075] in, This indicates the candidate box label assignment operation. This operation is based on the candidate box set. With the set of true target boxes The matching relationship between them will determine the true category labels. The candidate boxes are assigned to corresponding candidate boxes. If a candidate box successfully matches the ground truth box, its training label is the category label corresponding to the matching ground truth box; if it fails to match, the candidate box is labeled as the background category. This results in a set of candidate boxes. The corresponding candidate training label set .

[0076] Step 5: Analyze the candidate box set B and the candidate labels. Perform bounding box decoupling encoding and label semantic encoding to obtain the initial diffusion state. .

[0077] Specifically, the candidate boxes Each bounding box in the code is split into a center coordinate branch and a dimension branch, that is, the bounding box... Decomposed into center coordinates and dimensional parameters .

[0078] The center coordinate branch and the dimension branch are represented as follows:

[0079]

[0080] in, This represents the initial state of the central coordinate branch. This indicates the initial state of the size branch.

[0081] candidate labels Convert to a smooth one-hot vector and map to a continuous semantic embedding to obtain the initial state of the label branch:

[0082] in, This indicates a smooth one-hot encoding operation for the label. This represents the semantic embedding mapping of tags.

[0083] This yields the initial diffusion state, which is composed of the center coordinates, size parameters, and label semantics:

[0084] Traditional diffusion detection methods typically only diffuse and denoise the bounding box coordinates, treating the target category as a separate, post-processing sub-branch. This results in a lack of guidance from category semantic information during high-noise stages, failing to effectively constrain the bounding box regression process and causing localization errors. The method provided in this application addresses this issue by using candidate labels... Mapping to continuous semantic embeddings and relating them to geometric states , The initial diffusion state, composed of both elements, makes category semantic information no longer secondary, but a core state component equally important as geometric information, participating in the entire diffusion and denoising process. This solves the problem of insufficient coupling between category recognition and localization. Furthermore, by decoupling the bounding box into two independent branches—center coordinates and size parameters—it creates conditions for setting differentiated noise scheduling strategies for branches with different statistical distributions, avoiding the perturbation mismatch problem caused by uniform scheduling and improving the detection accuracy of multi-scale targets.

[0085] Step 6: Initial diffusion state The diffusion time steps are randomly sampled and noise is injected to obtain the noisy diffusion state. .

[0086] Specifically, from the diffusion time step set Random sampling time step :

[0087] Gaussian noise was sampled for the center coordinate branch, the size branch, and the label semantic branch, respectively.

[0088] Since the center coordinates and size parameters differ in their statistical distributions, separate cumulative noise scheduling functions are set for the center coordinate branch and the size branch respectively:

[0089] in, Indicates the first Each branch at time step The cumulative noise figure, This represents the cosine scheduling offset parameter for the corresponding branch. This represents the total number of diffusion steps. For the label semantic branch, the corresponding label noise scheduling coefficient is set. .

[0090] Subsequently, forward diffusion noise addition is performed on each of the three branches:

[0091] For the bounding box geometric branches, the decoupled forward diffusion process can also be represented as:

[0092] For the semantic branch of the tag, its forward diffusion process can be represented as:

[0093] This step yields the training time. Noisy joint diffusion state:

[0094] At the same diffusion time step After the center coordinate branch, size branch, and label semantic branch each complete the positive noise addition, the three are combined into the training time. The noisy joint diffusion state.

[0095] Step 7: Analyze the multi-scale feature F and the noisy diffusion state. Perform semantic-geometric joint denoising prediction to obtain the denoising prediction result.

[0096] Specifically, firstly, the time step is obtained by decoding based on the noisy center coordinate branch and the noisy size branch. Noisy candidate boxes:

[0097] Then, the noisy label semantics are embedded. By fusing with visual features, the fused features are obtained:

[0098] in, Indicates multi-scale features Visual features extracted from candidate bounding box regions. Represents the characteristic projection matrix, This indicates a feature splicing operation.

[0099] The diffusion decoder is in the current noisy joint diffusion state. Multi-scale features and time step As input, and combined with by The current candidate box obtained by decoding and fusion characteristics Clean state estimation results for the center coordinate branch, size branch, and label semantic branch:

[0100] in, The parameter is The diffusion detection network.

[0101] Step 8: Calculate the training loss and update the model parameters based on the denoising prediction results.

