Remote sensing image slice library image classification method, model and image sampling method

Abstract

The invention belongs to the technical field of image data processing, and discloses a remote sensing image slice library image classification method, a remote sensing image slice library image classification model and an image sampling method. Comprising the following steps: separating an image in an image slice library into a background image and a target image by adopting a target and background separation strategy; extracting background features from the background image, extracting target features from the target image, and extracting overall features from the background image and the target image by adopting a contrast regularization processing method; and classifying the images according to the background features, the target features and the overall features. The problems of low image classification speed and low accuracy of the remote sensing image slice library are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of image data processing, and in particular relates to an image classification method, a model and an image sampling method for a remote sensing image slice library. Background Art

[0002] The massive growth of remote sensing image data has brought new challenges to the relevant departments in building efficient and intelligent remote sensing analysis platforms. In order to meet these challenges, the storage, rapid processing and publishing capabilities of data are keeping pace with the times. Among them, the fast and effective sampling technology of slice library images has become a key means to improve the efficiency of remote sensing image data analysis.

[0003] The existing rapid sampling technology of remote sensing image slice library images performs well in classification speed and sample extraction accuracy, especially when processing large-scale remote sensing image data. Compared with traditional manual classification and rule-based sampling methods, it can significantly improve processing efficiency and reduce human errors. However, it still faces some challenges in practical applications and has the following disadvantages: 1) High computational cost and long processing time. The complexity of image classification and representative sample output will lead to significant computational requirements, especially when processing extremely large amounts of data, requiring a large amount of high-performance computing resources, thereby increasing computational costs and processing time.

[0004] 2) Data preprocessing is complex. Data preprocessing is crucial in the rapid sampling of remote sensing image slice libraries. Currently, it mainly includes radiation correction, geometric correction, image enhancement and denoising, etc. However, these data preprocessing technologies are complex and require precise operations and professional skills.

[0005] 3) Classification accuracy is greatly affected by the training data set. Supervised classification techniques such as support vector machines (SVM), maximum likelihood methods, and decision trees all rely on high-quality training data. If the training data is insufficient or inaccurate, the classification results will be inaccurate, affecting the representativeness of the entire sample and the accuracy of the sampling. In addition, it is also difficult to conduct quantitative statistics for each classification. Summary of the invention

[0006] The purpose of the present invention is to provide a remote sensing image slice library image classification method, model and image sampling method to solve the problems of slow speed and low accuracy of remote sensing image slice library image classification.

[0007] The present invention is achieved through the following technical solutions: A remote sensing image slice library image classification method, using a multi-stage hierarchical adaptive partitioning network model to classify images in the image slice library; The following steps are involved: S1, using the target and background separation strategy to separate the images in the image slice library into background images and target images; S2, using contrast regularization processing method, extracting background features from the background image, extracting target features from the target image, and extracting overall features from the background image and the target image; S3. Classify the image according to background features, target features and overall features.

[0008] In some embodiments, step S1 includes: Convert images to grayscale and blur them; The blurred image is processed by Gaussian blur to enhance the contrast between the background and the target in the blurred image; According to the grayscale image and the blurred image after Gaussian blur processing, the image is separated into a background image and a target image through a mask.

[0009] In some embodiments, in the contrast regularization processing method, the extracted target features and overall features are used to generate positive sample pairs, the extracted background features and target features are used to generate negative sample pairs, and the feature distance of the positive sample pairs is shortened, while the feature distance of the negative sample pairs is increased.

[0010] In some embodiments, it also includes using a self-guided network method to enhance the background features, target features and overall features, and automatically identifying and emphasizing the commonalities of the background features, target features and overall features between images of the same type.

[0011] In some embodiments, in the contrast regularization processing method, the loss function of the multi-stage hierarchical adaptive partitioning network model is: , Where N is the number of input samples; C is the number of categories; Represents the true label of sample i. If the sample belongs to the cth category, then =1, otherwise 0; is the model's predicted probability that sample i belongs to class c; β is a weight coefficient used to balance the cross entropy loss and the contrast regularization term; is the contrast regularization term, used to calculate the background feature F b , target feature F t and the overall characteristic F m The relationship between.

