Migration correction methods, systems and storage media for high-resolution remote sensing image annotation

Abstract

The embodiment of the application provides a kind of high-resolution remote sensing image annotation migration correction method, system and storage medium, belong to remote sensing image processing field.The method comprises: obtaining sample image and target image, wherein, sample image and target image all include at least one target object, and sample image is configured with the annotation information of target object;Target image is input into the prediction model trained in advance, and the target image prediction probability graph is obtained;Determine sample patch according to annotation information, and determine the matching area of sample patch in prediction probability graph;The similarity value of each pixel point in at least one sample patch and matching area is calculated, and each sample patch is matched based on the similarity value in matching area with maximum weight value overall constraint, to obtain target matching position, and finally obtain the target annotation result of target object in target image.The application can improve the accuracy of the same area, different time remote sensing image data annotation migration correction.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing, and in particular to a method, system, and storage medium for migration correction of high-resolution remote sensing image annotations. Background Technology

[0002] Remote sensing refers to non-contact, long-distance detection technology. In today's society, remote sensing technology plays an increasingly important role, providing strong technical and data support for various fields such as geographic information systems and spatial planning, agricultural production and natural resource management, weather forecasting and disaster monitoring, as well as infrastructure construction and traffic management.

[0003] When processing remote sensing imagery, data annotation is typically required to help operators better understand or further utilize the imagery. Once data annotation has been completed for remote sensing imagery of the same area at a specific time, analyzing remote sensing imagery of the same area at another time can reveal changes in surface information over different time periods. Related technologies involve annotation transfer of remote sensing imagery data annotations; however, differences in imaging angles and the sensors used for acquisition reduce the accuracy of this data annotation transfer. Summary of the Invention

[0004] The main objective of this application is to propose a method, system, and storage medium for the migration correction of high-resolution remote sensing image annotations, which can improve the accuracy of the migration correction of remote sensing image data annotations in the same area but at different times.

[0005] To achieve the above objectives, a first aspect of this application proposes a migration correction method for high-resolution remote sensing image annotation. The method includes: acquiring a sample image and a target image, wherein both the sample image and the target image include at least one target object, and the sample image is configured with annotation information of the target object; inputting the target image into a pre-trained prediction model to obtain pixel probability values ​​of each pixel in the target image as the target object, and establishing a prediction probability map of the target image based on each pixel probability value; determining the sample patch where the target object is located in the sample image based on the annotation information, and determining the matching region of the sample patch in the prediction probability map; calculating the similarity value between at least one sample patch and each pixel in the matching region, and constraining the matching of each sample patch in the matching region based on the similarity value to obtain a target matching position, so that each sample patch after constraint matching corresponds as a whole to the target object in the target image; and obtaining the target annotation result of the target object in the target image based on the target matching position.

[0006] In some embodiments, a target patch is obtained based on the target annotation result of the target image. The target patch includes a target patch feature value and target patch annotation information corresponding to the target patch feature value. The sample patch includes a sample patch feature value and sample patch annotation information corresponding to the sample patch feature value. After obtaining the target annotation result of the target object in the target image based on the target matching position, the method further includes: obtaining a weight parameter value for augmentation; performing a difference operation between the weight parameter value and a unit weight to obtain a difference weight parameter value; and comparing the weight parameter value with the sample patch feature value. Multiply the values ​​to obtain a first sub-value; multiply the difference weight parameter value with the target patch feature value to obtain a second sub-value; add the first sub-value and the second sub-value to obtain first augmented information; multiply the weight parameter value with the sample patch annotation information to obtain a third sub-value; multiply the difference weight parameter value with the target patch annotation information to obtain a fourth sub-value; add the third sub-value and the fourth sub-value to obtain second augmented information; based on the first augmented information and the second augmented information, an augmented and improved sample is obtained, which is used to iteratively improve the target annotation result.

[0007] In some embodiments, determining the sample patch in the sample image where the target object is located based on the annotation information includes: acquiring the positioning reference point and resolution parameters of the target image; adjusting the positioning and resolution of the sample image based on the positioning reference point and the resolution parameters to make the coordinate position and resolution of the adjusted sample image consistent with those of the target image, thereby obtaining a preprocessed sample image; performing feature transformation on the preprocessed sample image based on the image features of the target image to obtain a transformed sample image, and obtaining the transformed annotation information of the sample image based on the transformed sample image; and determining the sample patch in the sample image where the target object is located based on the transformed annotation information.

[0008] In some embodiments, the step of performing feature transformation on the preprocessed sample image based on the image features of the target image to obtain a transformed sample image includes: acquiring first spectral information of the preprocessed sample image and second spectral information of the target image; if the first spectral information and the second spectral information are within a preset spectral range, performing a relative radiometric correction operation on the preprocessed sample image to obtain the transformed sample image; if the first spectral information and the second spectral information are not within the preset spectral range, performing a linear quantization operation on the preprocessed sample image so that the linearly quantized first spectral information and the second spectral information are within the preset spectral range to obtain the transformed sample image; or, performing a style transfer operation on the preprocessed sample image based on a pre-trained network model to obtain the transformed sample image.

[0009] In some embodiments, the step of inputting the target image into a pre-trained prediction model to obtain pixel probability values ​​for each pixel in the target image as the target object, and establishing a prediction probability map of the target image based on each pixel probability value, includes: acquiring a pre-trained prediction model; inputting the target image into the prediction model, performing probability prediction on each pixel in the target image according to the prediction model, and obtaining a prediction probability value for each pixel; if the prediction probability value is within a preset probability range, determining that the pixel is a region label of the target image; calculating the vegetation index of the target image, and determining a green space mask of the target image based on the vegetation index; expanding the prediction region based on the region label and the green space mask, and obtaining a prediction probability map corresponding to the target image based on the expanded prediction region; and obtaining a prediction probability map of the target image based on the prediction region.

[0010] In some embodiments, the pre-trained prediction model is trained through the following steps: obtaining a preset prediction training set, the prediction training set including multiple training sample images for training, and setting corresponding training labels for the training sample images; selecting any one of the training sample images from the prediction training set and inputting it into the prediction model to obtain training prediction probability values ​​for each sample region of the training sample image; calculating the prediction loss value of the prediction model based on the training prediction probability value and the training labels, and adjusting the parameters of the prediction model based on the prediction loss value to obtain the trained prediction model.

[0011] In some embodiments, determining the matching region of the sample patch in the prediction probability map includes: acquiring geographic coordinate information; outputting the sample patch to the prediction probability map according to the geographic coordinate information to obtain a constraint matching map; acquiring a preset search radius, wherein the search radius is set according to the prediction probability map; and determining the matching region of the sample patch on the constraint matching map according to the search radius, wherein the matching region includes four sub-matching regions.

[0012] In some embodiments, after calculating the similarity value between at least one of the sample patches and each pixel in the matching region, the method further includes: obtaining a preset distance threshold; determining the pixel corresponding to the maximum similarity value of each sub-matching region as a sub-candidate matching point for each sub-matching region; selecting the sub-candidate matching point with the largest similarity value among the sub-candidate matching points as a baseline matching point; if the distance between the remaining sub-candidate matching points and the baseline matching point is less than the distance threshold, determining the sub-candidate matching point as an error matching point; and determining the candidate matching point corresponding to the sample patch based on the sub-candidate matching point and the error matching point.

[0013] In some embodiments, after determining the candidate matching point corresponding to the sample patch based on the sub-candidate matching point and the error matching point, the method further includes: if the distance between the first candidate matching point in the first matching region and the second candidate matching point in the second matching region is less than the preset distance threshold, determining that the first candidate matching point and the second candidate matching point are the same candidate matching point; and determining that the sample patch corresponding to the same candidate matching point includes the first sample patch corresponding to the first candidate matching point and the second sample patch corresponding to the second candidate matching point.

[0014] In some embodiments, constraining the matching of each sample patch in the matching region based on the similarity value to obtain the target matching position includes: determining a weighted bipartite graph based on the candidate matching points corresponding to each sample patch and the sample patch corresponding to each candidate matching point, using the similarity of each candidate matching point as the weight, wherein the weighted bipartite graph is used to characterize the weight relationship between each sample patch and each candidate matching point; and determining the target matching position based on each sample patch and each candidate matching point from the weighted bipartite graph using the overall maximum weight matching as a constraint condition.

