Lightweight remote sensing satellite image segmentation method and system based on texture analysis module
By employing a lightweight remote sensing satellite image segmentation method based on a texture analysis module, and utilizing optimized neural network models and high-definition cameras, the problem of imbalance between computational complexity and texture feature utilization in remote sensing image segmentation on spaceborne platforms is solved, achieving efficient and reliable image segmentation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-10
- Publication Date
- 2026-03-31
AI Technical Summary
Existing remote sensing image segmentation methods have high computational complexity on resource-constrained spaceborne processing platforms, making it difficult to achieve efficient real-time processing. Furthermore, lightweight models that ignore local texture features lead to a decrease in the ability to distinguish texture-sensitive ground features, affecting segmentation accuracy and robustness.
A lightweight remote sensing satellite image segmentation method based on a texture analysis module is adopted. An optimized neural network model is used to extract image features. Image data is acquired by combining a high-definition camera, a satellite-borne GPS receiver, and a star sensor. Different types of texture features are separated through candidate frame recognition and texture analysis module to output accurate segmentation results. Security is ensured through an encrypted communication unit.
It reduces computational complexity, enhances the ability to recognize ground textures, improves image segmentation efficiency and accuracy, is suitable for resource-constrained spaceborne processing platforms, and ensures the real-time performance and security of segmentation.
Smart Images
Figure CN121767652A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing satellite image segmentation technology, and more specifically, it relates to a lightweight remote sensing satellite image segmentation method and system based on a texture analysis module. Background Technology
[0002] Remote sensing satellite imagery, as a crucial tool for acquiring information about the Earth's surface, is widely used in fields such as land resource surveys, urban planning, environmental monitoring, agricultural yield estimation, and disaster assessment. With the rapid development of satellite remote sensing technology, the spatial and spectral resolution of remote sensing images has continuously improved, resulting in data exhibiting characteristics of massive data volume, high dimensionality, and complexity. Against this backdrop, achieving accurate identification and segmentation of various ground features in remote sensing images has become critical. Remote sensing image segmentation aims to perform semantic classification on each pixel and precisely extract ground features such as buildings, roads, water bodies, and vegetation. Traditional methods rely on manually designed features (such as gray-level co-occurrence matrices and Gabor texture features) combined with shallow classifiers (such as support vector machines), but their feature representation capabilities are limited, making it difficult to cope with the challenges of diverse ground features and large scale variations, resulting in limited segmentation accuracy and generalization ability. In recent years, deep learning technologies such as convolutional neural networks (CNNs) have significantly improved image segmentation performance. Models such as U-Net and DeepLab achieve powerful feature extraction and context modeling capabilities through encoder-decoder structures. However, these mainstream models are complex in structure, have a large number of parameters, and high computational overhead, making them difficult to deploy in resource-constrained scenarios such as spaceborne processing platforms, thus limiting their real-time processing capabilities.
[0003] Furthermore, different land features in remote sensing images exhibit significant texture differences, such as the striped texture of farmland and the geometric texture of urban building complexes. Existing lightweight models often neglect explicit modeling of local texture features when reducing computational costs, leading to a decrease in the ability to distinguish texture-sensitive land features and affecting overall accuracy and robustness. Given the trend towards intelligent and real-time remote sensing technology, existing methods struggle to balance model lightweighting with the utilization of texture features. Therefore, designing a lightweight remote sensing satellite image segmentation method and system based on a texture analysis module is crucial for improving automation and intelligence levels, enhancing the deployment capabilities of edge devices, and ensuring segmentation accuracy and efficiency. Summary of the Invention
[0004] The technical problem to be solved by this invention is to provide a lightweight remote sensing satellite image segmentation method and system based on a texture analysis module. It utilizes an optimized neural network model for efficient image feature extraction, reduces computational complexity, enhances the ability to recognize ground textures, and provides a more efficient and reliable solution for remote sensing image segmentation.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a lightweight remote sensing satellite image segmentation method based on a texture analysis module, comprising the following steps:
[0006] Step 1: Use a high-definition camera to capture optical images of the target area according to a predetermined trajectory and attitude, and record the parameters in the onboard GPS receiver and star sensor corresponding to the photos in real time;
[0007] Step 2: Use a candidate frame recognizer to sample the optical image of the target area captured by the high-definition camera. Use a random sampling algorithm to select key frames that appear in the optical image and output candidate frame images.