[0012] The present invention also relates to a multi-stage hierarchical adaptive partitioning network model, which is used for classifying images in the above-mentioned remote sensing image slice library image classification method, comprising: A separation processing module is used to separate the images in the image slice library into background images and target images through a target and background separation strategy; A feature extraction module, used to extract background features in the background image, extract target features in the target image, and extract overall features in the background image and the target image through a contrast regularization processing method, including a background branch, a target branch, and a main branch, wherein the background branch is used to extract background features in the background image, the target branch is used to extract target features in the target image, and the main branch is used to extract overall features in the background image and the target image; The contrast classification module is used to classify images according to background features, target features and overall features.

[0013] In some embodiments, a common feature recognition module is also included, which is used to enhance the background features, target features and overall features through a self-guided network method, and automatically identify and emphasize the commonalities of background features, target features and overall features between images of the same type.

[0014] The present invention also relates to a remote sensing image slice library image sampling method, which samples images in the image slice library, comprising the steps of: Using one of the above remote sensing image slice library image classification methods to classify the images; The image file of the image is stored in the local file system, the metadata information of the image is stored in the relational database, and a mapping between the image file and the metadata information is established; The images are retrieved and sampled according to the image files in the local file system, the metadata information in the relational database and the total sampling quantity to obtain a representative sample.

[0015] In some embodiments, the process further includes the step of correcting image classification errors in a database based on user feedback information, wherein the user feedback information is correction information fed back by the user when the user finds that a representative sample in the image sampling has classification errors.

[0016] In some embodiments, the step of iteratively updating the image according to the error correction information is further included, including: Incremental training is performed on the multi-stage hierarchical adaptive partitioning network model by local updating based on the error correction information; After completing the incremental training, the multi-stage hierarchical adaptive partitioning network model reclassifies the images in the entire image slice library.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. The present invention adopts a target and background separation strategy and a contrast regularization processing method through a multi-stage hierarchical adaptive partitioning network model to perform effective and accurate image classification, which significantly improves the speed and accuracy of image classification in the remote sensing image slice library.

[0018] 2. The image files of the images are stored in the local file system, and the metadata information of the images is stored in the relational database in a separate storage method, which realizes fast calling and efficient storage, makes query and management more flexible, reduces the storage pressure of large-scale data, improves the retrieval efficiency of sampling, and ensures the efficiency and scalability of processing large-scale data during the sampling process.

[0019] 3. Image classification error correction and iterative update are performed based on user feedback information. When image classification errors are found during image sampling, the classification and metadata information of the image in the database can be automatically adjusted. The multi-stage hierarchical adaptive partitioning network model is incrementally trained through error correction information to continuously optimize the classification performance of the model, improve the classification speed and accuracy of the model, thereby improving the accuracy of image sampling and achieving continuous improvement of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 This is a flow chart of the image sampling method for the existing remote sensing image slice library.

[0022] Figure 2 This is a flow chart of the image classification method for the remote sensing image slice library in Example 1 of the present invention.

[0023] Figure 3 This is a flow chart of the target and background separation strategy in Example 1 of the present invention.

[0024] Figure 4 This is a flow chart of the image sampling method for the remote sensing image slice library in Example 3 of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0026] The existing sampling technology solutions for remote sensing image slice libraries mainly focus on improving processing efficiency and sample accuracy. The existing technology starts with image preprocessing, such as correction and normalization, to ensure data quality; then, automatic classification and annotation technologies, such as machine learning and supervised classification, are used to identify and classify surface features; in terms of sampling strategies, random sampling and stratified sampling are used according to image features to ensure the representativeness of the samples; in addition, feature extraction technology enhances the understanding of image content and optimizes sampling results; data compression technology is also used to process large-scale data and speed up analysis; sampling methods are continuously optimized through feedback from actual applications. This solution has the potential to be widely used in fields such as land monitoring and urban planning, and is expected to improve the application efficiency and accuracy of remote sensing image data.

[0027] Reference Figure 1 , Flowchart of the existing rapid sampling method for remote sensing image slice libraries,The existing rapid sampling technology for remote sensing image slice libraries mainly includes two key steps: image classification and output of representative samples. However, the existing technology has certain difficulties in quantitative statistics of each classification. The sampling process based on these two steps can be briefly described as: 1) Sample library construction: Remote sensing image data is obtained from different satellites, drones or other remote sensing equipment. Remote sensing image data is the basis for analysis and processing, and contains important visual information of the surface.