[0015] In some embodiments, determining the target matching position from the weighted bipartite graph based on each sample patch and each of the candidate matching points includes: if the similarity value of the sample patch at the target matching position is not lower than a preset similarity threshold, obtaining the corresponding sample patch; if the sample patch has no target matching position, obtaining the deleted sample patch; if the similarity value of the sample patch is lower than the similarity threshold, obtaining the deleted sample patch or the newly added sample patch; and obtaining the target annotation result based on the corresponding sample patch, the deleted sample patch, and the newly added sample patch, wherein the target annotation result includes the target annotation and the certainty corresponding to the target annotation, and the certainty is used to characterize the credibility measure between the target annotation and the actual annotation result.

[0016] To achieve the above objectives, a second aspect of this application proposes a migration correction system for high-resolution remote sensing image annotation. The system includes: an acquisition module for acquiring a sample image and a target image, wherein both the sample image and the target image include at least one target object, and the sample image is configured with annotation information for the target object; a probability prediction module for inputting the target image into a pre-trained prediction model to obtain pixel probability values ​​for each pixel in the target image as the target object, and establishing a prediction probability map of the target image based on each pixel probability value; a matching module for determining a sample patch in the sample image where the target object is located based on the annotation information, and determining a matching region for the sample patch in the prediction probability map; and a calculation module for calculating a similarity value between at least one sample patch and each pixel in the matching region, and constraining the matching of each sample patch in the matching region based on the similarity value to obtain a target matching position, so that each constrained and matched sample patch corresponds as a whole to the target object in the target image.

[0017] The result module is used to obtain the target annotation result of the target object in the target image based on the target matching position.

[0018] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0019] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0020] The present application proposes a high-resolution remote sensing image annotation migration correction method, system, and storage medium. The high-resolution remote sensing image annotation migration correction method can be applied to a high-resolution remote sensing image annotation migration correction system. This application acquires sample images and target images, both including at least one target object, and the sample images contain annotation information for the target object. Next, the target image is input into a pre-trained prediction model to obtain a predicted probability map of the target image. Then, constraint matching is performed on the predicted probability map of the target image based on the annotation information to determine the target matching position, so that each sample patch after constraint matching corresponds to the target object in the target image as a whole. Finally, based on the target matching position, the target annotation result of the target object in the target image is obtained, which can improve the accuracy of annotation migration correction for remote sensing image data from the same region but different time phases. Attached Figure Description

[0021] Figure 1A This is a schematic diagram of a remote sensing target object provided in an embodiment of this application;

[0022] Figure 1B This is a schematic diagram of a remote sensing target object from a rear-view perspective provided in an embodiment of this application;

[0023] Figure 1C This is a schematic diagram of a forward-looking image of a remotely sensed target object provided in an embodiment of this application;

[0024] Figure 2 This is a schematic diagram illustrating an application scenario of the migration correction system for high-resolution remote sensing image annotation provided in this application embodiment;

[0025] Figure 3 This is an optional flowchart of the high-resolution remote sensing image annotation migration correction method provided in the embodiments of this application;

[0026] Figure 4 This is a schematic diagram of a sample image provided in an embodiment of this application;

[0027] Figure 5 This is a schematic diagram of the target image provided in the embodiments of this application;

[0028] Figure 6 This is a schematic diagram of sample image annotation provided in the embodiments of this application;

[0029] Figure 7A This is a schematic diagram of target image annotation provided in an embodiment of this application;

[0030] Figure 7B This is another schematic diagram of target image annotation provided in an embodiment of this application;

[0031] Figure 8This is another optional flowchart of the high-resolution remote sensing image annotation migration correction method provided in the embodiments of this application;

[0032] Figure 9 This is an optional hybrid augmentation flowchart of the migration correction method for high-resolution remote sensing image annotation provided in the embodiments of this application;

[0033] Figure 10 yes Figure 3 A flowchart of an implementation of step S103 in the process;

[0034] Figure 11 yes Figure 10 A flowchart of an implementation of step S303 in the process;

[0035] Figure 12 yes Figure 3 A flowchart of an implementation of step S102 in the process;

[0036] Figure 13 This is another optional flowchart of the high-resolution remote sensing image annotation migration correction method provided in the embodiments of this application;

[0037] Figure 14 yes Figure 3 Another implementation flowchart of step S103 in the process;

[0038] Figure 15 This is a schematic diagram of constraint matching for sample patches provided in an embodiment of this application;

[0039] Figure 16 yes Figure 3 A flowchart of the implementation following step S105;

[0040] Figure 17 yes Figure 16 A flowchart of the implementation following step S805;

[0041] Figure 18 This is a flowchart of an alternative matching point merging method for the migration correction of high-resolution remote sensing image annotation provided in this application embodiment;

[0042] Figure 19 yes Figure 3 A flowchart of an implementation of step S105 in the process;

[0043] Figure 20 yes Figure 19 A flowchart of an implementation of step S1002;

[0044] Figure 21 This is a schematic diagram of the functional module for migration correction of high-resolution remote sensing image annotation provided in the embodiments of this application;

[0045] Figure 22 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0049] First, let's analyze some of the terms used in this application:

[0050] Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0051] Natural Language Processing (NLP): NLP uses computers to process, understand, and utilize human language (such as Chinese and English). NLP is a branch of artificial intelligence and an interdisciplinary field of computer science and linguistics, often referred to as computational linguistics. NLP includes syntactic analysis, semantic analysis, and discourse understanding. It is commonly used in machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, intent recognition, information extraction and filtering, text classification and clustering, sentiment analysis, and opinion mining. It involves data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computation.

[0052] Remote sensing refers to non-contact, long-distance detection technology. It generally refers to the detection of the electromagnetic radiation and reflection characteristics of objects using sensors / remote sensors. Remote sensing enables the detection of target features from a distance and without contact with the target object, through instruments such as remote sensors that are sensitive to electromagnetic waves.

[0053] Remote sensing technology specifically includes satellite remote sensing and airborne remote sensing. Both satellite remote sensing and airborne remote sensing are used to acquire information about the Earth's surface for purposes such as urban planning, land change monitoring, and disaster assessment.

[0054] With the continuous development of technology, the use of satellite and aerial remote sensing to monitor building tops at different times is becoming increasingly common for building structure assessment, urban planning and construction, disaster detection and response, and other purposes. Furthermore, after photographing building tops, data annotation is required to distinguish between building top areas and non-building top areas. To avoid unnecessary waste of human resources, the corresponding annotations for building tops are typically migrated between remote sensing images of the same area but from different times.

[0055] However, due to differences in imaging time, imaging angle, and the use of different sensors, the relative position of the building top in different images can vary significantly, which makes annotation migration difficult.

[0056] For example, such as Figure 1A As shown, Figure 1A This is a schematic diagram of a remote sensing target object provided in an embodiment of this application. Building 1 and building 2 are two target buildings (hereinafter referred to as "objects") in the remote sensing image, and object 1 and object 2 have different heights, with object 1 being taller than object 2. Figure 1B As shown, Figure 1BThis is a schematic diagram of a remote sensing target object from a rearward perspective provided in an embodiment of this application. In remote sensing imaging, when viewed from above, object 1 and object 2 are on the same horizontal line as the imaging reference point (hereinafter referred to as...). Figure 3 In the embodiments described, this is also used as the imaging reference point. If the imaging angle during remote sensing is biased towards the rear of the imaging reference point, due to the different heights of object 1 and object 2, in the remote sensing image generated from the rear-view image, the taller object 1 will appear higher in the final image compared to the shorter object 2. Figure 1C As shown, Figure 1C This is a schematic diagram of a forward-looking image of a remotely sensed target object provided in an embodiment of this application. Similarly, in remote sensing imaging, because object 1 and object 2 have different heights, in the remote sensing image generated from the forward-looking image, the taller object 1 will appear lower in the final image compared to the shorter object 2. It should be noted that polar-orbiting satellites generally fly from north to south, and looking north is a rear-view perspective; therefore, in... Figure 1B and Figure 1C In the middle, north is at the top and south at the bottom, with north being the back.

[0057] In other words, when remotely sensing objects at different altitudes, different imaging angles will affect the final remote sensing image, further reducing the accuracy of the data annotation results when transferring data annotations from remote sensing images of the same area at different times. Furthermore, the sensors used to capture remote sensing images at different times are not necessarily the same. Since remote sensing images captured by different sensors have different spectral response characteristics and spatial resolutions, these factors will also affect the data annotation transfer. Therefore, when transferring data annotations from remote sensing images of the same area at different times, the accuracy of the data annotation transfer is easily reduced due to differences in imaging angles and the sensors used for capture.