[0008] Step 3: Crop the candidate frame images according to the image specifications, input the candidate frame image blocks into the remote sensing image analysis model, use the backbone network to extract multi-scale image features, use the texture analysis module to separate different types of texture features, and output the accurate segmentation results of the image.
[0009] Step 4: Post-process the segmentation results, stitch together the segmentation results corresponding to each image block according to the cropping order, and output the complete segmentation result;
[0010] Step 5: Compare the segmentation results with the standard results to adjust the parameters of the lightweight convolutional Transformer network;
[0011] Step Six: Freeze the parameters of the lightweight convolutional Transformer network model immediately after training;
[0012] Step 7: After loading the model parameters, use the lightweight convolutional Transformer network model to analyze the continuous frames captured by the high-definition camera, and repeat steps 2 to 4.
[0013] A lightweight remote sensing satellite image segmentation system based on a texture analysis module includes an image acquisition unit, an image processing unit, and a result feedback unit. The image acquisition unit is connected to the input end of the image processing unit, and the result feedback unit is connected to the output end of the image processing unit.
[0014] The image acquisition unit includes a high-definition camera, a satellite-borne GPS receiver, and a star sensor. The high-definition camera is configured to acquire a series of candidate frames captured by the satellite as it scans the target area according to a pre-planned sequence. The satellite-borne GPS receiver and the star sensor are used to acquire the satellite's orbital and attitude information.
[0015] The image processing unit includes a candidate frame recognizer, a memory, and a remote sensing image analyzer. The candidate frame recognizer is configured to select candidate frames from a series of images acquired by the image acquisition unit and randomly sample the selected candidate frames. The memory is configured to store candidate frame images and high-resolution images captured in the satellite operation plan. The remote sensing image analyzer is configured to receive candidate frame images from the memory, extract cropped image patches using a backbone network, separate different types of texture features using a texture analysis module, classify different texture features using a terrain segmentation module, and output pixel-level terrain segmentation results. The segmentation results are stored in the memory.
[0016] The result feedback unit includes a post-processor. The result feedback unit is configured to reselect the sampling target and execute the segmentation process when the difference between the candidate frames output by the terrain segmentation module exceeds a set range.
[0017] Preferably, the candidate frame recognizer selects candidate frames indiscriminately, using equidistant selection to select frames in a fixed sequence, with the equidistant interval set manually, and the interval being 2 frames.
[0018] Preferably, the candidate frame recognizer selects candidate frames according to actual needs, and sets the candidate frame set with a small number of frames to select all.
[0019] Preferably, the satellite orbital and attitude parameters stored in the memory are automatically stored after the remote sensing image analyzer finishes reading them, and a log is retained.
[0020] Preferably, the segmentation system further includes an encrypted communication unit connected to the output of the result feedback unit. The encrypted communication unit is used to encrypt key segmentation results, and the encryption algorithm used has satellite orbital parameters and attitude information as the main parameters.
[0021] The beneficial effects of adopting the above technical solution are as follows: This invention utilizes an optimized neural network model for efficient image feature extraction, which not only reduces computational complexity but also enhances the ability to recognize ground textures, improves image segmentation efficiency and accuracy, and provides a more efficient and reliable solution for remote sensing image segmentation. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the architecture of the partitioning system.
[0023] Figure 2 This is a schematic diagram of the workflow of the segmentation system.
[0024] Figure 3 This is a schematic diagram of the backbone network.
[0025] Figure 4This is a schematic diagram of the texture analysis module.
[0026] Figure 5 This is a flowchart of the segmentation method.
[0027] In the diagram: 1. Image acquisition unit, 11. Onboard GPS receiver, 12. Star sensor, 13. High-definition camera; 2. Image processing unit, 21. Candidate frame recognizer, 22. Memory, 23. Remote sensing image analyzer; 3. Result feedback unit, 31. Post-processor; 4. Encrypted communication unit. Detailed Implementation
[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0029] like Figure 1 As shown, the segmentation system includes an image acquisition unit 1, an image processing unit 2, and a result feedback unit 3. The image acquisition unit 1 is connected to the input of the image processing unit 2, and the result feedback unit 3 is connected to the output of the image processing unit 2. The image acquisition unit 1 includes an onboard GPS receiver 11, a star sensor 12, and a high-definition camera 13, which is configured to acquire high-definition images of the target area. The image processing unit 2 includes a candidate frame recognizer 21, a memory 22, and a remote sensing image analyzer 23. The candidate frame recognizer 21 is configured with a random sampling model, which randomly samples continuously captured image frames using a fixed sampling method and outputs the results to the memory 22 for storage. The memory 22 is configured to store candidate frames acquired by the high-definition camera 13, onboard GPS parameters, and star sensor attitude parameters. The remote sensing image analyzer 23 is configured with a multi-scale feature extraction module, a texture analysis module, and a terrain segmentation module. After reasonably cropping the candidate frames in the memory 22, the remote sensing image analyzer 23 performs accurate segmentation on all cropped images and finally outputs the segmentation results. The result feedback unit 3 includes a post-processor 31. When there are significant differences between the segmentation results of candidate frames, the post-processor 31 chooses to restart the sampling process or weight the segmentation results of each candidate frame to obtain the final output result. In specific implementation, the image acquisition unit 1, image processing unit 2, and result feedback unit 3 form an integrated control system, such as an AI-assisted recognition system, which can realize automated control systems for image recognition, judgment, learning, and understanding.