[0028] 2) Data preprocessing. Perform radiation and atmospheric correction to ensure the spectral authenticity of remote sensing images; perform geometric correction through geocoding and orthorectification techniques to correct image deformation and position errors and improve geographic accuracy; at the same time, apply image enhancement and denoising techniques to improve image quality and enhance feature recognition.

[0029] 3) Image classification: By using supervised classification techniques such as support vector machines (SVM), maximum likelihood and decision trees, as well as machine learning methods, surface features in remote sensing image data can be efficiently classified.

[0030] 4) Representative sample output: After the image classification is completed, feature extraction is used to optimize the classification results and automatically select representative samples from each category.

[0031] Example 1 A remote sensing image slice library image classification method, referring to Figure 2 , a multi-stage hierarchical adaptive partitioning network model is used to classify images in the image slice library; the following steps are included: S10, separating the image in the image slice library into a background image and a target image by using a target and background separation strategy; Reference Figure 3, convert the image to grayscale image and blur the image.

[0032] Grayscale images preserve texture and morphological features and help reduce the interference of color and lighting.

[0033] The blurred image is processed by Gaussian blur to enhance the contrast between the background and the target in the blurred image, reduce the noise and unnecessary details in the image, and thus enhance the distinction between the target and the background.

[0034] According to the grayscale image and the blurred image after Gaussian blur processing, the image is separated into a background image and a target image through a mask. The mask is used to accurately separate the target and background areas, the target image is used to extract key target features, and the background image is used to capture auxiliary information in the scene.

[0035] S11, using a contrast regularization processing method to extract background features from the background image, extract target features from the target image, and extract overall features from the background image and the target image; S12. Classify the image according to background features, target features and overall features.

[0036] Through the multi-stage hierarchical adaptive partitioning network model, the target and background separation strategy and contrast regularization processing method are adopted to perform effective and accurate image classification, which significantly improves the speed and accuracy of image classification in the remote sensing image slice library.

[0037] Example 2 The present invention also relates to a multi-stage hierarchical adaptive partitioning network model, which is used for classifying images in the above-mentioned remote sensing image slice library image classification method, comprising: 1) Separation processing module, used to separate the images in the image slice library into background images and target images through the target and background separation strategy.

[0038] In the target and background separation strategy, the background branch is used to analyze the background information in the image, the target branch is used to analyze the most distinguishing target features in the image, and the main branch is used to comprehensively analyze the image and integrate the overall features.

[0039] The multi-branch structure of the multi-stage hierarchical adaptive partitioning network model MLAPNet can effectively process scenes that are visually similar but of different categories, significantly reducing misclassification and improving classification accuracy.

[0040] 2) A feature extraction module, which is used to extract background features in the background image, extract target features in the target image, and extract overall features in the background image and the target image through a contrast regularization processing method, including a background branch, a target branch and a main branch. The background branch is used to extract background features in the background image, the target branch is used to extract target features in the target image, and the main branch is used to extract overall features in the background image and the target image.

[0041] The multi-stage hierarchical adaptive partitioning network model adopts a multi-branch structure to independently process the information in the target image and the background image. By utilizing the image background information to improve the accuracy of target recognition, it can better cope with the complexity and variability of remote sensing data, ensure the stability of the multi-stage hierarchical adaptive partitioning network model, and significantly improve the classification accuracy.

[0042] By contrasting regularization methods, the differences between different classes are strengthened and the consistency within the same class is enhanced in multiple branches of the multi-stage hierarchical adaptive partitioning network model, which promotes the learning of more robust and discriminative feature expressions. At the same time, each branch can analyze the input remote sensing data from different perspectives, thereby enriching the overall feature expression.

[0043] In the contrastive regularization processing method, the extracted target features and overall features are used to generate positive sample pairs, and the extracted background features and target features are used to generate negative sample pairs. The feature distance of the positive sample pairs is shortened, and the feature distance of the negative sample pairs is increased. These feature pairs are used to calculate similarity in contrastive learning.