[0058] Based on this, embodiments of this application provide a method, system, and storage medium for migration correction of high-resolution remote sensing image annotations. The method involves acquiring sample images and target images, where each image includes at least one target object, and the sample image contains annotation information for the target object. Next, the target image is input into a pre-trained prediction model to obtain a predicted probability map of the target image. Then, constraint matching is performed on the predicted probability map of the target image based on the annotation information to determine the target matching position, ensuring that each sample patch after constraint matching corresponds to the target object in the target image. Finally, based on the target matching position, the target annotation result of the target object in the target image is obtained, which can improve the accuracy of annotation migration correction for remote sensing image data from the same region but different time phases.

[0059] For example, such as Figure 2 As shown, Figure 2 This is a schematic diagram illustrating an application scenario of the high-resolution remote sensing image annotation migration correction system provided in this application embodiment. The system includes a first subsystem 11, a second subsystem 12, and a terminal server 10 for controlling each subsystem. The terminal server 10 acquires sample images and target images from the areas where the first subsystem 11 and the second subsystem 12 are located, and performs data processing based on the acquired sample images and target images to obtain the target annotation result of the target object in the target image. It should be noted that the terminal server 10 includes at least one subsystem; that is, the terminal server 10 can be connected only to the first subsystem 11, or it can be connected to a third, fourth, or more subsystems, depending on the actual needs of the operator. This application embodiment does not impose specific limitations on this.

[0060] Based on this, the migration correction method for high-resolution remote sensing image annotation in the embodiments of this application can be described through the following examples.

[0061] In some embodiments, the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that utilize digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0062] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0063] The high-resolution remote sensing image annotation migration correction method provided in this application relates to the field of artificial intelligence technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the high-resolution remote sensing image annotation migration correction method, but is not limited to the above forms.

[0064] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0065] It should be noted that in all specific embodiments of this application, when processing data related to the remote sensing area, such as remote sensing image information, is required, permission or consent from relevant personnel within the remote sensing image area will be obtained first. For example, when acquiring remote sensing images in area A, permission or consent from relevant personnel in area A will be obtained first. Furthermore, the collection, use, and processing of this remote sensing data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require the acquisition of sensitive information in the area, such as building information, the necessary remote sensing data for the proper functioning of these embodiments will only be acquired after obtaining separate permission or consent from relevant personnel.

[0066] like Figure 3 As shown, Figure 3This is an optional flowchart of the high-resolution remote sensing image annotation migration correction method provided in the embodiments of this application. Figure 3 The method may include, but is not limited to, steps S101 to S105.

[0067] Step S101: Obtain sample image and target image, wherein both sample image and target image include at least one target object, and the sample image is configured with annotation information of the target object;

[0068] In some embodiments, remote sensing technology is typically used to image the same area at different time phases to obtain sample images and target images, wherein the sample images and target images include the target object of interest. In this embodiment, the roof of a building will be used as the target object for detailed description.

[0069] In some embodiments, both the sample image containing the target object (the top of the building) and the target image are orthophoto images, meaning that the relevant remote sensing sensor will take a vertical image from directly above the building. Furthermore, the sample image often includes annotation information for the top of the building, which is used to characterize the area representing the top of the building.

[0070] For example, such as Figure 4 , Figure 5 As shown, Figure 4 This is a schematic diagram of a sample image provided in an embodiment of this application. Figure 5 This is a schematic diagram of the target image provided in an embodiment of this application. Figure 4 The sample image in the image is a panchromatic image formed by GF-7 at a rear-view resolution of 0.6 meters on March 14, 2020 (there is also a 2.4-meter resolution multispectral image with visible light plus near-infrared 4 bands; generally, panchromatic and multispectral images are superimposed or fused for input). Figure 5 The target image in the image is a panchromatic image formed by GF-2 at a resolution of 0.8 meters on September 15, 2017 (another 3.2-meter resolution multispectral image with visible light and near-infrared 4 bands is also available; the panchromatic and multispectral images are superimposed or fused for input).

[0071] Understandable Figure 4 Sample images and Figure 5 The target images in the two images were taken from the same area using different sensors, and the imaging angles of the two remote sensing images are different. Figure 4 The sample images in the sample are more likely to be taken from a rear-view perspective, while Figure 5The target image is more likely to be taken from a forward-facing perspective. In other words, the image resolution, season, pixel range, and imaging angle of the sample image and the target image are different. Under such circumstances, it is necessary to transform and transfer the sample image to be similar to the target image in terms of resolution, positioning, color, and texture, and to accurately transfer the annotation information of the sample image in order to obtain the annotation information of the building top of the target object in the target image.

[0072] For example, such as Figure 6 , Figure 7A As shown, Figure 6 This is a schematic diagram of sample image annotation provided in the embodiments of this application. Figure 7A This is a schematic diagram of target image annotation provided in an embodiment of this application, wherein, Figure 6 This is the annotation result of overlaying sample annotations from the sample image onto the GF-7 panchromatic image. In this embodiment, the annotation area corresponds to the top of the building (excluding the building facade); while Figure 7A The annotations in the sample image are transformed and then superimposed onto the GF-2 panchromatic image. It can be observed that... Figure 7A The labeling information on the tops of the taller buildings on the left is basically correct, while the labeling information on the tops of the slightly shorter buildings on the right is inconsistent. In other words, from... Figure 6 , Figure 7A As can be seen, when the annotation information of the sample image is transferred to the target image, a disaster of detail occurs, and even a mismatch occurs. Overall, the transfer of annotation information is inaccurate.

[0073] To address the aforementioned problems, this application proposes a high-resolution remote sensing image annotation migration correction method. This method solves the problem in related technologies where the accuracy of remote sensing image annotation migration decreases due to differences in imaging angles and the sensors used for acquisition. Furthermore, after applying the high-resolution remote sensing image annotation migration correction method proposed in this application, Figure 6 After migration correction, the target image obtained from the sample images is as follows: Figure 7B As shown, Figure 7B This is another target image annotation diagram provided in the embodiments of this application. It can be understood that the target image annotation after migration and correction by this method has higher accuracy and can better fit the target object in the target image.

[0074] In some embodiments, remote sensing images of the same area at different times are acquired to obtain sample images and target images, wherein the sample images and target images include at least the same target object, and the sample images may be later in time than the target images, or they may be earlier in time than the target images.

[0075] In some embodiments, the sample image is further configured with annotation information of the target object corresponding to the sample image. This annotation information is used to characterize the data information of the corresponding region, such as in... Figure 6 In the text, the annotation information indicates that the marked area is the rooftop of the building.

[0076] Step S102: Input the target image into the pre-trained prediction model to obtain the pixel probability value of each pixel in the target image as the target object, and establish the prediction probability map of the target image based on the pixel probability value.

[0077] In some embodiments, the target image can be input into a pre-trained prediction model, which predicts the probability value of each pixel in the target image being labeled, and can obtain a prediction probability map of the target image based on the probability value of each pixel. This prediction probability map can roughly predict the area of ​​the building top, but the building prediction area represented by the prediction probability map cannot be used as the final target labeling result of the target image, as this is inaccurate. It is also necessary to perform constraint matching based on the labeling information in the sample image to obtain the final target labeling result. Furthermore, constraint matching based on the prediction probability map and the labeling information in the sample image can achieve accurate transfer of the labeling information of the sample image.

[0078] In some embodiments, the prediction model is a semantic segmentation model, which can be selected from FCN (Fully Convolutional Networks), UNet, DeepLab (Deep Convolutional Networks), or Mask R-CNN (Mask Region Convolutional Neural Network) according to actual needs.

[0079] Step S103: Determine the sample patch where the target object is located in the sample image based on the annotation information, and determine the matching area of ​​the sample patch in the prediction probability map;

[0080] In some embodiments, since the sample image is configured with annotation information, sample patches of the target object can be determined based on the annotation information. For example, Figure 6 The sample includes three patches. Next, the sample patches are mapped to the predicted probability map to obtain the matching regions of the samples. The purpose of this is to allow for further adjustments based on the obtained results, in order to obtain more accurate target annotations for the target image.

[0081] In some embodiments, sample patches can also be obtained by taking the bounding rectangle from the annotation information.

[0082] Step S104: Calculate the similarity value between at least one sample patch and each pixel in the matching region, and constrain the matching of each sample patch in the matching region based on the similarity value to obtain the target matching position, so that the constrained matching sample patch corresponds to the target object in the target image as a whole.

