CN-122023785-A - SAR image speckle perception rotation target detection method based on physical prior

CN122023785ACN 122023785 ACN122023785 ACN 122023785ACN-122023785-A

Abstract

The invention relates to the technical field of SAR data processing and target detection, and discloses a SAR image speckle perception rotation target detection method based on physical priori. And constructing a conditional diffusion model by combining the equivalent noise layer of the SAR image, the physical priori such as contrast and the like, modeling the rotation frame prediction as a conditional diffusion denoising process, and guiding the gradual optimization of frame parameters by using the physical priori. And the Gaussian-gas tan distance is introduced as regression loss, and the problems of angular period discontinuity and deformation are effectively relieved by combining shape consistency constraint. And finally, further optimizing the detection result through geometric consistency constraint, and outputting a high-precision rotation target detection image after post-processing enhancement. The precision, the robustness and the physical consistency of SAR image rotation target detection under complex scenes are remarkably improved.

Inventors

LI XIAOYU
LI LU
ZHENG XIANGYUN
LIU YING
ZHAO ZUOPENG
XU SHANSHAN

Assignees

中国矿业大学
江苏比特达信息技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (9)

1. The SAR image speckle perception rotation target detection method based on physical priori is characterized by comprising the following steps: collecting SAR image data and related information of a target scene, wherein the SAR image data comprises SAR images under different radar frequency bands, resolution ratios and imaging modes; preprocessing the acquired SAR image, extracting target region characteristics through low-rank sparse decomposition, and inhibiting background noise; Combining physical priori information of the SAR image, and constructing a conditional diffusion model of the rotating frame by using the diffusion model; guiding a denoising process of the rotating frame through the physical prior information, and modeling the prediction of the rotating frame as a gradual denoising process of conditional generation; Introducing Gaussian-gas range regression loss, and combining shape consistency constraint to relieve regression instability caused by discontinuous angle period and inconsistent target shape; Optimizing the regression process of the rotating frame through geometric consistency constraint and shape consistency loss function, and ensuring the accurate positioning and shape stability of the rotating target; and storing a final detection result, carrying out post-processing on the image, enhancing image details by utilizing high-frequency information, removing residual noise, and outputting a high-precision SAR detection image.
2. The method for detecting a speckle perceived rotating target of an SAR image based on physical prior according to claim 1, wherein the method for preprocessing the acquired SAR image, extracting the target region features through low-rank sparse decomposition, and suppressing the background noise specifically comprises: decomposing an original SAR image into sparse components representing a target area and low-rank components representing background noise by adopting a low-rank sparse decomposition model, extracting target structural features and inhibiting the background noise, and then extracting a multi-scale feature map from the preprocessed image by using a convolutional neural network backbone network.
3. The method for detecting the speckle perceived rotating target of the SAR image based on the physical prior as set forth in claim 2, wherein the method for constructing the conditional diffusion model of the rotating frame by using the diffusion model in combination with the physical prior information of the SAR image specifically comprises the following steps: calculating equivalent noise layers and contrast of the target image and the background image as physical prior information; Constructing a physical priori condition vector comprising the logarithmically transformed input image, the equivalent noise layer and contrast of the target area, the equivalent noise layer and contrast of the background area and the target category semantic label; parameterizing the rotating frame into center point coordinates, width, height and rotation angle, and constructing the noisy rotating frame by adding gaussian noise to establish a diffusion process.
4. The method for detecting a speckle-aware rotating target of an SAR image based on physical prior according to claim 3, wherein the denoising process of the rotating frame is guided by the physical prior information, and the prediction of the rotating frame is modeled as a step-by-step denoising process of conditional generation, specifically comprising: Constructing a conditional diffusion type rotary decoding terminal, wherein the decoding terminal takes a noisy frame parameter of the current time step, a corresponding multi-scale feature map, a physical priori condition vector and a diffusion time step as inputs for predicting frame noise of the current step; in the reasoning stage, starting from an initial frame set containing random noise, gradually updating frame parameters through multi-step iterative denoising, so that the frame is converged from a random noise state to accurate prediction of the target position, shape and angle; and after each step of updating, resampling local condition information from the feature map and the physical prior vector according to the current frame position, and dynamically adjusting the subsequent guide signals.
5. The physical prior-based SAR image speckle perceived rotation target detection method as set forth in claim 4, wherein Gaussian-gas-range regression loss is introduced, and shape consistency constraint is combined to alleviate regression instability caused by angle cycle discontinuity and target shape inconsistency, and the method specifically comprises: Modeling the rotating frame as two-dimensional Gaussian distribution, adopting Gaussian-gas distance as main regression loss, and simultaneously optimizing the matching of a center point, a size and an angle by minimizing the mean value difference of Gaussian distribution corresponding to the prediction frame and the real frame; introducing aspect ratio smoothing loss, encouraging adjacent denoising step predicted frames to have smooth aspect ratio variation; and introducing physical priori shape loss based on the target category priori, and enabling the aspect ratio of the constraint prediction frame to conform to the reasonable range of the target category corresponding to the target category priori.
6. The method for detecting a speckle-aware rotating target of an SAR image based on physical prior according to claim 5, wherein the regression process of the rotating frame is optimized by geometric consistency constraint and shape consistency loss function, and the method specifically comprises the steps of: And adding geometric consistency regular term constraint in training, and forcing the predicted result after disturbance to be consistent with the average value frame of multiple predictions by applying multiple micro disturbance to the same input.
7. The physical prior-based SAR image speckle perceived rotating target detection method of claim 6, wherein the training process is optimized with a two-stage training strategy: The method comprises a first stage of extracting a backbone network and a denoising diffusion model by fixed features, and independently training a rotating frame solution wharf by using detection loss, and a second stage of carrying out end-to-end fine tuning by combining the whole processing flow, wherein a total loss function is obtained by weighted summation of denoising loss, detection loss and geometric consistency loss.
8. The method for detecting a speckle-aware rotating target of an SAR image based on physical prior according to claim 7, wherein the steps of storing the final detection result, performing post-processing on the image, enhancing image details and removing residual noise by using high-frequency information, and outputting a high-precision SAR detection image comprise: Screening the overlapping prediction frame by adopting a non-maximum suppression algorithm based on Gaussian-gas range distance, and filtering false positive detection by combining a confidence score and a physical consistency score of network prediction; Fusing high-frequency components of an original SAR image with the preprocessed smooth image, and retaining or enhancing edge and texture details of a target while inhibiting speckle; and applying guided filtering to non-target areas in the output image to further smooth background residual noise.
9. The SAR image speckle perceived rotating target detection method based on physical prior of claim 8, wherein, The output high-precision SAR detection image comprises two results, wherein one result is a vector detection file and comprises the position, the size, the direction, the category and the confidence information of each rotating target, and the other result is an enhanced SAR detection visual image, and the image is superimposed with an accurate rotating frame mark and enhanced through post-processing.