[0044] In the comparison regularization method, the loss function of the multi-stage hierarchical adaptive partitioning network model is: , Where N is the number of input samples; C is the number of categories; Represents the true label of sample i. If the sample belongs to the cth category, then =1, otherwise 0; is the model's predicted probability that sample i belongs to class c; β is a weight coefficient used to balance the cross entropy loss and the contrast regularization term; is the contrast regularization term, used to calculate the background feature F b , target feature F t and the overall characteristic F m The relationship between them can optimize the feature learning process.

[0045] 3) Contrast classification module, which is used to classify images according to background features, target features and overall features.

[0046] 4) Common feature recognition module, which is used to enhance the background features, target features and overall features through a self-guided network method, and automatically recognize and emphasize the commonalities of background features, target features and overall features between images of the same type.

[0047] The multi-stage hierarchical adaptive partitioning network model further refines and optimizes feature representation through a self-guided network method. The multi-stage hierarchical adaptive partitioning network model better adapts to the complex variability of remote sensing data by automatically identifying and emphasizing the common features between samples in the same category, especially in data environments with high intrinsic variability and complexity, ensuring the stable performance of the multi-stage hierarchical adaptive partitioning network model and improving the accuracy of classification.

[0048] Example 3 The present invention also relates to a remote sensing image slice library image sampling method, referring to Figure 4 , samples the images in the image slice library, including the following steps: S21, obtain multi-source remote sensing images and build an image slice library; By acquiring remote sensing images from various data sources such as satellites, drones, and ground sensors, the remote sensing images are cropped into a uniform size to build an image slice library covering a variety of mission scenarios.

[0049] The sources of remote sensing image data cover a variety of geographical environments and application scenarios, such as urban planning, agricultural monitoring, forest protection, etc., ensuring that remote sensing image data represent potential mission requirements.

[0050] Cropping remote sensing images to a uniform size eliminates the need for complex radiation, atmospheric and geometric correction operations, simplifies the traditional cumbersome data preprocessing steps, simplifies the complexity of data processing, speeds up processing, lowers the technical threshold, and significantly improves processing efficiency. It is suitable for the rapid processing of large-scale remote sensing image data.

[0051] S22, using a multi-stage hierarchical adaptive partitioning network model to classify images in the image slice library; The image classification method in Example 1 is used to classify the images, which will not be described in detail here.

[0052] The multi-stage hierarchical adaptive partitioning network model needs to be trained before image classification. In order to meet the input specifications of the multi-stage hierarchical adaptive partitioning network model, images cropped to a uniform size in the image slice library are used to construct a training data set that meets the requirements. The multi-stage hierarchical adaptive partitioning network model is systematically trained using the training data set of labeled remote sensing image scene classification to ensure that the multi-stage hierarchical adaptive partitioning network model can accurately identify and classify different surface features.

[0053] S23, storing the image file of the image in the local file system, storing the metadata information of the image in the relational database, and establishing a mapping between the image file and the metadata information; specifically including: S231. Database preparation.

[0054] The database includes a relational database and a local file system. The relational database is used to store metadata information of the image, including image name, path, classification, upload date, etc. The local file system is used to store image files. By storing metadata information and image files separately, the storage pressure of large-scale data is reduced, and the overall performance and scalability of data storage are improved. Classification information no longer relies on folder storage, but is managed through many-to-many relationships in a relational database. Classification is formulated by domain experts or business needs based on specific application scenarios, and classification labels such as "forest", "city", and "water body" can be defined according to actual needs. New classification labels can be dynamically expanded according to business needs or technological changes. As application scenarios change or new targets are added, classifications can be added in the category table to flexibly adapt to more application scenarios.

[0055] S232: Build associations between images and categories.

[0056] In order to achieve the mapping between images and multiple categories, a many-to-many relationship model is used to create an intermediate table to connect images and categories and associate images with categories.

[0057] When an image belongs to two categories at the same time, first insert the image metadata information, then insert the category data information, and then associate the image with the category through the intermediate table. For example, if an image belongs to both "forest" and "water body" categories, first insert the image metadata information, then insert "forest" and "water body", then associate the image with the "forest" category, and associate the image with the "water body" category.