[0083] In some embodiments, when there is only one sample patch in the sample image, the top-left pixel of the sample patch is selected as the target pixel, and the similarity between the target pixel and each pixel in the matching region is calculated to obtain multiple candidate matching points. After filtering out error matching points, the pixel with the highest similarity value is determined as the target matching position of the sample patch. When the sample image contains multiple sample patches, the top-left pixel of each sample patch is also selected as the target pixel, and the similarity between the target pixel and each pixel in the corresponding matching region is calculated to obtain multiple candidate matching points. After filtering out error matching points, the target matching position of each sample patch is determined. Furthermore, the target matching position enables each sample patch after migration matching to correspond as a whole with the target object in the target image, rather than a local correspondence (e.g., ...). Figure 6 (The situation in the middle).

[0084] Step S105: Based on the target matching location, obtain the target annotation result of the target object in the target image.

[0085] In some embodiments, target annotation results for each target object in the target image can be obtained based on the target matching location. When the target annotation results are superimposed on the target image, they can correspond to the target objects as a whole, that is, the overall matching accuracy is high.

[0086] It should be noted that the high-resolution remote sensing image annotation migration correction method proposed in this application is also applicable to remote sensing images of the same area but at different times captured by the same remote sensing sensor. Alternatively, when the annotation information in the sample image corresponds to the top area of ​​a building and the facade area of ​​a building, the high-resolution remote sensing image annotation migration correction method proposed in this application can also be used. The embodiments in this application are only described as preferred embodiments and are not intended to impose specific limitations.

[0087] Understandably, this application acquires sample images and target images, both of which include at least one target object, and the sample images contain annotation information for the target object. Next, the target image is input into a pre-trained prediction model to obtain a predicted probability map of the target image. Then, constraint matching is performed on the predicted probability map of the target image based on the annotation information to determine the target matching position, so that each sample patch after constraint matching corresponds to the target object in the target image as a whole. Finally, based on the target matching position, the target annotation result of the target object in the target image is obtained, which can improve the accuracy of annotation migration correction for remote sensing image data from the same region but different time phases.

[0088] like Figure 8 As shown, Figure 8 This is another optional flowchart of the high-resolution remote sensing image annotation migration correction method provided in the embodiments of this application. Figure 8 The method may include, but is not limited to, steps S201 to S205.

[0089] Step S201: Obtain the weight parameter values ​​used for expansion;

[0090] In some embodiments, target patches can be obtained based on the target annotation results of the target image. That is, the target annotation results are the overall annotation of the target image, including at least one target patch corresponding to a target annotation. Each target patch includes a target patch feature value and the corresponding target patch annotation information. Similarly, sample patches include sample patch feature values ​​and the corresponding sample patch annotation information. The sample patch feature values, sample patch annotation information, target patch feature values, and target patch annotation information can be used as a basis for augmentation to obtain augmented samples. These augmented samples can be used to iteratively enhance the sample image and the target image to further improve the accuracy of annotation transfer.

[0091] For example, the target annotation result represents the area in the target image that is not labeled as the top of a building. Each area labeled as the top of a building is a target patch. Characterizing the feature values ​​of sample patches in a sample image, where the feature values ​​of sample patches can represent the shape, spectrum, and texture of the corresponding sample patches, etc. The sample annotation information represents the sample patch and indicates whether the labeled area is the top of a building; similarly, the target patch feature value and target annotation information of the target image are obtained. .

[0092] In some embodiments, the weight parameter values ​​can be determined based on the beta distribution. get.

[0093] Step S202: Perform a difference calculation between the weight parameter value and the unit weight to obtain the difference weight parameter value;

[0094] In some embodiments, the unit weight is 1, and the difference between the unit weight 1 and the weight parameter value obtained in step S201 is calculated, i.e., (1- This allows us to obtain the difference weight parameter value.

[0095] Step S203: Multiply the weight parameter value by the sample patch feature value to obtain the first sub-value; multiply the difference weight parameter value by the target patch feature value to obtain the second sub-value; add the first sub-value and the second sub-value to obtain the first augmented information.

[0096] In some embodiments, parameters are taken according to the beta distribution. ,generate :

[0097]

[0098] in, These are the weight parameter values. As the first sub-value, As the second sub-value, This is the first piece of additional information.

[0099] Step S204: Multiply the weight parameter value with the sample patch annotation information to obtain the third sub-value; multiply the difference weight parameter value with the target patch annotation information to obtain the fourth sub-value; add the third sub-value and the fourth sub-value to obtain the second augmented information.

[0100] In some embodiments, parameters are taken according to the beta distribution. ,generate :

[0101]

[0102] in, These are the weight parameter values. As the third sub-value, It is the fourth sub-value. This is the second piece of additional information.

[0103] It is understandable that in steps S203 and S204... To achieve the same result, the sample image and the target image need to be weighted and mixed under the same weight parameters. Afterwards, parameters can be taken multiple times according to the beta distribution. In which, each time the amount taken They can be different, and depend on the newly acquired ones. The sample image and the target image are then weighted and mixed again to obtain the further augmented sample.

[0104] Step S205: Based on the first augmented information and the second augmented information, the augmented sample after augmentation and improvement is obtained. The augmented sample is used to iteratively improve the target annotation results.

[0105] In some embodiments, the augmented image is determined to be an augmented sample. Alternatively, the augmented sample can be iterated multiple times, that is, the augmented sample can be used as the target image and then mixed and augmented with the sample image again to improve the quality of the output sample set.

[0106] In some embodiments, such as Figure 9 As shown, Figure 9 This is an optional hybrid augmentation flowchart of the high-resolution remote sensing image annotation migration correction method provided in the embodiments of this application. Figure 9 In this context, the sample image is the original sample set, and the target image is the new image. The target annotation results for the new image can be obtained through the following steps:

[0107] Step 1: Preprocessing to ensure that the coordinates and resolution of the sample are consistent with those of the new image. The preprocessing operation includes extracting the green area mask.

[0108] Step 2: Sample set feature transformation. Use relative radiometric correction or add a data style transfer algorithm to transform the image data blocks of the original sample set so that the spectral and texture representation of the transformed image blocks is consistent with the representation of the new image.

[0109] Step 3: Train a classifier using the converted sample set to predict the probability distribution map of the building tops in the new image, and process the predicted probability map with a conditional random field and add morphological dilation filtering.

[0110] Step 4: Constrain matching on the probability map for each sample image, solve the maximum weight matching problem using the bipartite graph, obtain the preliminary matching position of each sample patch, determine the deleted patch and the added patch, where the deleted patch corresponds to the demolished building, and the added patch corresponds to the added building, and output the target annotation result and the corresponding certainty of the target image of the migration correction.

[0111] Step 5: The results output from Step 2 and Step 4 can be mixed and augmented in a one-to-one correspondence, and Step 3 and Step 4 can be re-executed with the augmented samples to iteratively improve and output a high-quality migration correction sample set and corresponding certainty. It should be noted that iterative enhancement can also be performed multiple times according to the characteristics of the image data and the quality of the results to obtain a high-quality migration sample set and certainty.

[0112] It is understandable that, based on the above steps S101 to S105, the accuracy of annotation migration can be further improved by mixing and augmenting the sample image and the target image.

[0113] like Figure 10 As shown, Figure 10 yes Figure 3 A flowchart of an implementation of step S103 is provided. In some embodiments, step S103 may include steps S301 to S304:

[0114] Step S301: Obtain the positioning reference points and resolution parameters of the target image;

[0115] In some embodiments, since the imaging angle of the target object in the sample image and the target image is different from that of the remote sensing sensor, the sample image needs to be preprocessed first so that the positioning and resolution of the preprocessed sample image are consistent with those of the target image.

[0116] In some embodiments, the positioning reference points and resolution parameters of the target image are obtained based on the target image to preprocess the sample image.

[0117] Step S302: Based on the positioning reference point and resolution parameters, the sample image is adjusted in terms of positioning and resolution so that the coordinate position and resolution of the adjusted sample image and the target image are consistent, thus obtaining the preprocessed sample image.

[0118] In some embodiments, the sample image and its corresponding annotation information are transformed through matching and geometric correction to achieve the same coordinate positioning and resolution as the target image. For example, the sample image and the target image can first be made to have essentially the same image size through several transformation methods such as rotation and translation. Then, the coordinate positioning of the sample image and the target image can be made consistent through special point positioning. Finally, the resolution can be adjusted by computer to make the resolutions of the two images consistent.