[0030] Specifically, the high-definition camera 13 is located at the tail of the satellite, while the onboard GPS receiver 11 and star sensor 12 are located in the satellite's internal electronics bay. The high-definition camera 13 is used to acquire high-definition images of the target area, and the onboard GPS receiver 11 and star sensor 12 are used to acquire the satellite's orbital and attitude information. The onboard GPS receiver 11 and star sensor 12 can be replaced by parameters pre-planned in the design.
[0031] The candidate frame recognizer 21 can also be configured as a random sampling algorithm to sample candidate frames from the high-definition image frames acquired by the image acquisition unit 1. The memory 22 is configured to store candidate frame images and satellite orbit and attitude information. The remote sensing image analyzer 23 is configured to receive the images from the memory 22, crop the high-definition images to a fixed size, such as 320*256, extract multi-scale feature information using the ConvTrans backbone network, separate different types of texture features through the texture analysis module, and finally perform pixel-level segmentation of the features by the terrain segmentation module. In this embodiment, accurate segmentation of the high-definition image by the remote sensing image analyzer 23 can obtain more accurate terrain information, eliminate invalid information in the image, reduce communication transmission costs, and ensure the stability of candidate frame segmentation through post-processing 31.
[0032] The image segmentation system also includes an encrypted communication unit 4. The encrypted communication unit 4 is connected to the output of the result feedback unit 3. To ensure the security of the processing results, the encrypted transmission unit 4 uses orbit information and attitude parameters as private key noise to encrypt the segmented image results, thus further ensuring the security of the accurate segmentation results. In this embodiment, the result feedback unit 3 is connected to the encrypted communication unit 4. When the result feedback unit 3 determines that the segmentation results of the candidate frames meet the requirements, the encrypted communication unit 4 compresses the stitched image segmentation results and uses satellite orbit information and attitude information as keys to encrypt the segmentation results. In specific implementations, the encrypted communication unit 4 only needs to use predetermined orbit information and attitude information for encryption, and uses orbit parameters and attitude parameters as encryption noise to encrypt the segmentation results. In the lightweight remote sensing image segmentation network, a texture analysis module is used. This multi-dimensional module is responsible for separating different types of textures in the network structure, which is an important guarantee for improving the model's segmentation accuracy.
[0033] The candidate frame recognizer 21 employs a random sampling algorithm, taking candidate image frames as input and outputting a sampled and simplified image. The remote sensing image analyzer 23 consists of a backbone network ConvTrans, a texture analysis module, and a terrain segmentation module. The number of candidate frames is determined by the sampling parameters, and candidate frames only need to be extracted during the high-definition camera's capture process; one image is extracted from each video as a keyframe. The candidate frames are input into the backbone network ConvTrans, which passes the main features to the texture analysis module.
[0034] like Figure 3 As shown, the design of the backbone network ConvTrans in the remote sensing image analyzer is as follows. Given a feature map... After cutting and channel separation,
[0035] (1)
[0036] Out_1, (2)
[0037] (3)
[0038] After Avgpooling, conv, GeLU, and Sigmoid, local feature information can be obtained. To obtain global feature information, an improved lightweight transformer module is used, where multi_head is set to 4.
[0039] (4)
[0040] (5)
[0041] (6)
[0042] (7)
[0043] like Figure 4 As shown, the most important texture analysis module is designed as follows: a multi-channel convolution is designed to separate texture features, which consist of fine-grained features obtained from convolution and coarse-grained features obtained from Transformer.