[0102] Specifically, this application compares the bounding boxes and category results predicted by the model with the true target boxes. Authentic Labels Perform set matching and calculate set prediction loss:

[0103] in

[0104]

[0105]

[0106] in, Represents the category cross-entropy loss. This represents the bounding box coordinate regression loss. This indicates the generalized intersection and comparison loss. , , These represent the weight coefficients of the corresponding loss terms; M represents the number of real targets participating in the set matching. Indicates the first Category labels for real targets; Indicates the relationship with the first The probability distribution of the predicted candidate target categories that match the real target; This indicates that after set matching, it matches the first... A true target bounding box Matching predicted bounding boxes; Indicates the relationship with the first A predicted bounding box that matches a real target; This represents the x-coordinate of the center point of the predicted bounding box; Indicates the first The x-coordinate of the center point of the actual target bounding box; This represents the ordinate of the center point of the predicted bounding box; Indicates the first The ordinate of the center point of the actual target bounding box; Indicates the width of the prediction bounding box; Indicates the first The width of a real target bounding box; Indicates the height of the predicted bounding box; Indicates the first The height of the actual target bounding box.

[0107] Combining the label denoising loss and the conditional prior supervision loss, we obtain the total model training loss:

[0108]

[0109] Where N represents the number of candidate targets participating in the label semantic denoising training; Indicates the first The candidate target in the first Smooth label values ​​across categories; The first term obtained from the semantic branch prediction of the label The candidate target belongs to the first The probability of each category, This represents the semantic denoising loss of the labels. Indicates prior supervised loss. and These represent the weight coefficients for the corresponding loss terms. Based on the total loss... For model parameters Perform reverse propagation and optimization updates.

[0110] Step 9: Repeat steps 1 to 8 until the model training converges, and obtain the trained SAR target detection model.

[0111] Specifically, during the training process, labeled SAR image samples are continuously input, and multi-scale feature extraction, conditional prior generation, diffusion noise addition, joint denoising prediction, and loss optimization are repeatedly performed until the model's detection performance on the validation set reaches the preset requirements or the number of training iterations reaches the preset threshold. The model parameters after training are saved, thereby obtaining the trained SAR target detection model, which is a decoupled dual-space directional diffusion model capable of performing SAR image target detection.

[0112] The model inference process provided in this application specifically includes the following steps. Step 1: Obtain unlabeled SAR images.

[0113] Specifically, acquire the unlabeled SAR image to be detected. The image does not contain manually annotated target bounding boxes and category labels.

[0114] Step 2: Perform multi-scale feature extraction on the unlabeled SAR image to obtain multi-scale features. .

[0115] Specifically, unlabeled SAR images Input the trained backbone network and feature pyramid network to obtain multi-scale features:

[0116] in, This represents the multi-scale features corresponding to the image to be detected. These multi-scale features are used for subsequent conditional prior box generation and backdiffusion denoising.

[0117] Step 3: Based on multi-scale features Predict the target center heatmap H and size map S, and generate conditional prior boxes. ; Multiscale features Input a priori network to obtain the target center heatmap and size map:

[0118] From the heat map of the target center Select the top scorers Each response location is used as a candidate target center, and in the size map Read the width and height of the corresponding position to form a conditional prior box:

[0119] Each conditional prior box can be represented as:

[0120] in, Indicates the first The center location of each candidate target and This indicates the width and height of the candidate target. This step allows the inference process to begin with candidate boxes that are closer to the real target region, rather than starting entirely with unconditional random boxes.

[0121] Step 4: Utilize conditional prior boxes An initial set of candidate boxes is constructed by combining random candidate boxes.

[0122] Specifically, let the total number of candidate boxes used in the inference phase be... , which includes A conditional prior box and A random candidate box, satisfying:

[0123] Concatenate the conditional prior boxes and random candidate boxes to obtain the initial candidate box set:

[0124] in, Indicates the selection of the preceding condition from the set of conditional prior boxes. A priori box, Indicates random generation One candidate box, This indicates a stitching operation. The proportion of the prior bounding box can be represented as:

[0125] Subsequently, the initial candidate box set is decoupled and encoded into a center coordinate branch and a size branch:

[0126] Since there are no true category labels during the inference phase, the label semantic branch is initialized to a Gaussian noise state:

[0127] Finally, the initial joint diffusion state of the inference phase is obtained:

[0128] Step 5: Process multi-scale features and initial diffusion state Perform back-diffusion iterative denoising to gradually update the results. .