[0119] Step S303: Perform feature transformation on the preprocessed sample image according to the image features of the target image to obtain the transformed sample image, and obtain the annotation information of the transformed sample image based on the transformed sample image.

[0120] In some embodiments, since remote sensing images are acquired by remote sensing sensors, which can obtain information in different bands by sensing electromagnetic radiation in different bands, this information can be divided into spectral information and spatial information (e.g., texture information). Therefore, in the case where the sample image and the target image are captured by different sensors in the embodiments of this application, it is still necessary to use the target image as a reference to perform feature transformation on the spectral and texture details of the sample image, so that the transformed sample image and the corresponding annotation information can be consistent with the target image, thereby improving the accuracy of data annotation migration correction.

[0121] Step S304: Based on the converted annotation information, determine the sample patch in the sample image where the target object is located.

[0122] In some embodiments, after feature transformation of the sample image, the transformed sample image and the corresponding annotation information can be obtained. Based on the annotation information, the sample patch of the target object in the sample image can be determined. The spectral and texture features of the pixels of the sample patch obtained after feature transformation are consistent with those in the target image. The purpose is to reduce the disturbance caused by interference information and improve the accuracy of data annotation migration correction.

[0123] like Figure 11 As shown, Figure 11 yes Figure 10 A flowchart of an implementation of step S303 is provided. In some embodiments, step S303 may include steps S401 to S404:

[0124] Step S401: Obtain the first spectral information of the preprocessed sample image and the second spectral information of the target image;

[0125] In some embodiments, using the target image as a reference image, feature transformation is performed on the sample image to ensure that the image features of the transformed sample image are consistent with those of the target image. Based on this, it is necessary to obtain the first spectral information of the sample image after preprocessing for positioning and resolution, and the second spectral information of the target image.

[0126] Step S402: If the first spectral information and the second spectral information are within the preset spectral range, perform relative radiometric correction on the preprocessed sample image to obtain the converted sample image.

[0127] In some embodiments, an effective similarity range is preset between the first spectral information of the sample image and the second spectral information of the target image. If the first spectral information and the second spectral information are within this range, it can be determined that the sample image and the target image are directly subjected to relative radiometric correction. Its function is to eliminate the brightness difference in remote sensing images caused by factors such as shooting angle and atmospheric influence, so as to improve the quality and reliability of remote sensing images and enable remote sensing data to more accurately reflect the changes and characteristics of surface information.

[0128] In some embodiments, a histogram-based relative radiometric correction method is used to perform histogram statistics on the pixel brightness of the sample image and the target image, obtaining the brightness value distribution of the corresponding sample image and the target image, respectively. A relative brightness ratio coefficient is calculated in the sample image and the target image, and brightness adjustment, i.e., relative radiometric correction, is performed based on this relative brightness ratio coefficient to enhance the brightness uniformity and stability of the converted sample image.

[0129] Step S403: If the first spectral information and the second spectral information are not within the preset spectral range, perform linear quantization on the preprocessed sample image so that the linearly quantized first spectral information and the second spectral information are within the preset spectral range, and obtain the converted sample image.

[0130] In some embodiments, if the first spectral information and the second spectral information are not within a preset effective similarity range, linear quantization processing of the sample image is required. Specifically, the quantization method expression is as follows: ,in The original pixel values ​​and the transformation result are given. The maximum and minimum values ​​of the sample image block bands. The maximum and minimum effective values ​​of the corresponding bands in the new image; if a certain band of the sample image block does not have a corresponding value in the new image, then the band remains unchanged; alternatively, all bands of the sample image block and the new image can be normalized to between 0 and 1 according to their respective ranges.

[0131] Step S404, or, based on the pre-trained network model, perform style transfer operation on the pre-processed sample image to obtain the transformed sample image.

[0132] In some embodiments, if the difference between the first spectral information and the second spectral information of the sample image and the target image is too large, such as far exceeding the preset effective similarity range, a deep learning network model for style transfer can be trained, and the trained network model can be used to transform the sample image so that its spectral and texture representation is similar to that of the target image.

[0133] In some embodiments, the sample image can be divided into multiple sample image blocks to form a sample image block set, which is used for training the network model. Furthermore, the sample image block set used for training the style transfer network model is selected according to the data characteristics: (1) the sample image block set can be randomly selected within the effective range; (2) it can also be selected with a fixed step size and random perturbation, with overlap within the effective range; (3) or, based on coordinate positioning, priority is given to selecting the same geographical location in the sample image and the target image, and the number of samples selected from the sample image and the new image should be as equal as possible.

[0134] In some embodiments, if the number of data bands in the sample image and the target image is small and the quantization range is small, such as only the visible light red, green and blue bands, and the quantization is to the byte level, then the generative adversarial network ColorMapGAN, which only converts colors, is suitable; under other data conditions, the generative adversarial network DualGAN or the novel Diffusion network can be selected.

[0135] like Figure 12 As shown, Figure 12yes Figure 3 A flowchart of an implementation of step S102 is provided. In some embodiments, step S102 may include steps S501 to S506:

[0136] Step S501: Obtain the pre-trained prediction model;

[0137] In some embodiments, a pre-trained prediction model is used to output the probability of target annotation in the target image. The prediction model is a semantic segmentation model, which can be selected based on the characteristics of the data. For general data, the UNet model can be used; when the sample image set is large enough, a network structure with more parameters can be used, or the Transformer semantic segmentation algorithm can be chosen to construct the prediction model.

[0138] Step S502: Input the target image into the prediction model, and perform probability prediction on each pixel in the target image according to the prediction model to obtain the predicted probability value of each pixel.

[0139] In some embodiments, the target image is input into a pre-trained prediction model, which can predict the probability of each pixel in the target image and obtain the probability value of each pixel. Based on the probability value of each pixel, the predicted probability map of the target image can be determined.

[0140] Step S503: If the predicted probability value is within the preset probability range, determine the pixel as the region label of the target image.

[0141] In some embodiments, a probability range is preset, which indicates that when the predicted probability value of a pixel falls within the range, the pixel corresponds to a region label of the target image. In the embodiments of this application, the probability range is set to 0.46 to 0.5.

[0142] Step S504: Calculate the vegetation index of the target image, and determine the green space mask of the target image based on the vegetation index.

[0143] In some embodiments, it is also necessary to calculate the vegetation index in the target image and determine the green area mask of the target image based on the vegetation index, so as to identify the green area of ​​the target image and eliminate the interference of green vegetation on remote sensing in subsequent processing. For example, if the target image contains standard red and near-infrared bands, and the growth status of green vegetation differs significantly from that in the sample image due to seasonal factors, using the vegetation index alone (…) (This refers to the near-outer red band and the infrared band.) Based on the vegetation index, a green space mask is formed. A green space mask value of 1 indicates a green area, while a value of 0 indicates a non-green area. The green space extraction principle is to extract as accurately as possible, although some green areas may not be extracted. Step S505: Based on the area labeling and the green space mask, the predicted area is expanded, and the expanded predicted area is then used...

[0144] In some embodiments, if a green area is extracted, the probability value of determining the predicted area where the green area is located is 0.

[0145] In some embodiments, since the sample image does not perfectly correspond to the target image in reality, the predicted probability value obtained in the prediction model is usually low and the region is coarse and fragmented. Therefore, the standard conditional random field algorithm using binary potential is often used to process it to obtain a more spatially continuous probability map. Then, grayscale morphological "dilation" filtering is used to appropriately expand the region with a large probability value.

[0146] In some embodiments, when solving a conditional random field for a binary potential, only the probability is obtained, without obtaining the labeled result; if the pydensecrf library is used, only the probability matrix returned by the inference function is obtained; and the structuring element of the dilation filter uses a square structure with a side length between 3 and 9, which works well.

[0147] Step S506: Based on the predicted area, obtain the predicted probability map corresponding to the target image.

[0148] In some embodiments, since the predicted probability value of the predicted area corresponding to the green area is 0, it means that the area cannot correspond to the target object. Based on the predicted probability value output for each pixel, the area label corresponding to each pixel is obtained. For example, if the predicted probability value of pixel A being the top of a building is 0.48, then pixel A is determined to be the top of a building, i.e., it is a building area label; if the predicted probability value of pixel B being the top of a building is 0.2, then pixel B is determined not to be the top of a building, i.e., it is not a building area label. Multiple pixels connected together constitute the predicted probability map corresponding to the target image.

