Fourier contour syntax-guided remote sensing target detection method and device, and medium

Abstract

The invention relates to a Fourier contour syntax guided remote sensing target detection method, equipment and a medium, belongs to the technical field of remote sensing image target detection, and can effectively detect obviously symmetrical targets such as airplanes or ships in a remote sensing scene lacking contour / target mask labels. According to the invention, in a small sample scene, detection of a target category and positioning of a target can be realized, and a closed contour of the target can also be output. Fourier low-frequency spectrum is used to realize contour coding, and only 128-dimensional descriptors are used to replace traditional discrete contour point prediction. Through a triple Fourier guidance mechanism formed by cyclic alignment loss, low-frequency amplitude regularization and symmetric constraint, the noise interference of a weak supervision false contour is effectively relieved, the model is forced to preferentially learn the global shape and symmetric structure of a target, and the adaptability to scale change, any rotation angle and complex background interference is remarkably improved. And meanwhile, the detection accuracy in a small sample scene is improved, and practical value and theoretical value are both considered.

Description

Fourier contour syntax-guided remote sensing target detection methods, equipment, and media Technical Field

[0001] This invention relates to a remote sensing target detection method, device, and medium guided by Fourier contour syntax, belonging to the field of remote sensing image target detection technology. Background Technology

[0002] Aircraft and vessel detection in remote sensing images has significant applications in fields such as maritime surveillance, military monitoring, and military reconnaissance. However, due to complex imaging conditions, varying target scales, and significant background interference, their detection still faces many challenges. Especially in small-sample scenarios, achieving high-precision contour-level detection using limited labeled data remains a key research challenge. Currently, mainstream target detection methods can be divided into two categories:

[0003] I. Traditional Image Processing Methods: Early methods largely relied on hand-designed features and thresholding techniques, such as edge detection operators (e.g., Canny, Sobel), morphological operations, and template matching. While these methods can extract target contour information to some extent, their performance is heavily dependent on image quality and scene stability. In practical applications, targets in remote sensing images are often affected by the environment and imaging conditions, making traditional methods prone to false positives and false negatives. Furthermore, the design of hand-designed features requires strong domain knowledge and is difficult to generalize to complex and ever-changing real-world environments.

[0004] II. Deep Learning-Based Methods: With the development of deep learning, object detection methods based on convolutional neural networks (CNNs) (such as Faster R-CNN and the YOLO series) have made significant progress. These methods, through end-to-end learning, can automatically extract discriminative features and exhibit superior performance on large sample datasets. However, in scenarios with scarce labeled data and small sample sizes, deep models are prone to overfitting, leading to a decline in generalization ability. Furthermore, most methods focus on bounding box-level detection, making it difficult to achieve pixel-level contour reconstruction, while contour information is crucial for certain target model recognition and behavior analysis.

[0005] In recent years, some studies have attempted to improve the accuracy of contour detection by introducing attention mechanisms, multi-scale feature fusion, and shape prior knowledge. However, existing methods often neglect explicit modeling of the target contour during feature representation, particularly in balancing low-frequency contour structure with high-frequency details. Furthermore, how to utilize frequency domain features (such as Fourier descriptors) to model the global and local properties of the contour remains an area for further exploration.

[0006] Therefore, in current small-sample remote sensing target detection tasks, there is still a need for a contour-level detection method that can effectively integrate contour features, utilize frequency domain prior knowledge, and has strong generalization ability, so as to reconstruct the target contour under limited annotation conditions. Summary of the Invention

[0007] In view of this, the present invention provides a remote sensing target detection method, system, device and medium guided by Fourier contour syntax, which can effectively detect obviously symmetrical targets such as aircraft or ships in remote sensing scenarios where contour / target mask annotation is lacking, with high detection accuracy and the ability to reconstruct the target contour.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows.

[0009] A remote sensing target detection method guided by Fourier contour syntax includes:

[0010] Acquire a remote sensing image containing symmetrical targets, the remote sensing image having oriented target bounding boxes labeled;

[0011] The remote sensing images are input into a pre-trained backbone network and a feature pyramid network to extract multi-scale feature maps.

[0012] Based on the multi-scale feature map, a directional candidate box is generated through a directional region proposal network, and a fixed-resolution region of interest feature is obtained through a rotation and alignment operation.