[0129] Specifically, from the diffusion termination time step Begin, according to The reverse denoising is performed sequentially. At each time step... First, the current candidate box is obtained by decoding based on the current noisy center branch and size branch:

[0130] Then, the visual features corresponding to the current candidate box are fused with the semantic embedding of the noisy label:

[0131] The diffusion decoder is in the current noisy joint diffusion state. Multi-scale features of the image to be detected and time step As input, and combined within the network by The current candidate box obtained by decoding and its fusion characteristics Predict the clean state estimate for the current time step:

[0132] in, This represents the completed training of the diffusion detection network, i.e., the semantic-geometric joint denoising network; where... This indicates the network parameters obtained during the training phase through backpropagation optimization in steps 8 and 9. Indicating the first step in the reverse diffusion process The current noisy joint diffusion state at each time step. This indicates the current backdiffusion time step number. , and These represent the clean state estimates predicted by the network at the current time step.

[0133] Based on the prediction results, noise estimates for the center coordinate branch, size branch, and label semantic branch are obtained respectively:

[0134] in, Indicates the branch type, which can be taken as... , or .when When, this formula is used for the central coordinate branch; when When, this formula is used for dimension branches; when When this formula is used for tag semantic branches, it indicates the first... The time step The current noisy state of each branch. For the center coordinate branch, Indicates the current noisy center coordinate state; for size branches, This indicates the current noisy width and height dimensions; for the tag semantic branch, This represents the current noisy category semantic embedding state. These states were initially determined by inference step 4. The initialization is obtained and continuously updated during the back diffusion process in reasoning step 5.

[0135] This indicates the proportion of clean states retained in the noisy states within the current branch. This indicates the proportion of noise components in the current branch. Both quantities are determined by the noise scheduling parameters. Calculated.

[0136] Then, perform backdiffusion updates on the center coordinate branch, size branch, and label semantic branch respectively:

[0137] in, This represents the update operation of a single branch at a backdiffusion time step. By repeatedly performing the above backdiffusion update, the model gradually restores the initial noisy state to a clean detection state. This indicates a splicing operation, used to recombine the update results of the three branches into a complete joint diffusion state. and These represent the current time step. The center coordinate branch, size branch, and label semantic branch states. They originate from the current joint state. At the initial moment of inference, these branches are constructed by inference step 4; in subsequent iterations, they are obtained by updating the results of the previous round of backdiffusion.

[0138] Step 6: Based on the final noise reduction status The target bounding box, class probability, and confidence score are obtained by decoding. After the backdiffusion iteration is completed, the final denoised state is obtained.

[0139] Specifically, based on the final denoising state Decode the target bounding box based on the center coordinate branch and the size branch:

[0140] in, This represents the set of target bounding boxes predicted by the model. Each predicted target box can be represented as:

[0141] in, Indicates the first Predicted target bounding boxes. Superscript This indicates the candidate target's number, or which predicted bounding box it is. The "..." above the symbol... "" indicates that the bounding box is the result predicted by the model.

[0142] Predicting category probabilities based on label semantic branches:

[0143] in, Indicates the first The probability distribution of each category for each candidate target. The predicted categories and corresponding confidence levels of each candidate target are represented as follows:

[0144]

[0145] in, Indicates the prediction category, This represents the prediction confidence level. Among them, Indicates category index, Indicates the first The candidate target belongs to the first The probability of a class.

[0146] Step 7: Based on the obtained target bounding box, class probability, and confidence level, filter the results to obtain the final detection result.

[0147] Specifically, a confidence threshold is applied to all candidate detection results to remove candidate targets with confidence scores below a preset threshold. Furthermore, non-maximum suppression or ensemble prediction filtering strategies can be used to remove duplicate detection boxes, resulting in the final set of detection results.

[0148] in, This represents the final set of detection results. Indicates the final target number to retain. Indicates the first The bounding box of the target's location. Indicates the first The categories of targets, Indicates the first The confidence level of each objective.