[0058] When an image does not have a suitable classification, a special unclassified is created and the image is associated with the unclassified through an intermediate table for subsequent manual review and processing.

[0059] S233. Processing and uploading of image files and classification files.

[0060] Image files are stored in a unified directory of the local file system instead of being placed in specific folders by category, and metadata information such as the file name, path, and category of each image is automatically written into the image_metadata table and image_category table in the relational database.

[0061] In the local file system, the classified folders processed by the multi-stage hierarchical adaptive partitioning network model are imported into a specific directory in the local file system. In this process, the metadata information such as the file name, path, and classification of each image will be automatically read and inserted into the relational database.

[0062] With this design, the physical file of the image only needs to be stored once, and the classification information is managed by the relational database regardless of how many categories it belongs to. This method not only saves storage space, but also improves retrieval efficiency. Users can flexibly query all images under a specific category through the relational database.

[0063] The separate storage structure of relational database and local file system is adopted, image files are stored in the file system, and metadata information is stored in the relational database, which realizes fast calling and efficient storage, makes query and management more flexible, reduces the storage pressure of large-scale data, improves the retrieval efficiency of sampling, and ensures the efficiency and scalability of processing large-scale data in the sampling process. Through this separate design, the classification information is decoupled from the image file, and the relational database manages the multi-classification information of the image, ensuring flexible query and management, while reducing the storage pressure of the relational database and improving the retrieval efficiency.

[0064] S24, searching and sampling images according to the image files in the local file system, the metadata information in the relational database and the total number of samples to obtain representative samples; specifically including: S241. Allocate the sampling quantity according to the category ratio based on the total sampling quantity input by the user.

[0065] According to the total number of samples, the number of images to be drawn in each category is calculated according to the proportion of the number of images in each category to the total number of images. For example, if the total number of samples entered by the user is 1,000, and the "forest" category accounts for 10% of the total number of images, 100 samples will be drawn from this category.

[0066] S242. The sampling quantity allocated according to the category ratio is retrieved and sampled in a stratified and grouped manner according to the resolution or scale within each category to obtain a representative sample.

[0067] Within each category, stratified and grouped sampling is performed according to different resolutions or scales to ensure that images of different resolutions are sampled fairly, making images at each resolution level representative and avoiding excessive sampling from images of a certain resolution, which would affect the balance and diversity of the overall sample.

[0068] S243. Dynamically load the sampled representative sample through the file path.

[0069] For example, use the returned image_path to display an image. This way, representative samples are displayed. Only the image path and related information need to be stored, without directly processing the image data, which reduces the burden on the database.

[0070] S25. Use a user interface to present a representative sample of sampling; The user can retrieve samples and display representative samples through the user interface, including the following steps: S251. The user logs in to the user interface by entering the user name and password, and enters the main interface after successful login.

[0071] S252. After entering the main interface, the user selects the "rule browsing" option or the "control query" option through the navigation bar or menu option according to the sampling requirements to enter the topic page.

[0072] S253. After entering the topic page, when the user enters keywords through the search bar to search for sampling, representative samples related to the keywords are automatically generated; for example, after entering "forest", "water body", etc., relevant image data information is quickly retrieved and relevant representative samples are output.

[0073] When you enter the topic page by selecting the "Rule Browse" option, the topic page presents a rule overview list, showing the category names of all image classifications. Users select the single category to be sampled through the category menu in the sidebar of the topic page, and then enter the sampling number to automatically generate a representative sample of the selected single category; or users select the "All Category Sampling" option in the sidebar of the topic page, and then enter the sampling number to automatically generate representative samples of all categories.

[0074] When entering the topic page by selecting the "Control Query" option, the user interface can be connected to the database, and the user can enter commands for precise retrieval sampling, and automatically generate representative samples related to the input commands. For example, display the image data of a specific path or modify the relevant information of a path image.

[0075] S25. After the sampling is completed, the user clicks the "Details" option, and the user interface will display detailed information of the representative sample, including the complete category name, storage path, release date, etc.

[0076] In addition, it will also display records such as whether the sample has been reclassified and the number of times the classification has changed, to provide detailed historical information and operation records of the sample slice.