[0013] Based on the real oriented bounding boxes in the training set, instance-level pseudo contours are generated through normalization correction, edge extraction, and contour resampling.

[0014] Based on the features of the region of interest, a lightweight fully connected network is used to predict the two-ended Fourier coefficients, which are then decoded into a fixed-length closed contour by inverse fast Fourier transform.

[0015] Cyclic alignment loss and low-frequency amplitude regularization loss are introduced to alleviate pseudo-contour noise interference and stabilize the contour learning process.

[0016] Based on the symmetry of the symmetric target, a symmetric constraint loss is designed to optimize the profile;

[0017] The directional detection loss is weighted and fused with the Fourier guided loss, which includes the cyclic alignment loss, low-frequency amplitude regularization loss and symmetry constraint loss. End-to-end joint training of the detection branch and the contour branch is achieved through gradient backpropagation.

[0018] During the inference phase, the above steps of feature extraction, candidate region generation, Fourier coefficient prediction and decoding are repeated on the test image to output the target's oriented bounding box, category information and reconstructed closed contour.

[0019] Specifically, the generation of instance-level pseudo-contours includes:

[0020] Calculate the corner coordinates of the true oriented bounding box, and correct the rotated region into an axis-aligned image patch through perspective transformation;

[0021] The corrected image patch is converted to grayscale and then smoothed using Gaussian filtering.

[0022] The Canny operator is used to extract the edge map, and the outer contour with the largest area is selected.

[0023] The outer contour is resampled into a fixed number of ordered points, and then mapped back to the original image coordinates through inverse perspective transformation to obtain a pseudo contour.

[0024] Specifically, the predicted two-ended Fourier coefficients are:

[0025] Based on the features of the region of interest, linear projection is used to predict the two-ended Fourier coefficients, preserving the positive and negative frequency components of the low-frequency spectrum to form a complex spectrum descriptor.

[0026] Specifically, the decoding of a fixed-length closed contour refers to:

[0027] The predicted Fourier coefficients are reconstructed into a complex spectrum, which is then decoded into an ordered sequence of contour points in a normalized space via inverse fast Fourier transform.

[0028] By inversely mapping the rotation matrix, scaling factor, and center offset, the contour points are transformed to the original image coordinates.

[0029] The cyclic alignment loss is calculated by cyclically shifting the contour point sequence to find the minimum error between the predicted contour and the target contour.

[0030] The low-frequency amplitude regularization loss is obtained by calculating the difference between the predicted contour and the target contour in the low-frequency spectrum amplitude in the Fourier domain.

[0031] The symmetric constraint loss is obtained by calculating the difference between the normalized contour obtained by decoding the predicted Fourier coefficients and the contour after y-axis reflection transformation.

[0032] The directional detection loss includes classification loss and directional bounding box regression loss.

[0033] The present invention also provides an electronic device, comprising: a processor; a memory for storing a program; wherein, when the program is executed by the processor, the processor causes the processor to execute a Fourier contour syntax-guided remote sensing target detection method according to the present invention.

[0034] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the Fourier contour syntax-guided remote sensing target detection method described in the present invention.

[0035] Beneficial effects

[0036] 1. This invention not only enables target category detection and target localization in small-sample scenarios, but also outputs the closed contour of the target. In this invention, Fourier low-frequency spectrum is used for contour encoding, replacing traditional discrete contour point prediction with only a 128-dimensional descriptor. Through a triple Fourier guidance mechanism consisting of cyclic alignment loss, low-frequency amplitude regularization, and symmetry constraints, noise interference from weakly supervised pseudo-contours is effectively mitigated, forcing the model to prioritize learning the global shape and symmetric structure of the target, significantly improving adaptability to scale changes, arbitrary rotation angles, and complex background interference. Simultaneously, it improves detection accuracy in small-sample scenarios, balancing practical and theoretical value.

[0037] 2. This invention significantly reduces annotation costs. In remote sensing scenarios, polygon or mask annotation of targets consumes a lot of manpower and time, and the annotation accuracy is difficult to guarantee under complex imaging conditions. However, this invention relies on the weakly supervised learning approach, which only needs to reuse existing directional bounding box annotations to generate pseudo-contour supervision signals, alleviating the problem of scarce contour annotations in remote sensing scenarios. The decoded contours are natural and smooth, perfectly adapting to the characteristics of scarce small sample data. Attached Figure Description

[0038] Figure 1 is a flowchart of the implementation of the method of the present invention.