[0149] This application, through the aforementioned training and inference processes, enables the joint prediction of target location, target category, and corresponding confidence level in SAR images. During the training phase, decoupled dual-space diffusion modeling using center coordinate branches, size branches, and label semantic branches allows the model to learn stable target bounding box recovery and category semantic recovery capabilities. During the inference phase, conditional prior boxes guide initial sampling, allowing the back-diffusion process to begin from more reliable candidate regions, thereby improving the accuracy and stability of SAR remote sensing target detection.

[0150] The method provided in this application predicts target center heatmaps and size maps based on multi-scale features, and selects candidate target centers from high-response locations. These are combined with corresponding width and height information to form conditional prior boxes, thus providing more reliable initial candidate regions for the subsequent backdiffusion process. The bounding box state is split into center coordinate branches and size branches, and diffusion modeling is performed using independent noise scheduling strategies for each. Simultaneously, category labels are mapped to continuous semantic embeddings and incorporated into the diffusion state, achieving joint modeling and collaborative denoising of bounding box geometric information and category semantic information, effectively improving the accuracy of target detection. The designed network is trained using a publicly available SAR target detection dataset. During training, diffusion time steps are randomly sampled, and noise is injected into the center branch, size branch, and label branch. The ensemble prediction loss, label denoising loss, and conditional prior loss are jointly optimized, enabling the network to accurately recover target location and category information under complex backgrounds and strong noise interference. By using a mixture of conditional prior boxes and random candidate boxes to construct the initial inference state, the joint recovery of target boxes and categories is gradually completed through backdiffusion iteration. The final detection result is further improved by combining confidence filtering or non-maximum suppression.

[0151] Figure 2 The diagram illustrates the detection results of the method provided in this application for different input images. From left to right, each column shows: the input image, the ground truth bounding box, and the bounding box detected by the method of this application. From top to bottom, each row shows: from... Figure 2 It can be seen that the method provided in this application can improve the accuracy and inference stability of SAR target detection in complex scenarios, and reduce the risk of missed detection and false detection. The present invention also provides a decoupled dual-space directional diffusion system for remote sensing target detection in SAR images, comprising: a model training module and a model inference module; The model training module includes: The acquisition unit is used to acquire labeled SAR image training samples; the training samples include: a set of SAR images and a set of real target bounding boxes corresponding to each SAR image. Real category tag set ; The extraction unit is used to extract multi-scale features from the labeled SAR image to obtain multi-scale features F; The generation unit is used to generate conditional priors from the multi-scale features F to obtain conditional prior boxes. ; Construction unit, used to construct based on the conditional prior box Construct a candidate box set B; The allocation unit is used to allocate according to the set of real category labels. Assign a category label to each candidate box to obtain the candidate labels. ; A splitting unit is used to split each bounding box in the candidate box set B into a center coordinate branch and a size branch. The initial states of the center coordinate branch and the size branch are respectively represented as follows:

[0152]

[0153] Transformation mapping unit, used to transform the candidate tags Convert to a smooth one-hot vector and map to a continuous semantic embedding to obtain the initial state of the label branch:

[0154] in, This indicates a smooth one-hot encoding operation for the label. Represents the semantic embedding mapping of tags; The joint unit is used to obtain the initial diffusion state based on the initial states of the center coordinate branch, the size branch, and the label branch. ; Diffusion unit, used for the initial diffusion state The diffusion time steps are randomly sampled and noise is injected to obtain the noisy diffusion state. ; A denoising unit is configured to, based on the multi-scale feature F and the noisy diffusion state... Semantic-geometric joint denoising prediction is performed to obtain the denoising prediction result; A computing unit is used to utilize the denoised prediction results and the set of true target bounding boxes. Real category tag set Calculate the loss function to update the model parameters, and iterate the training until convergence.

[0155] Experimental example: This experimental example uses a training platform equipped with an NVIDIA RTX 3090 GPU and is simulated using Python software and the PyTorch deep learning framework. The system employs ResNet-101 as the backbone network, combined with FPN for multi-scale feature extraction. Training is performed on three RTX 3090 GPUs, using AdamW as the optimizer with a learning rate of 2.5 × 10⁻⁵, weight decay of 1.0 × 10⁻⁴, a total batch size of 12, and approximately 450k training iterations.