Description

SAR image speckle perception rotation target detection method based on physical prior Technical Field The invention relates to the technical field of SAR data processing and target detection, in particular to a SAR image speckle perceived rotation target detection method based on physical priori. Background Synthetic Aperture Radar (SAR) is widely used for remote sensing monitoring, environmental observation and disaster assessment due to its imaging capability over the course of the day and around the clock. However, the SAR image inevitably introduces speckle noise during imaging, and the quality of the image and the target detection accuracy are affected. In SAR images, speckle noise has a high degree of spatial and statistical coupling with target features, especially in complex scenarios with low textures and weak edges, which makes balancing noise suppression and target structure retention particularly difficult. Conventional denoising algorithms, while capable of partially suppressing noise, are not effective in dealing with the spatial variability and complex edge features of speckle noise in SAR images. Although the deep learning method can automatically learn the complex mapping of noise and signals, most methods rely on large-scale labeling data, and have the defects in the aspect of physical mechanism fusion. Compared with the traditional method, the denoising method based on deep learning can automatically learn complex mapping of noise and signals from data and extract multi-level features in the image, so that the denoising effect is improved to a certain extent. However, these deep learning methods generally rely on a large amount of labeling data, and the introduction of physical knowledge is insufficient when dealing with speckle noise and target features, lacking efficient fusion with SAR image physical mechanisms. This results in limitations in generalization ability and physical consistency of the deep learning method in real scenes. Furthermore, detection of rotational targets in SAR images has been a difficulty in the field of image processing, especially in terms of prediction of rotational bounding boxes. Conventional target detection methods are generally based on a general visual detection framework, and these methods fail to sufficiently consider physical characteristics of SAR imaging, and thus have a disadvantage in modeling of a rotating target. Although existing research has made some progress in specific scenarios, most approaches still rely on large scale labeling data and have poor modeling capabilities in terms of angular regression and shape consistency of rotating objects. Particularly under the conditions of complex background and strong interference, the existing method is difficult to maintain good detection precision and robustness. Therefore, how to combine physical priori information to effectively improve the performance of the model in SAR image denoising and rotation target detection is still a problem to be solved. Disclosure of Invention The invention aims to provide a SAR image speckle perception rotation target detection method based on physical priori, which solves the problems of low detection precision, poor robustness and insufficient physical consistency of a rotation target in a complex scene caused by lack of effective fusion with an SAR imaging physical mechanism in the prior art. In order to achieve the above purpose, the invention provides a SAR image speckle perception rotation target detection method based on physical priori, comprising the following steps: collecting SAR image data and related information of a target scene, wherein the SAR image data comprises SAR images under different radar frequency bands, resolution ratios and imaging modes; preprocessing the acquired SAR image, extracting target region characteristics through low-rank sparse decomposition, and inhibiting background noise; Combining physical priori information of the SAR image, and constructing a conditional diffusion model of the rotating frame by using the diffusion model; guiding a denoising process of the rotating frame through the physical prior information, and modeling the prediction of the rotating frame as a gradual denoising process of conditional generation; Introducing Gaussian-gas range regression loss, and combining shape consistency constraint to relieve regression instability caused by discontinuous angle period and inconsistent target shape; Optimizing the regression process of the rotating frame through geometric consistency constraint and shape consistency loss function, and ensuring the accurate positioning and shape stability of the rotating target; and storing a final detection result, carrying out post-processing on the image, enhancing image details by utilizing high-frequency information, removing residual noise, and outputting a high-precision SAR detection image. Preprocessing an acquired SAR image, extracting target region features through low-ra