[0039] Figure 2 is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to specific embodiments.

[0041] This invention provides a remote sensing target detection method guided by Fourier contour syntax, the specific process of which is shown in Figure 1, including the following steps:

[0042] Step 1: For remote sensing target detection guided by Fourier contour syntax, which includes symmetry syntax, remote sensing images containing symmetrical targets such as aircraft or ships are collected. The remote sensing images have oriented target boxes so that they can be used and trained during data preprocessing and training.

[0043] Among them, the remote sensing target detection guided by Fourier contour syntax requires the use of symmetric syntax. Therefore, remote sensing images containing symmetrical targets such as airplanes or ships are collected, and these remote sensing images are equipped with oriented target boxes so that they can be used and trained during data preprocessing and training.

[0044] Step 2, Multi-scale feature extraction: Input the remote sensing image into the pre-trained backbone network and feature pyramid network (FPN) to extract multi-scale feature maps;

[0045] In this embodiment, remote sensing images are input into a ResNet-50 pre-trained backbone network to extract low-level features containing basic information such as texture and edges. Then, the feature pyramid network (FPN) upsamples and laterally fuses the features at different levels to generate multi-scale feature maps. This can capture the detailed information of small targets as well as cover the global structure of large targets, providing comprehensive feature support for subsequent detection.

[0046] Step 3: Oriented candidate region generation: Oriented candidate boxes are generated through the Oriented RPN network, and after the Rotated RoIAlign operation, the region of interest (ROI) features with a fixed resolution are output.

[0047] Specifically, based on the multi-scale feature map, Oriented RPN generates anchor boxes with rotation angles through a sliding window mechanism, while predicting the foreground probability and directional bounding box regression offset of each anchor box to adapt to the detection requirements of arbitrary target poses. Then, rotation non-maximum suppression is used to filter overlapping candidate boxes, selecting high-confidence directional candidate regions. Finally, the Rotated RoIAlign operation is used to perform bilinear interpolation sampling on each candidate region, uniformly mapping candidate regions of different sizes and angles to 7×7 fixed resolution ROI features, effectively avoiding spatial quantization errors, ensuring feature space consistency, and providing stable and high-quality input features for subsequent Fourier contour branching and directional detection heads.

[0048] Step 4, Weakly Supervised Pseudo-Contour Generation: Based on the real oriented bounding boxes in the training set, instance-level pseudo-contours are generated through normalization correction, edge extraction, and contour resampling, serving as the supervision target for the Fourier contour branch; the specific implementation method is as follows:

[0049] First, for the ground truth bounding boxes in the training set, the coordinates of their four corner points are calculated. The corresponding rotated regions are warped into axis-aligned image patches through perspective transformation, and 8 pixels of padding are added to preserve complete edge information and eliminate the interference of rotation posture on subsequent processing. Next, the corrected image patches are converted into grayscale images, and noise is smoothed with a 3×3 Gaussian filter. Then, edge maps are extracted using the Canny operator (low threshold 60, high threshold 160) to capture the boundary information of the target. Subsequently, the outer contour with the largest area is selected from the edge map, and finally the contour is uniformly resampled into 128 points. After inverse perspective transformation, it is mapped back to the original image coordinates to obtain instance-level pseudo contours, which are used as the supervision target of the Fourier contour branch.

[0050] Step 5, Fourier contour branch training: Based on the ROI features, predict the two-ended Fourier coefficients through a lightweight fully connected network, and decode them into a fixed-length closed contour through inverse fast Fourier transform (IFFT).

[0051] In this embodiment, the shared fully connected layer of the Fourier contour branch multiplexing directional detection head constructs a lightweight network structure. Based on the linear projection of ROI features, it predicts the two-ended Fourier coefficients, retaining the positive and negative frequency components of the low-frequency spectrum (e.g., taking 64 frequency points from ±1 to ±32), ultimately obtaining a compact 128-dimensional descriptor (each frequency point contains a real and an imaginary part). Subsequently, the predicted Fourier coefficients are reconstructed into a complex spectrum of length 128, decoded by inverse fast Fourier transform (IFFT) into a closed contour in normalized space (unified into 128 ordered points), and then the contour points are transformed to the original image coordinates through the inverse mapping of rotation matrix, scaling factor, and center offset.