[0156] This experimental example uses the SSDD and SAR-AIRcraft-1.0 datasets for algorithm validation. The SSDD dataset is a commonly used SAR ship detection dataset, containing 1160 images; the SAR-AIRcraft-1.0 dataset is a high-resolution SAR aircraft detection dataset, containing 4368 images and 16463 labeled aircraft instances, covering 7 categories. The experiment follows the standard data partitioning and evaluation protocols of relevant public benchmarks. To demonstrate the effectiveness of the proposed method, Recall, Precision, F1, and AP50 were used as evaluation criteria on the public SSDD dataset, and AP50 and AP75 were used as evaluation criteria on the SAR-AIRcraft-1.0 dataset. This experimental example selects DiffusionDet, DiffDet4SAR, and RT-DETRv3 networks as representative comparison algorithms. The DiffusionDet network was proposed by "Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo, DiffusionDet: Diffusion Model for ObjectDetection, in Proceedings of the IEEE / CVF International Conference on Computer Vision (ICCV), 2023, pp. 19830–19843." The DiffDet4SAR network was proposed by "JieZhou, Chao Xiao, Bo Peng, Zhen Liu, Li Liu, Yongxiang Liu, and Xiang Li, DiffDet4SAR: Diffusion-Based Aircraft Target Detection Network for SARImages, IEEE Geoscience and Remote Sensing Letters, vol. 21, pp. 1–5, 2024." The RT-DETRv3 network was proposed by “S. Wang, C. Xia, F. Lv, and Y. Shi, RT-DETRv3: Real-time End-to-End Object Detection with Hierarchical Dense PositiveSupervision, arXiv preprint arXiv:2409.08475, 2024.”By comparing the values ​​of the above evaluation metrics, the performance of various algorithms in SAR target detection tasks can be intuitively assessed. The comparison results are shown in Tables 1 and 2. Table 1 shows the comparison results of different detection algorithms on the SAR-AIRcraft-1.0 dataset.

[0157] Table 2 shows the comparison results of different detection algorithms on the SSDD dataset.

[0158] Generally speaking, higher values ​​for AP50, AP75, R, P, and F1 indicate better algorithm performance on the detection task. As shown in Tables 1 and 2, on two publicly available datasets, the method provided in this application outperforms other comparative algorithms on multiple evaluation metrics.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A decoupled double spatial direction diffusion method for SAR image remote sensing target detection, characterized in that, include: Model training process and model inference process; The model training process includes: Step 1: obtain a labeled SAR image training sample; the training sample comprises: a SAR image set, a real target boundary box set corresponding to each SAR image , a real class label set ​ Step 2: Perform multi-scale feature extraction on the labeled SAR image to obtain multi-scale features F; Step 3: Generate conditional priors for the multi-scale features F to obtain conditional prior boxes. ; Step 4: Based on the aforementioned conditional prior box Construct a candidate box set B; and based on the true category label set... Assign a category label to each candidate box to obtain the candidate labels. ; Step 5: Split each bounding box in the candidate box set B into a center coordinate branch and a size branch. The initial states of the center coordinate branch and the size branch are represented as follows: The candidate labels Convert to a smooth one-hot vector and map to a continuous semantic embedding to obtain the initial state of the label branch: in, This indicates a smooth one-hot encoding operation for the label. Represents the semantic embedding mapping of tags; Step 6: Based on the initial states of the center coordinate branch, the size branch, and the label branch, obtain the initial diffusion state. ; Step 7: For the initial diffusion state The diffusion time steps are randomly sampled and noise is injected to obtain the noisy diffusion state. ; Step 8: Based on the multi-scale feature F and the noisy diffusion state Semantic-geometric joint denoising prediction is performed to obtain the denoising prediction result; Step 9: Utilize the denoised prediction results and the set of true target bounding boxes Real category tag set Calculate the loss function to update the model parameters, and iterate the training until convergence.

2. The method according to claim 1, characterized in that, Step 3 includes: The multi-scale features Input condition prior network to predict target center heatmap and target size drawing , is represented as: in, This indicates a conditional prior network. Represents the heat map of the target center. Represents the target dimensions; From the heat map of the target center Select the top scorers Each response location is used as a candidate target center, and in the size map The width and height information are read from the corresponding positions to form a set of conditional prior boxes. : in, This indicates the operation of generating candidate boxes based on the response location and size map information from the heatmap. This indicates the number of candidate centers selected.