[0149] Understandably, first outputting a predicted probability map of the target image and then adjusting and correcting the data annotations can avoid the errors caused by simply migrating the original sample annotations and improve the accuracy of data annotation migration and correction for remote sensing images.

[0150] like Figure 13 As shown, Figure 13 This is another optional flowchart of the high-resolution remote sensing image annotation migration correction method provided in the embodiments of this application. Figure 13The method may include, but is not limited to, steps S601 to S603.

[0151] Step S601: Obtain a preset prediction training set, which includes multiple training sample images for training, and set corresponding training labels for the training sample images.

[0152] In some embodiments, the prediction model can be obtained from pre-training. Specifically, a prediction training set for training can be obtained and trained and adjusted using training labels of training sample images included in the prediction training set, wherein the training labels contain accurate verification results of the corresponding training sample images.

[0153] Step S602: Select any training sample image from the prediction training set and input it into the prediction model to obtain the training prediction probability value of each sample region of the training sample image.

[0154] In some embodiments, any training sample image is selected from the prediction training set and input into the prediction model to obtain the training prediction probability value of each sample region of the training sample image.

[0155] Step S603: Calculate the prediction loss value of the prediction model based on the training prediction probability value and training labels, and adjust the parameters of the prediction model according to the prediction loss value to obtain the trained prediction model.

[0156] In some embodiments, the obtained training prediction probability values ​​and training labels are compared, the difference between the two is calculated, which is the prediction loss value of the prediction model, and the parameters of the prediction model are adjusted according to the prediction loss value to obtain the trained prediction model. The trained prediction model can accurately predict the probability value of the region that needs to be labeled in the target image to a certain extent.

[0157] like Figure 14 As shown, Figure 14 yes Figure 3 Another implementation flowchart of step S103 in the figure, in some embodiments, step S103 may further include steps S701 to S404:

[0158] Step S701: Obtain geographic coordinate information;

[0159] In some embodiments, geographic coordinate information of sample images and target images is obtained, and this geographic coordinate information is used to map sample patches in the sample images to the predicted probability map of the target images.

[0160] Step S702: Based on the geographic coordinate information, the sample patches are output to the prediction probability map to obtain the constraint matching map;

[0161] In some embodiments, based on the obtained geographic coordinate information, sample patches of the sample image are output to the prediction probability map to obtain a constraint matching map. The sample patches in the constraint matching map roughly correspond to the area labels formed in the prediction probability map, but further movement corrections are needed for the sample patches to improve the accuracy of remote sensing image data label migration correction.

[0162] Step S703: Obtain the preset search radius, wherein the search radius is set according to the prediction probability map;

[0163] In some embodiments, since the region probability expansion is performed when generating the prediction probability map, the area of ​​the corresponding region label in the prediction probability map is larger than the area of ​​the sample patch. Therefore, the sample patch in the constraint matching map needs to be fine-tuned in the corresponding region label so that the final generated target label can accurately correspond to the target object.

[0164] In some embodiments, the sample patches need to be labeled and corrected within a preset search range, where the search range can be set according to the probability prediction map. For example, region label A in the probability prediction map corresponds to sample patch A, and the distance r between the top-left corner pixel of sample patch A and the top-left corner pixel of region label A is determined as the search radius.

[0165] Step S704: Based on the search radius, determine the matching region of the sample patch on the constraint matching map, wherein the matching region includes four sub-matching regions.

[0166] In some embodiments, the top left corner pixel of sample patch A is taken as the center pixel, and r is taken as the search radius. The matching area of ​​the sample patch on the constraint matching map is determined to be a matrix with a length and width of (2r+1) pixels.

[0167] For example, such as Figure 15 As shown, Figure 15 This is a schematic diagram of constraint matching for sample patches provided in an embodiment of this application. Taking the target object as the top of a building as an example, Top 1, Top 2, and Top 3 are the tops of three buildings in the constraint matching diagram. Among them, the matching area of ​​Top 3 is... Figure 15 A pixel matrix with a length and width of (2r+1) is formed. This (2r+1) pixel matrix is ​​divided into four blocks, each with at least r rows and r columns. The center pixel (the top left corner pixel) is assigned to any neighboring block, resulting in four sub-matching regions. Sample patches are then adjusted within these four sub-matching regions to improve the accuracy of the target image's corresponding annotation map.

[0168] like Figure 16 As shown, Figure 16 yes Figure 3A flowchart following step S105 is provided. In some embodiments, step S801 to S805 may also be included after step S105:

[0169] Step S801: Obtain the preset distance threshold;

[0170] In some embodiments, multiple candidate matching points can be identified within the matching region. However, two matching points that are too close may actually be the same matching point, but due to some error perturbation, they appear as duplicates. Therefore, a distance threshold is set, and when the distance between two matching points is less than the set distance threshold, one of them is filtered out to reduce errors.

[0171] Step S802: Determine the pixel point corresponding to the maximum similarity value of each sub-matching region, which is the sub-candidate matching point of each sub-matching region;

[0172] In some embodiments, such as Figure 15 As shown, each sample patch in the sample image may be irregular, which is not conducive to the formation of target annotations. Therefore, an outer rectangle can be used, with the area inside the outer rectangle representing the building top and the area outside the outer rectangle representing non-building tops. Based on this outer rectangle, the pixel at the top left corner is used as the reference pixel. The pixel similarity of each pixel in the matching area is calculated, and the sub-candidate matching points with the highest similarity values ​​in the four sub-matching areas are determined. Thus, four sub-candidate matching points are obtained.

[0173] In some embodiments, similarity can only be calculated using absolute difference sum, Euclidean distance or similar measures, and not using correlation coefficient or cosine similarity. This is because using measures such as cosine similarity would result in pixels with probabilities close to 0 being counted as highly similar pixels.

[0174] In some embodiments, the method for calculating the sum of absolute differences is as follows: Sum the mask pixels of the sample patches to determine the effective pixel value as 1. This represents the probability value corresponding to a valid pixel. The closer the probability value is to 1, the more likely the corresponding pixel is the top of the building.

[0175] In some embodiments, the similarity values ​​can be uniformly transformed to values ​​within the range [0,1], where a similarity value closer to 0 indicates less similarity, and a similarity value closer to 1 indicates greater similarity. Furthermore, when using the sum of absolute differences as a similarity measure, the normalized similarity can be calculated as: , where N is the number of valid pixels.

[0176] Step S803: Select the sub-candidate matching point with the highest similarity value among all sub-candidate matching points as the baseline matching point;

[0177] In some embodiments, the sub-candidate matching point with the highest similarity value is selected from four sub-candidate matching points as the baseline matching point.

[0178] In some embodiments, if there are multiple identical maximum similarity values, the one closest to the center point of the reference pixel of the sample patch is selected as the reference matching point.

[0179] Step S804: If the distance between the remaining sub-candidate matching points and the baseline matching point is less than the distance threshold, the sub-candidate matching points are determined as error matching points.

[0180] In some embodiments, if among the remaining three sub-candidate matching points, there is one whose distance to the benchmark matching point is less than a preset distance threshold, the sub-candidate matching point is determined to be an error matching point and is filtered out.

[0181] Step S805: Determine the candidate matching points corresponding to the sample patches based on the sub-candidate matching points and error matching points.

[0182] In some embodiments, candidate matching points for each sample patch are ultimately determined based on sub-candidate matching points whose distance from the benchmark matching point is greater than a distance threshold, and error matching points whose distance from the benchmark matching point is less than a distance threshold. Furthermore, it is understood that each sample patch will ultimately receive 1 to 4 candidate matching points, meaning each sample patch may correspond to one or more candidate matching points.

[0183] like Figure 17 As shown, Figure 17 yes Figure 16 A flowchart following step S805 is provided. In some embodiments, step S901 to S902 may also be included after step S805:

[0184] Step S901: If the distance between the first candidate matching point in the first matching region and the second candidate matching point in the second matching region is less than a preset distance threshold, determine that the first candidate matching point and the second candidate matching point are the same candidate matching point.

[0185] In some embodiments, when the distance between candidate points contained in each matching region is less than a preset distance threshold, the candidate points need to be merged.

[0186] Step S902, and determine that the sample patches corresponding to the same candidate matching point include the first sample patch corresponding to the first candidate matching point and the second sample patch corresponding to the second candidate matching point.