[0052] Step 6, Fourier top-down guidance: Introduce cyclic alignment loss and low-frequency amplitude regularization to alleviate pseudo-contour noise interference and stabilize the contour learning process;

[0053] Specifically, two types of losses guide the model to stably learn the contour. Cyclic alignment loss resolves ambiguity regarding the starting point of the closed contour. Cyclic shifting of the contour point sequence is used to find the minimum error between the predicted and target contours, avoiding interference from starting point differences in shape learning. The specific formula is as follows:

[0054]

[0055] Low-frequency amplitude regularization focuses on the low-frequency components of the Fourier transform (corresponding to the global shape), forcing the prediction to match the low-frequency spectrum amplitude of the target, guiding the model to prioritize learning the global structure of the target, and weakening the local noise of pseudo-contours. The specific formula is as follows:

[0056]

[0057] Where P represents the contour points and N represents the total number of contour points. This represents the cyclic shift operator, where 's' represents the number of shift steps and 'Z' represents the Fourier coefficients. The set of indices representing the Fourier coefficients.

[0058] Step 7, Symmetry Constraint: Utilize the symmetry of the target to be detected to impose symmetry constraints on the target and design a loss function to optimize the contour;

[0059] Specifically, a y-axis reflection operation is performed on the normalized contour obtained by inverse fast Fourier transform decoding. The similarity between the reflected contour and the original predicted contour is calculated using the Chamfer distance, and this similarity difference is converted into a symmetry constraint loss. The specific formula is as follows:

[0060]

[0061] Where CD is the Chamfer distance, and C is the normalized set of contour target points.

[0062] Since the weakly supervised pseudo-contour contains local noise, and the symmetric structure of the target to be detected is a stable prior, this loss can correct the asymmetric deviation of the contour, making the reconstruction result more consistent with the real structure of the target.

[0063] Step 8, Joint Optimization of Multiple Losses: The directional detection loss (classification + bounding box regression) and the Fourier guided loss are weighted and fused together, and the detection branch and contour branch are jointly trained end-to-end through gradient backpropagation;

[0064] Specifically, collaborative training is achieved through joint orientation detection and Fourier contour branching. The orientation detection loss comprises classification loss and orientation bounding box regression loss, responsible for supervising the accurate identification of target categories and the position of the orientation boxes. The Fourier guided loss integrates cyclic alignment loss, low-frequency amplitude regularization, and symmetry constraints to jointly constrain the shape rationality of the contour. After weighting and fusing the two types of losses according to preset weights, the parameters of the detection head and contour branch are simultaneously updated through gradient backpropagation, achieving end-to-end joint training. This allows detection accuracy and contour reconstruction quality to mutually constrain and synergistically improve in small-sample scenarios.

[0065] Step 9, Inference Stage Output: For the test image, after feature extraction, directional candidate region generation, Fourier coefficient prediction and IFFT decoding, the target's directional bounding box, category information and reconstructed closed contour are output.

[0066] Specifically, for the test image, the preceding steps of the training process are reproduced: after preprocessing, the data is input into the backbone network and FPN to extract multi-scale features. Oriented RPN generates oriented candidate boxes, which are then filtered by NMS rotation and RotatedRoIAlign to obtain ROI features at a fixed resolution. Subsequently, the trained Fourier contour branch is invoked to predict the bi-ended Fourier coefficients based on the ROI features. These coefficients are then decoded by IFFT into normalized closed contours and mapped back to the original image coordinates. Finally, the oriented bounding boxes, class labels, and reconstructed closed contours of the target are output simultaneously.