3. The method according to claim 1, characterized in that, Step 6 includes: From diffusion time step set Random sampling time step : Gaussian noise was sampled for the center coordinate branch, the size branch, and the label semantic branch, respectively. Set separate cumulative noise scheduling functions for the center coordinate branch and the dimension branch: in, Indicates the first Each branch at time step The cumulative noise figure, This represents the cosine scheduling offset parameter for the corresponding branch. Indicates the total number of diffusion steps; for each label semantic branch, sets the corresponding label noise scheduling coefficient. ; Forward diffusion noise is applied to each of the three branches to obtain the training time. Noisy joint diffusion state: Same diffusion time step After the center coordinate branch, size branch, and label semantic branch each complete the positive noise addition, the three are combined into the training time. The noisy joint diffusion state.

4. The method according to claim 3, characterized in that, Step 7 includes: The time step is obtained by decoding the noisy center coordinate branch and the noisy size branch. Noisy candidate boxes: Embed noisy tag semantics By fusing with visual features, fused features are obtained. : in, Indicates multi-scale features Visual features extracted from candidate bounding box regions. Represents the characteristic projection matrix, Indicates feature concatenation operation; The diffusion decoder is in the current noisy joint diffusion state. Multi-scale features and time step As input, and combined with by The current candidate box obtained by decoding and fusion characteristics Clean state estimation results for the center coordinate branch, size branch, and label semantic branch: in, The parameter is A semantic-geometric joint denoising network.

5. The method according to claim 4, characterized in that, The loss function in step 8 is as follows: in, Describing set prediction loss This represents the semantic denoising loss of the labels. Indicates prior supervised loss. and These represent the weighting coefficients of the corresponding loss terms; Among them, ensemble prediction loss Based on the predicted bounding boxes and category results, compare them with the true target boxes. Authentic Labels The set matching calculation yielded the following: in, Represents the cross-entropy loss of categories. This represents the bounding box coordinate regression loss. This indicates the generalized intersection and comparison loss. , , These represent the weight coefficients of the corresponding loss terms; M represents the number of real targets participating in the set matching. Indicates the first Category labels for real targets; Indicates the relationship with the first The probability distribution of the predicted candidate target categories that match the real target; This indicates that after set matching, it matches the first... A true target bounding box Matching predicted bounding boxes; Indicates the relationship with the first A predicted bounding box that matches a real target; This represents the x-coordinate of the center point of the predicted bounding box; Indicates the first The x-coordinate of the center point of the actual target bounding box; This represents the ordinate of the center point of the predicted bounding box; Indicates the first The ordinate of the center point of the actual target bounding box; Indicates the width of the prediction bounding box; Indicates the first The width of a real target bounding box; Indicates the height of the predicted bounding box; Indicates the first The height of the actual target bounding box; Label semantic denoising loss as follows: Where N represents the number of candidate targets participating in the label semantic denoising training; Indicates the first The candidate target in the first Smooth label values ​​across categories; The first term obtained from the semantic branch prediction of the label The candidate target belongs to the first The probability of each category; Conditional prior monitoring loss Based on the conditional prior box With the true target bounding box The comparison yielded the following results: 。 6. The method according to claim 1, characterized in that, The model reasoning process includes: Acquire the SAR image to be detected; Multi-scale feature extraction is performed on the SAR image to be detected to obtain multi-scale features. ; Based on the multi-scale features Generate conditional prior boxes ; According to the conditional prior box An initial set of candidate boxes is obtained by combining random candidate boxes. ; The initial candidate box set Decoupling coding for center coordinate branch and size branches Initialize the label semantic branch to a Gaussian noise state. ; According to the central coordinate branch and size branches and Gaussian noise state Determine the initial joint diffusion state ; Based on the multi-scale features and initial joint diffusion state Semantic-geometric joint denoising prediction is performed, and the final denoised state is obtained by progressively updating. ; Using the final denoising state The target bounding box is obtained by decoding based on the center coordinate branch and the size branch. Class probability and confidence level; For the target bounding box The results are filtered using category probability and confidence level to obtain the final detection result; the final detection result includes: the target's location bounding box, the target category, and the target's confidence level.