[0077] Using a user interface to present representative samples can simplify data processing, improve classification efficiency, dynamically output representative samples of sampling, reduce operational complexity and training requirements, optimize user operating experience, greatly improve operational convenience and the accuracy of representative sample extraction, and overcome the problems of high computing costs, data processing complexity, and professional training requirements of traditional technologies, making sampling technology easier to popularize and use.

[0078] S26, performing error correction on image classification in the database according to user feedback information; S261. Obtain user feedback information.

[0079] User feedback information is the correction information fed back by the user when the representative sample is found to be misclassified in the image sampling. The correction message includes the wrong category and the expected correct category. For example, an image should be classified as "city" but is mistakenly classified as "forest".

[0080] After the user performs sampling, if the user finds that the classification of the representative samples in the sampling results is incorrect, the user can enter the feedback interface through the "Feedback Option" at the bottom of the special page. The feedback interface provides a detailed feedback filling guide. After the user enters the feedback information according to the instructions, the input content will be automatically formatted and transmitted to the back-end service. The back-end service automatically receives and stores the feedback information submitted by the user.

[0081] S262: Extract metadata information, misaligned category information, and corresponding correct category information of the misclassified image according to the user feedback information.

[0082] After receiving user feedback, the backend service automatically extracts key information, including metadata such as the file name of the misclassified image, the wrong category, and the correct category provided by the user. For example, the category of an image is identified as "forest", but the user feedback shows that the image should actually belong to the "city" category.

[0083] S263. Update the classification labels of the misclassified images in the database according to the correct category information.

[0084] According to the correct category information provided by the user, the classification label of the image in the relational database is directly updated without moving the image file. For example, if the category of an image is updated from "forest" to "city", the classification information in the relational database will be adjusted synchronously. At this time, there is no need to change the file path. The image is still stored in the local file system, and the classification is managed by the relational database.

[0085] S26. After updating the classification labels, the metadata information of the misclassified images in the relational database is synchronously updated.

[0086] After the classification label is updated, the metadata information in the relational database will be updated synchronously to ensure that the classification, description and other information of the image are consistent with the actual situation. In this way, the classification label and other metadata information of the image will be updated in a timely manner in the relational database.

[0087] S26. After the error correction update of the erroneously classified image is completed, the error correction update information and the change information before and after the error correction update are recorded and tracked.

[0088] After the image classification adjustment is completed, the updated information is passed to the backend service to record the changes between the original classification and the new classification of the image. These updates are recorded and tracked to further analyze the impact of user feedback on classification accuracy.

[0089] By performing image classification error correction based on user feedback information, when image classification errors are found during the image sampling process, the classification and metadata information of the images in the database can be automatically adjusted, thereby improving the accuracy of image sampling.

[0090] S27. Iteratively update the image according to the error correction information.

[0091] The multi-stage hierarchical adaptive partitioning network model is incrementally trained by local updating based on the error correction information.

[0092] In order to improve the accuracy of image classification in the remote sensing image slice library, the multi-stage hierarchical adaptive partitioning network model is incrementally trained based on error correction information. The incorrectly classified samples and the correctly classified samples are added back to the training data set, and the weights of the multi-stage hierarchical adaptive partitioning network model are adjusted to adapt to the new classification information. The specific process of incremental training includes adding the incorrectly classified samples and the correctly classified samples back to the training data set, loading the weights of the current multi-stage hierarchical adaptive partitioning network model, and training through local updates instead of starting from scratch. A small number of training rounds are performed using new data to ensure that the multi-stage hierarchical adaptive partitioning network model can adapt to the new classification rules and improve classification accuracy.

[0093] After the multi-stage hierarchical adaptive partitioning network model completes incremental training, it reclassifies the images in the entire image slice library. To ensure that the multi-stage hierarchical adaptive partitioning network model adapts to the new feedback information and accurately classifies similar samples, a secondary classification check is performed on the classified images to confirm that the new model can correctly distinguish the images that were originally misclassified.

[0094] The image classification is iteratively updated based on user feedback information, the classification performance of the multi-stage hierarchical adaptive partitioning network model is continuously optimized, the classification speed and accuracy of the multi-stage hierarchical adaptive partitioning network model are improved, thereby improving the accuracy of image sampling, achieving continuous improvement of the multi-stage hierarchical adaptive partitioning network model, and ensuring continuous improvement of classification accuracy.