[0044] (8)
[0045] (9)
[0046] (10)
[0047] (11)
[0048] (12)
[0049] (13)
[0050] Fine-grained and coarse-grained features are mapped through different convolution channels. Finally, they are summed together, facilitating feature separation in the upsampling section (terrain segmentation module). After training, texture feature separation becomes stable, which helps the model grasp the overall image and improves the stability of the remote sensing image analyzer.
[0051] like Figure 5 As shown, the segmentation method includes the following steps:
[0052] S1. Image and key parameter acquisition: Image acquisition unit 1 acquires orbital parameters from onboard GPS receiver 11, satellite attitude parameters from star sensor 12, and target area images from high-definition camera 13;
[0053] S2. Candidate frame selection: The candidate frame recognizer 21 of the image processing unit 2 takes the target area photographic image input by the image acquisition unit 1 as input, removes the photos taken to ensure the robustness of the results and irrelevant photos taken after leaving the target area, and outputs the candidate frame image using the candidate frames filtered by the random sampling algorithm.
[0054] S3. Candidate Frame Storage: Candidate frames containing target area information sampled by the image processing unit 21 are stored in the memory 22 as input to the remote sensing image analyzer 23. The memory 22 also contains orbital and attitude parameters from the time of capture. Specifically, the memory 22 writes the images output by the candidate frame recognizer 21 into the memory in the order they were captured by the high-definition camera in the image acquisition unit, and then transmits them sequentially to the remote sensing image analyzer 23 after writing.
[0055] S4. Remote sensing image segmentation: The candidate frame images in the memory are input into the remote sensing image analyzer 23. Through the cropping step in the remote sensing image analyzer 23, the high-resolution images are cropped into low-resolution images that meet the batch processing requirements using the satellite orbit parameters and attitude parameters in the memory 22, and then extracted. Multi-scale image features are extracted through the backbone network, and different types of textures are separated through the texture analysis module. Finally, the terrain segmentation module is used to obtain pixel-level accurate segmentation results.
[0056] S5. Post-processing of segmentation results: The segmentation results are spliced together according to the cropping order. If the differences meet the standards, smoothing will be performed to clean up outliers and noise points, and the final segmentation result will be obtained.
[0057] S6. If there are significant differences between the precise segmentation results obtained by the remote sensing image analyzer for the candidate frames, return to step S2 and obtain an image of the appropriate target region by resampling.
[0058] S7. Encrypted Transmission: After obtaining stable candidate frame image segmentation results, a specific encryption tool is used to stably encrypt the segmentation results. An encryption operator is used to add noise to the pixel values of the segmentation results before transmitting them to the satellite's communication module.
[0059] In this embodiment, the security of remote sensing satellite image segmentation task can be improved by using the encrypted communication unit 4. Since the size of the segmentation tag is smaller than the original image, and the encryption step does not bring additional communication costs, the efficiency of image segmentation task completion can be guaranteed.
[0060] Compared with existing technologies, the lightweight remote sensing satellite image segmentation method and system based on a texture analysis module provided by this invention, during the process of the remote sensing satellite passing over a predetermined orbit and target area, involves the image acquisition unit's high-definition camera acquiring high-definition image frames of the target area, the image processing unit receiving the sampled image frames, and using a remote sensing image analyzer to obtain the terrain segmentation results of the image frames; when the difference between the segmentation results of candidate frames exceeds a threshold, the result feedback unit prompts the sampling algorithm to be updated, initiating a new round of image segmentation. This achieves stable segmentation of the target area image, and a post-processor is used to regularize the segmented fragments to form a complete segmentation result. Furthermore, an encrypted communicator is used to encrypt the segmentation result, improving the communication efficiency and security level of the remote sensing satellite.
[0061] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A lightweight remote sensing satellite image segmentation method based on a texture analysis module, characterized by, The method comprises the following steps: Step 1: using a high-definition camera to shoot optical images of a target area along a predetermined track and pose, and recording the parameters of a satellite-borne GPS receiver and a star sensor corresponding to the photos in real time; Step 2: using a candidate frame identifier to sample the optical images of the target area shot by the high-definition camera, using a random sampling algorithm to select key frames appearing in the optical images, and outputting candidate frame images; Step 3: cropping the candidate frame images according to image specifications, inputting the candidate frame image blocks into a remote sensing image analysis model, using a backbone network to extract multi-scale image features, using a texture analysis module to separate different types of texture features, and outputting accurate segmentation results of the images; Step 4: post-processing the segmentation results, splicing the segmentation results corresponding to each image block in the order of cropping, and outputting complete segmentation results; Step 5: comparing the segmentation results with standard results to adjust parameters of a lightweight convolutional Transformer network; Step 6: immediately freezing the parameters of the lightweight convolutional Transformer network model after training; Step 7: loading the model parameters, using the lightweight convolutional Transformer network model to analyze continuous frames shot by the high-definition camera, and repeating steps 2 to 4.