7. The method according to claim 6, characterized in that, The final denoising state The acquisition includes: From the diffusion termination time step Begin, according to Inverse denoising is performed sequentially at each time step. According to the current noisy center branch and size branches Decode to obtain the current candidate box : Current candidate box Corresponding visual features With the current noisy tag semantics Embedding is used for fusion to obtain fused features. : The diffusion decoder is in the current noisy joint diffusion state. Multi-scale features of the image to be detected and time step As input, and combined within the network by The current candidate box obtained by decoding and its fusion characteristics Predict the clean state estimate for the current time step: in, Indicating the first step in the reverse diffusion process The current noisy joint diffusion state at each time step. Indicates the current backdiffusion time step number. , and These represent the clean state estimation results predicted by the network at the current time step; Based on the clean state estimation results, noise estimates for the center coordinate branch, size branch, and label semantic branch are obtained respectively: in, Indicates the branch type, including: central branch ,size and tag semantic branches ; For noise scheduling parameters, This indicates the proportion of clean states retained in the noisy states within the current branch. This indicates the proportion of noise components in the current branch; Perform backdiffusion updates on the center coordinate branch, size branch, and label semantic branch respectively to finally obtain the denoised state. : in, This represents the update operation of a single branch at a back-diffusion time step. By repeatedly executing the above back-diffusion update, the initial noisy state is gradually restored to a clean detection state. This indicates a splicing operation, used to recombine the update results of the three branches into a complete joint diffusion state. , and These represent the current time step. The state of the center coordinate branch, the size branch, and the label semantic branch.

8. The method according to claim 7, characterized in that, The use of the final denoising state The target bounding box is obtained by decoding based on the center coordinate branch and the size branch. Class probabilities and confidence levels include: Using the final denoising state Decode the target bounding box based on the center coordinate branch and the size branch: in, This represents the final set of predicted target bounding boxes. Each predicted target box can be represented as: in, Indicates the first One predicted target bounding box, This indicates the coordinates of the center point of the predicted target bounding box. These represent the width and height of the predicted target bounding box, respectively; Predicting category probabilities based on label semantic branches: in, Indicates the first The probability distribution of each category for each candidate target, the th... The predicted categories and corresponding confidence levels of each candidate target are represented as follows: in, Indicates the prediction category, Indicates the prediction confidence level. Indicates the first The candidate target belongs to the first The probability of a class.

9. The method according to claim 8, characterized in that, The target bounding box The results are filtered using category probability and confidence level to obtain the final detection results, which include: All candidate detection results are filtered using a confidence threshold to remove candidate targets with confidence scores below a preset threshold; duplicate detection boxes are removed using non-maximum suppression or ensemble prediction filtering strategies to obtain the final detection result set. in, This represents the final set of detection results. Indicates the final target number to be retained. Indicates the first The bounding box of the target's location. Indicates the first The categories of targets, Indicates the first The confidence level of each objective.

10. A decoupled dual-space directional diffusion system for target detection in SAR image remote sensing, characterized in that, include: Model training module and model inference module; The model training module includes: The acquisition unit is used to acquire labeled SAR image training samples; the training samples include: a set of SAR images and a set of real target bounding boxes corresponding to each SAR image. Real category tag set ; The extraction unit is used to extract multi-scale features from the labeled SAR image to obtain multi-scale features F; The generation unit is used to generate conditional priors from the multi-scale features F to obtain conditional prior boxes. ; Construction unit, used to construct based on the conditional prior box Construct a candidate box set B; The allocation unit is used to allocate according to the set of real category labels. Assign a category label to each candidate box to obtain the candidate labels. ; A splitting unit is used to split each bounding box in the candidate box set B into a center coordinate branch and a size branch. The initial states of the center coordinate branch and the size branch are respectively represented as follows: Transformation mapping unit, used to transform the candidate tags Convert to a smooth one-hot vector and map to a continuous semantic embedding to obtain the initial state of the label branch: in, This indicates a smooth one-hot encoding operation for the label. Represents the semantic embedding mapping of tags; The joint unit is used to obtain the initial diffusion state based on the initial states of the center coordinate branch, the size branch, and the label branch. ; Diffusion unit, used for the initial diffusion state The diffusion time steps are randomly sampled and noise is injected to obtain the noisy diffusion state. ; A denoising unit is configured to, based on the multi-scale feature F and the noisy diffusion state... Semantic-geometric joint denoising prediction is performed to obtain the denoising prediction result; A computing unit is used to utilize the denoised prediction results and the set of true target bounding boxes. Real category tag set Calculate the loss function to update the model parameters, and iterate the training until convergence.

Images