[0095] The above sampling method makes the sampling technology more efficient and easy to popularize, significantly improves the processing speed, reduces the technical complexity, greatly improves the work efficiency, saves labor costs, and has important application value in the field of remote sensing image processing.

[0096] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A remote sensing image slice library image classification method, characterized in that: A multi-stage hierarchical adaptive partitioning network model is used to classify images in the image slice library; the following steps are included: S1, using the target and background separation strategy to separate the images in the image slice library into background images and target images; S2, using contrast regularization processing method, extracting background features from the background image, extracting target features from the target image, and extracting overall features from the background image and the target image; S3. Classify the image according to background features, target features and overall features.

2. The remote sensing image slice library image classification method according to claim 1, characterized in that: Step S1 includes: Convert images to grayscale and blur them; The blurred image is processed by Gaussian blur to enhance the contrast between the background and the target in the blurred image; According to the grayscale image and the blurred image after Gaussian blur processing, the image is separated into a background image and a target image through a mask.

3. The remote sensing image slice library image classification method according to claim 1, characterized in that: In the contrast regularization processing method, the extracted target features and overall features are used to generate positive sample pairs, and the extracted background features and target features are used to generate negative sample pairs, and the feature distance of the positive sample pairs is shortened, while the feature distance of the negative sample pairs is increased.

4. The remote sensing image slice library image classification method according to claim 1, characterized in that: It also includes using a self-guided network method to enhance the background features, target features and overall features, and automatically identifying and emphasizing the commonalities of background features, target features and overall features between images of the same type.

5. The remote sensing image slice library image classification method according to claim 1, characterized in that: In the comparison regularization method, the loss function of the multi-stage hierarchical adaptive partitioning network model is: , Where N is the number of input samples; C is the number of categories; Represents the true label of sample i. If the sample belongs to the cth category, then =1, otherwise 0; is the model's predicted probability that sample i belongs to class c; β is a weight coefficient used to balance the cross entropy loss and the contrast regularization term; is the contrast regularization term, used to calculate the background feature F b , target feature F t and the overall characteristic F m The relationship between.

6. A multi-stage hierarchical adaptive partitioning network model, characterized in that: Used to classify images according to a remote sensing image slice library image classification method according to any one of claims 1 to 5, comprising: A separation processing module is used to separate the images in the image slice library into background images and target images through a target and background separation strategy; A feature extraction module, used to extract background features in the background image, extract target features in the target image, and extract overall features in the background image and the target image by a contrast regularization processing method, comprising a background branch, a target branch, and a main branch, wherein the background branch is used to extract background features in the background image, the target branch is used to extract target features in the target image, and the main branch is used to extract overall features in the background image and the target image; The contrast classification module is used to classify images according to background features, target features and overall features.

7. A multi-stage hierarchical adaptive partitioning network model according to claim 6, characterized in that: It also includes a common feature recognition module for enhancing the background features, target features and overall features through a self-guided network method, and automatically identifying and emphasizing the commonalities of the background features, target features and overall features between images of the same type.

8. A remote sensing image slice library image sampling method, characterized in that: Sampling images in the image slice library includes the following steps: Classify the images using a remote sensing image slice library image classification method according to any one of claims 1 to 5; The image file of the image is stored in the local file system, the metadata information of the image is stored in the relational database, and a mapping between the image file and the metadata information is established; The images are retrieved and sampled according to the image files in the local file system, the metadata information in the relational database and the total sampling quantity to obtain a representative sample.

9. A remote sensing image slice library image sampling method according to claim 8, characterized in that: The method also includes the step of correcting image classification errors in a database according to user feedback information, wherein the user feedback information is correction information fed back by a user when a representative sample is found to have classification errors during image sampling.

10. A remote sensing image slice library image sampling method according to claim 9, characterized in that: The method further includes the step of iteratively updating the image according to the error correction information, including: Incremental training is performed on the multi-stage hierarchical adaptive partitioning network model by local updating based on the error correction information; After completing the incremental training, the multi-stage hierarchical adaptive partitioning network model reclassifies the images in the entire image slice library.

Images