[0187] For example, the first candidate matching point of the first matching region corresponds to sample patch 1, and the first candidate matching point includes candidate point 1, candidate point 2 and candidate point 3. The second candidate matching point of the second matching region corresponds to sample patch 2, and the second candidate matching point includes candidate point 4 and candidate point 5. If the distance between candidate point 1 and candidate point 4 is less than a preset distance, then candidate point 1 and candidate point 4 are merged into the same candidate point. At this time, the sample patches corresponding to the same candidate point are sample patch 1 and sample patch 2. Thus, one candidate point may correspond to one or more sample patches.

[0188] like Figure 18 As shown, Figure 18 This is a flowchart of a candidate matching point merging method for migration correction of high-resolution remote sensing image annotation provided in this application embodiment. After sample patches are preprocessed, they can be initially located according to set parameters. Then, for each sample patch (single patch), a nearest neighbor search matching is performed within the selected search range, the similarity value of each pixel is calculated, and each similarity value is normalized. Next, the maximum and minimum values ​​of four sub-candidate matching regions are selected and filtered to obtain the candidate matching points for each single patch. After that, for each single patch, the close candidate matching points are merged to finally form a many-to-many relationship between single patches and candidate matching points. The overall constraint matching is performed using the maximum weight of a weighted bipartite graph, and the demolition and addition of buildings are judged. Finally, the target matching location is determined and output.

[0189] like Figure 19 As shown, Figure 19 yes Figure 3 A flowchart of an implementation of step S105 is provided. In some embodiments, step S105 may include steps S1001 to S1002:

[0190] Step S1001: Based on the candidate matching points corresponding to each sample patch and the sample patch corresponding to each candidate matching point, a weighted bipartite graph is determined with the similarity of each candidate matching point as the weight. The weighted bipartite graph is used to characterize the weight relationship between each sample patch and each candidate matching point.

[0191] In some embodiments, through steps S701 to S704, S801 to S805, and S901 to S902 described above, it can be determined that one sample patch can correspond to one or more candidate matching points, and one candidate matching point can correspond to one or more sample patches, thereby forming a many-to-many relationship between candidate matching points and sample patches. The similarity of each candidate matching point can be used as a weight to determine a weighted bipartite graph. Subsequently, virtual weights can be added based on the specific number of candidate matching points and sample patches to ensure that the number of sample patches and candidate matching points is consistent.

[0192] Step S1002: Using the overall maximum weighted matching as a constraint, determine the target matching position based on each sample patch and each candidate matching point from the weighted bipartite graph.

[0193] In some embodiments, based on the determined weighted bipartite graph, the maximum overall weight can be used as the matching condition to finally determine the correspondence between each sample patch and the candidate matching point, thereby obtaining the target matching location.

[0194] In some embodiments, since the sample patches are matched as a whole, when a single sample patch selects the one with the highest similarity value among its candidate matching points, the overall sample patch may not necessarily be in the optimal matching state. Therefore, based on the weight relationship between each sample patch and each candidate matching point, the maximum weighted matching result that enables each sample patch to correspond to the target object as a whole, rather than a local correspondence, is calculated when each sample patch corresponds to different candidate matching points, thereby achieving the migration correction of data annotations in remote sensing images.

[0195] For example, based on the weight matrix of a weighted bipartite graph, the Kuhn-Munkres algorithm is used to find the maximum weight matching of the bipartite graph. That is, under the condition of filling virtual weights and making the number of sample patches equal to the number of candidate points, each sample patch corresponds to a unique candidate matching point, and the overall weight is maximized. Furthermore, the virtual filling values ​​are removed from the output results. Some sample patches may not have corresponding candidate matching points, and some candidate matching points may not have corresponding sample patches.

[0196] Understandably, the sample annotations at the target matching location can correspond highly to the target object, achieving high-accuracy data annotation transfer and correction of remote sensing images.

[0197] like Figure 20 As shown, Figure 20 yes Figure 19 A flowchart of an implementation of step S1002 is provided. In some embodiments, step S1002 may include steps S1101 to S1104:

[0198] Step S1101: If the similarity value of the sample patch at the target matching location is not lower than the preset similarity threshold, the corresponding sample patch is obtained;

[0199] In some embodiments, a similarity threshold is preset, and samples with a similarity value higher than the target matching point are identified as corresponding samples of the target image.

[0200] Step S1102: If the sample patch has no target matching location, the deleted sample patch is obtained;

[0201] In some embodiments, if a sample patch has no target matching point, it is determined that the original building corresponding to the sample patch has been demolished, and thus a deleted sample patch is obtained.

[0202] Step S1103: If the similarity value of the sample patch is lower than the similarity threshold, the sample patch is deleted or a new sample patch is added.

[0203] In some embodiments, if the similarity value of a sample patch is lower than the similarity threshold, it is determined that the object corresponding to the sample patch is not the target object, and the sample patch is deleted; or if a building is added, a new sample patch is added.

[0204] Step S1104: Based on the corresponding sample patches, deleted sample patches, and newly added sample patches, the target annotation result is obtained. The target annotation result includes the target annotation and the certainty corresponding to the target annotation. The certainty is used to characterize the credibility measure between the target annotation and the actual annotation result.

[0205] In some embodiments, corresponding sample patches and newly added sample patches are retained, and sample patches are screened out and deleted to obtain target annotation results. The target annotation results are used to characterize the data annotation of the target object corresponding to the target image.

[0206] In some embodiments, the target annotation result also includes target annotations and their corresponding degrees of certainty. For example, the degree of certainty is determined by the weights between sample patches and candidate matching points in a weighted bipartite graph, and special values ​​are used to annotate the degree of certainty of newly added buildings. The overall output is then the target annotation result. The degree of certainty serves as an effective indicator of the reliability of sample patches during manual inspection, further improving the accuracy of the target annotation result.

[0207] like Figure 21 As shown, Figure 21 This is a schematic diagram of the functional modules for migration correction of high-resolution remote sensing image annotations provided in this application embodiment. This application embodiment also provides a migration correction system for high-resolution remote sensing image annotations, which can implement the above-mentioned migration correction method for high-resolution remote sensing image annotations. The high-resolution remote sensing image annotation migration correction system includes:

[0208] The acquisition module 1201 is used to acquire sample images and target images, wherein both the sample images and the target images include at least one target object, and the sample images are configured with annotation information of the target objects.

[0209] The probability prediction module 1202 is used to input the target image into a pre-trained prediction model to obtain the pixel probability value of each pixel in the target image as the target object, and to establish a prediction probability map of the target image based on the pixel probability value.

[0210] The matching module 1203 is used to determine the sample patch where the target object is located in the sample image based on the annotation information, and to determine the matching area of ​​the sample patch in the prediction probability map;

[0211] The calculation module 1204 is used to calculate the similarity value between at least one sample patch and each pixel in the matching region, and constrain the matching of each sample patch in the matching region based on the similarity value to obtain the target matching position, so that the constrained matching sample patch corresponds to the target object in the target image as a whole.

[0212] Result module 1205 is used to obtain the target annotation result of the target object in the target image based on the target matching position.

[0213] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described migration correction method for high-resolution remote sensing image annotation. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0214] like Figure 22 As shown, Figure 22 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device includes:

[0215] The processor 1301 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0216] The memory 1302 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1302 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1302 and is called and executed by the processor 1301 to execute the high-resolution remote sensing image annotation migration correction method of the embodiments of this application.

[0217] The input / output interface 1303 is used to implement information input and output;

[0218] The communication interface 1304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0219] Bus 1305 transmits information between various components of the device (e.g., processor 1301, memory 1302, input / output interface 1303, and communication interface 1304);

[0220] The processor 1301, memory 1302, input / output interface 1303 and communication interface 1304 are connected to each other within the device via bus 1305.

[0221] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described high-resolution remote sensing image annotation migration correction method.

[0222] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include 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, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0223] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0224] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0225] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0226] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0227] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0228] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0229] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0230] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0231] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0232] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0233] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A migration correction method for high-resolution remote sensing image annotation, characterized in that, The method includes: Acquire sample images and target images, wherein both the sample images and the target images include at least one target object, and the sample images are configured with annotation information of the target objects; The target image is input into a pre-trained prediction model to obtain the pixel probability value of each pixel in the target image as the target object, and a prediction probability map of the target image is established based on each pixel probability value. The target object is located in the sample patch in the sample image based on the annotation information, and the matching region of the sample patch is determined in the prediction probability map. The sample patch is then corrected and adjusted in the corresponding matching region. Calculate the similarity value between at least one of the sample patches and each pixel in the matching region, and determine a weighted bipartite graph based on the candidate matching points corresponding to each of the sample patches and the sample patches corresponding to each candidate matching point, using the similarity of each candidate matching point as the weight. The weighted bipartite graph is used to characterize the weight relationship between each of the sample patches and each candidate matching point. Using the overall maximum weighted matching as a constraint, the target matching position based on each sample patch and each of the candidate matching points is determined from the weighted bipartite image, so that each sample patch after constraint matching corresponds to the target object in the target image as a whole. Based on the target matching location, the target annotation result of the target object in the target image is obtained.