[0067] This application also provides an electronic device. Figure 2 shows the structure of the electronic device provided in this embodiment. For example, the electronic device 20 may include a processor 21, a memory 22, and a transmission device 23. The processor is used to execute a Fourier contour syntax-guided remote sensing target detection method mentioned in the above embodiments. The processor and memory can be connected via a bus or other means, taking a bus connection as an example. The transmission device can be connected to the processor and memory via wired or wireless means. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the Fourier contour syntax-guided remote sensing target detection method in this application embodiment. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing a Fourier contour syntax-guided remote sensing target detection method in the above method embodiments. The memory may include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The one or more modules stored in the memory, when executed by the processor, perform a Fourier contour syntax-guided remote sensing target detection method according to the embodiments.

[0068] In another aspect, this application also provides a computer-readable storage medium, which may be the computer-readable storage medium included in the apparatus described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into the device. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art. The computer-readable storage medium stores one or more programs, which are used by one or more processors to execute a Fourier contour syntax-guided remote sensing target detection method described in this application.

[0069] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing target detection method guided by Fourier contour syntax, characterized in that, include: Acquire a remote sensing image containing symmetrical targets, the remote sensing image having oriented target bounding boxes labeled; The remote sensing image is input into a pre-trained backbone network and a feature pyramid network to extract multi-scale feature maps. Based on the multi-scale feature maps, a directional candidate box is generated through a directional region proposal network, and a fixed-resolution region of interest feature is obtained through rotational alignment. Based on the real directional bounding boxes in the training set, instance-level pseudo contours are generated through normalization correction, edge extraction, and contour resampling. Based on the region of interest features, a lightweight fully connected network predicts the two-ended Fourier coefficients, which are then decoded into a fixed-length closed contour through inverse fast Fourier transform. Cyclic alignment loss and low-frequency amplitude regularization loss are introduced to alleviate pseudo contour noise interference and stabilize the contour learning process. Based on the symmetry of the symmetrical target, a symmetry constraint loss is designed to optimize the contour. The directional detection loss is weighted and fused with a Fourier guided loss that includes the cyclic alignment loss, low-frequency amplitude regularization loss, and symmetry constraint loss, and end-to-end joint training of the detection branch and the contour branch is achieved through gradient backpropagation. In the inference phase, the above feature extraction, candidate region generation, Fourier coefficient prediction, and decoding steps are repeated on the test image to output the target's directional bounding box, category information, and reconstructed closed contour.

2. The method according to claim 1, characterized in that, The specific steps for generating instance-level pseudo-contours include: calculating the corner coordinates of the real oriented bounding box, correcting the rotated region into an axis-aligned image block through perspective transformation; converting the corrected image block into a grayscale image and performing Gaussian filtering for smoothing; extracting the edge map using the Canny operator and selecting the outer contour with the largest area; resampling the outer contour into a fixed number of ordered points, mapping them back to the original image coordinates through inverse perspective transformation to obtain the pseudo-contour.

3. The method according to claim 1 or 2, characterized in that, The prediction of the two-ended Fourier coefficients specifically involves: predicting the two-ended Fourier coefficients by linear projection based on the features of the region of interest, retaining the positive and negative frequency components of the low-frequency spectrum, and forming a complex spectrum descriptor.

4. The method according to claim 3, characterized in that, The decoding of a fixed-length closed contour specifically involves: reconstructing the predicted Fourier coefficients into a complex spectrum, decoding them into an ordered sequence of contour points in a normalized space via inverse fast Fourier transform; and converting the contour points to the original image coordinates through the inverse mapping of rotation matrix, scaling factor, and center offset.

5. The method according to claim 4, characterized in that, The cyclic alignment loss calculates the minimum error between the predicted contour and the target contour by cyclically shifting the contour point sequence.

6. The method according to claim 4 or 5, characterized in that, The low-frequency amplitude regularization loss is obtained by calculating the difference between the predicted contour and the target contour in the low-frequency spectrum amplitude in the Fourier domain.

7. The method according to claim 6, characterized in that, The symmetry constraint loss is obtained by calculating the difference between the predicted contour and the contour after reflection transformation.

8. The method according to claim 6 or 7, characterized in that, The directional detection loss includes classification loss and directional bounding box regression loss.

9. An electronic device, characterized in that, include: processor; A memory for storing a program; wherein, when the program is executed by the processor, the processor performs a Fourier contour syntax-guided remote sensing target detection method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a Fourier contour syntax-guided remote sensing target detection method as described in any one of claims 1 to 8.

Images