2. A lightweight remote sensing satellite image segmentation system based on texture analysis module for implementing the method as claimed in claim 1, characterized in that, The system comprises an image acquisition unit, an image processing unit and a result feedback unit, the image acquisition unit is connected with the input end of the image processing unit, and the result feedback unit is connected with the output end of the image processing unit.
3. The lightweight remote sensing satellite image segmentation system based on texture analysis module according to claim 2, characterized in that, The image acquisition unit comprises a high-definition camera, a satellite-borne GPS receiver and a star sensor, the high-definition camera is configured to acquire a series of candidate frames shot by the satellite according to a predetermined plan when the satellite sweeps through a target area; the satellite-borne GPS receiver and the star sensor are configured to acquire orbit information and attitude information of the satellite; The image processing unit comprises a candidate frame identifier, a storage and a remote sensing image analyzer, the candidate frame identifier is configured to select candidate frames from a series of images acquired by the image acquisition unit, and randomly sample the selected candidate frames, the storage is configured to store candidate frame images and high-definition images shot according to a satellite operation plan, and the remote sensing image analyzer is configured to receive the candidate frame images in the storage, extract cropped image blocks using a backbone network, separate different types of texture features using a texture analysis module, classify the different texture features through a terrain segmentation module, and output pixel-level terrain segmentation results, the segmentation results are stored in the storage; The result feedback unit comprises a post-processor, the result feedback unit is configured to reselect a sampling target and execute a segmentation process when the difference between the candidate frames output by the terrain segmentation module exceeds a set range.
4. The texture analysis module based lightweight remote sensing satellite image segmentation system according to claim 2, wherein, The candidate frame identifier randomly selects candidate frames using equidistant selection, selects frames in a fixed sequence, and manually sets an equidistant interval of 2 frames. 5.The lightweight remote sensing satellite image segmentation system based on texture analysis module of claim 2, wherein, The candidate frame identifier selects candidate frames according to actual needs, and sets a candidate frame set with a small number of frames to be selected completely. 6.The lightweight remote sensing satellite image segmentation system based on texture analysis module according to claim 2, wherein, The satellite orbit parameters and attitude parameters stored in the storage are automatically stored after being read by the remote sensing image analyzer, and logs are kept.
7. The lightweight remote sensing satellite image segmentation system based on texture analysis module according to claim 2, wherein, The segmentation system further comprises an encryption communication unit connected with the output end of the result feedback unit, the encryption communication unit being used for encrypting the key segmentation result, and the encryption algorithm used taking the orbit parameters and attitude information of the satellite as main parameters.
8. The lightweight remote sensing satellite image segmentation system based on texture analysis module according to claim 3, characterized in that, The candidate frame identifier adopts a random sampling algorithm, inputs a candidate image frame, and outputs a sampled and simplified picture; the remote sensing image analyzer is composed of a backbone network ConvTrans, a texture analysis module, and a terrain segmentation module; wherein the number of candidate frames is determined by the sampling parameters, and the candidate frames only need to be extracted during the shooting process of the high-definition camera, and one image is extracted from each video as a key frame; the candidate frame is input into the backbone network ConvTrans to pass the main features to the texture analysis module.
9. The lightweight remote sensing satellite image segmentation system based on texture analysis module according to claim 8, characterized in that, The design of the backbone network ConvTrans in the remote sensing image analyzer is as follows: given a feature map , after cropping and channel separation, , (1) Out_1, (2) （3） After Avgpooling, conv, GeLU, and Sigmoid, local feature information is obtained; in order to obtain global feature information, an improved lightweight transformer module is used, wherein the multi_head is set to 4; （4） （5） （6） （7）。 10. The lightweight remote sensing satellite image segmentation system based on texture analysis module according to claim 9, characterized in that, The texture analysis module is designed as follows: a multi-channel convolution is designed to separate the texture features, and the fine-grained features obtained by convolution and the coarse-grained features obtained by the transformer are combined; （8） （9） （10） （11） （12） （13） The fine-grained features and the coarse-grained features are mapped through different convolution channels; finally, they are added together, which is easy for the up-sampling part to separate the features; after training, the texture feature separation tends to be stable, which is conducive to the model's grasp of the overall image and improves the stability of the remote sensing image analyzer.