2. The high-resolution remote sensing image annotation migration correction method according to claim 1, characterized in that, Based on the target annotation results of the target image, target patches are obtained, each target patch including target patch feature values ​​and target patch annotation information corresponding to the target patch feature values; sample patches include sample patch feature values ​​and sample patch annotation information corresponding to the sample patch feature values. After obtaining the target annotation result of the target object in the target image based on the target matching position, the method further includes: Obtain the weight parameter values ​​used for augmentation; The difference between the weight parameter value and the unit weight is calculated to obtain the difference weight parameter value. Multiply the weight parameter value by the sample patch feature value to obtain a first sub-value; multiply the difference weight parameter value by the target patch feature value to obtain a second sub-value; add the first sub-value and the second sub-value to obtain the first augmented information. Multiply the weight parameter value by the sample patch annotation information to obtain the third sub-value; multiply the difference weight parameter value by the target patch annotation information to obtain the fourth sub-value; add the third sub-value and the fourth sub-value to obtain the second augmented information. Based on the first augmented information and the second augmented information, an augmented sample with augmentation and improvement is obtained, and the augmented sample is used to iteratively improve the target annotation result.

3. The high-resolution remote sensing image annotation migration correction method according to claim 1, characterized in that, Determining the sample patch where the target object is located in the sample image based on the annotation information includes: Obtain the positioning reference points and resolution parameters of the target image; Based on the positioning reference point and the resolution parameter, the sample image is adjusted in terms of positioning and resolution so that the coordinate position and resolution of the adjusted sample image and the target image are consistent, thus obtaining the preprocessed sample image; Based on the image features of the target image, feature transformation is performed on the preprocessed sample image to obtain the transformed sample image, and the annotation information of the transformed sample image is obtained based on the transformed sample image. Based on the converted annotation information, the sample patch in the sample image where the target object is located is determined.

4. The high-resolution remote sensing image annotation migration correction method according to claim 3, characterized in that, The step of performing feature transformation on the preprocessed sample image based on the image features of the target image to obtain the transformed sample image includes: Obtain the first spectral information of the preprocessed sample image and the second spectral information of the target image; If the first spectral information and the second spectral information are within a preset spectral range, a relative radiometric correction operation is performed on the preprocessed sample image to obtain the converted sample image; If the first spectral information and the second spectral information are not within the preset spectral range, a linear quantization operation is performed on the preprocessed sample image so that the linearly quantized first spectral information and the second spectral information are within the preset spectral range, thereby obtaining the converted sample image. Alternatively, a style transfer operation can be performed on the pre-processed sample image based on a pre-trained network model to obtain the transformed sample image.

5. The high-resolution remote sensing image annotation migration correction method according to claim 1, characterized in that, The step of inputting the target image into a pre-trained prediction model to obtain the pixel probability value of each pixel in the target image as the target object, and establishing a prediction probability map of the target image based on each pixel probability value, includes: Obtain a pre-trained prediction model; The target image is input into the prediction model, and the probability of each pixel in the target image is predicted according to the prediction model to obtain the predicted probability value of each pixel. If the predicted probability value is within a preset probability range, the pixel is determined to be a region label of the target image; Calculate the vegetation index of the target image, and determine the green space mask of the target image based on the vegetation index; Based on the area label and the green space mask, the predicted area is expanded, and the expanded predicted area is used as a basis for calculation. Based on the predicted region, a predicted probability map corresponding to the target image is obtained.

6. The migration correction method for high-resolution remote sensing image annotation according to claim 1, characterized in that, The pre-trained prediction model is trained through the following steps: Obtain a preset prediction training set, which includes multiple training sample images for training, and set corresponding training labels for the training sample images. Select any one of the training sample images from the prediction training set and input it into the prediction model to obtain the training prediction probability value of each sample region of the training sample image; Based on the training prediction probability value and the training label, the prediction loss value of the prediction model is calculated, and the parameters of the prediction model are adjusted according to the prediction loss value to obtain the trained prediction model.

7. The migration correction method for high-resolution remote sensing image annotation according to claim 1, characterized in that, Determining the matching region of the sample patch in the predicted probability map includes: Obtain geographic coordinate information; Based on the geographic coordinate information, the sample patches are output to the predicted probability map to obtain the constraint matching map; Obtain a preset search radius, wherein the search radius is set according to the predicted probability map; Based on the search radius, the matching region of the sample patch on the constraint matching map is determined, wherein the matching region includes four sub-matching regions.

8. The migration correction method for high-resolution remote sensing image annotation according to claim 7, characterized in that, After calculating the similarity value between at least one of the sample patches and each pixel in the matching region, the method further includes: Obtain the preset distance threshold; The pixel point corresponding to the maximum similarity value of each sub-matching region is determined as the sub-candidate matching point of each sub-matching region; Select the sub-candidate matching point with the largest similarity value among all the sub-candidate matching points as the benchmark matching point; If the distance between the remaining sub-candidate matching point and the benchmark matching point is less than the distance threshold, the sub-candidate matching point is determined as an error matching point; Based on the sub-candidate matching points and the error matching points, the candidate matching points corresponding to the sample patches are determined.

9. The migration correction method for high-resolution remote sensing image annotation according to claim 8, characterized in that, After determining the candidate matching points corresponding to the sample patch based on the sub-candidate matching points and the error matching points, the method further includes: If the distance between the first candidate matching point in the first matching region and the second candidate matching point in the second matching region is less than the preset distance threshold, the first candidate matching point and the second candidate matching point are determined to be the same candidate matching point. The sample patches corresponding to the same candidate matching point are determined to include the first sample patch corresponding to the first candidate matching point and the second sample patch corresponding to the second candidate matching point.

10. The high-resolution remote sensing image annotation migration correction method according to claim 1, characterized in that, Determining the target matching location from the weighted bipartite graph based on each sample patch and each of the candidate matching points includes: If the similarity value of the sample patch at the target matching location is not lower than a preset similarity threshold, the corresponding sample patch is obtained; If the sample patch has no target matching location, the sample patch is deleted; If the similarity value of the sample patch is lower than the similarity threshold, the sample patch is either deleted or a new sample patch is added. The target annotation result is obtained based on the corresponding sample patch, the deleted sample patch, and the newly added sample patch. The target annotation result includes the target annotation and the certainty corresponding to the target annotation. The certainty is used to characterize the credibility measure between the target annotation and the actual annotation result.

11. A migration correction system for high-resolution remote sensing image annotation, characterized in that, The system includes: An acquisition module is used to acquire sample images and target images, wherein both the sample images and the target images include at least one target object, and the sample images are configured with annotation information of the target objects; The probability prediction module is used to input the target image into a pre-trained prediction model to obtain the pixel probability value of each pixel in the target image as the target object, and to establish a prediction probability map of the target image based on each pixel probability value. The matching module is used to determine the sample patch where the target object is located in the sample image based on the annotation information, and to determine the matching region of the sample patch in the prediction probability map, wherein the sample patch is corrected and adjusted in the corresponding matching region; The calculation module is used to calculate the similarity value between at least one of the sample patches and each pixel in the matching region, and to determine a weighted bipartite graph based on the candidate matching points corresponding to each of the sample patches and the sample patches corresponding to each candidate matching point, using the similarity of each candidate matching point as the weight. The weighted bipartite graph is used to characterize the weight relationship between each of the sample patches and each candidate matching point. Using the overall maximum weighted matching as a constraint, the target matching position based on each sample patch and each of the candidate matching points is determined from the weighted bipartite image, so that each sample patch after constraint matching corresponds to the target object in the target image as a whole. The result module is used to obtain the target annotation result of the target object in the target image based on the target matching position.

12. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the migration correction method for high-resolution remote sensing image annotation as described in any one of claims 1 to 10.

13. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the migration correction method for high-resolution remote sensing image annotation as described in any one of claims 1 to 10.

Images