A method and device for preparing samples of remote sensing image ground feature elements

Abstract

The present invention discloses a method and device for making samples of remote sensing image ground object elements. The method includes: pre-training a backbone network based on a large open source data set to obtain the pre-training weights of the model; combining positive and negative click coding data with sample images and inputting them into the backbone network to extract feature images; training a segmentation subnetwork based on the annotation information of feature images and sample images to complete a model training process; processing the training data set using the trained model to output a predicted mask; automatically generating and updating simulated differential positive and negative click coding data based on the real sample annotation mask and the predicted mask; iteratively training the model N times to obtain a target model; inputting the image data to be annotated into the model, combining interactive click coding, and obtaining high-precision annotated sample data. The present invention can reduce the cost of producing deep learning sample data and improve the problem of lack of sample label data for semantic segmentation of remote sensing images.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a method and a device for preparing samples of remote sensing image ground object elements. Background Art

[0002] In recent years, with the rapid development of earth observation technology, multi-source high-resolution satellite remote sensing image data has increased significantly, providing rich data resources for fast and effective object recognition, while also putting forward higher demands on image processing capabilities.

[0003] Since remote sensing images have special spectral and spatial characteristics, general image segmentation algorithms are difficult to achieve ideal results in remote sensing data processing. As one of the three basic tasks in the field of computer vision, semantic segmentation realizes the classification and positioning of targets at the pixel level, and has been widely used in medical image analysis, unmanned vehicle driving, robotics and other fields. With the continuous improvement of the model, relying on the ability of convolutional neural networks in automatic learning and feature extraction, the advantages of deep learning semantic segmentation technology in the field of remote sensing continue to emerge, and it has played an important role in the intelligent remote sensing information extraction of land features such as buildings, roads, and water bodies.

[0004] The combination of semantic segmentation and remote sensing tasks provides a new opportunity for the rapid and accurate acquisition of surface feature information. However, the parameters of deep learning models driven by data models and based on large samples are huge in scale. From model construction to parameter optimization, they are closely related to data, and are more dependent on the quality and quantity of samples than traditional methods. Although there are many ways to obtain data in the field of remote sensing, data processing and labeling often require a lot of manpower and material resources. How to quickly achieve large-scale sample production is one of the urgent problems to be solved in remote sensing target extraction.

[0005] At present, the interactive method of assisting deep learning sample production has received more and more attention, and has also achieved initial success in the application of target segmentation in natural images. However, remote sensing images are essentially different from natural images. On the one hand, the acquisition of remote sensing images is seriously affected by factors such as lighting and weather; on the other hand, there are more types of targets and more complex backgrounds in remote sensing images. Therefore, the model trained by natural image datasets is difficult to be directly applied to remote sensing images. How to effectively solve the problem of producing sample label data in high-resolution remote sensing images is a crucial factor in promoting the further development of deep learning semantic segmentation in the field of remote sensing. Summary of the invention

[0006] The technical problem solved by the present invention is to overcome the deficiencies of the prior art and provide a method for preparing samples of ground object elements in remote sensing images.

[0007] The technical solution of the present invention is:

[0008] In a first aspect, an embodiment of the present invention provides a method for preparing samples of ground object elements in remote sensing images, comprising:

[0009] Preprocess the acquired high-resolution remote sensing images to obtain a training data set;

[0010] Pre-train the backbone network of the model based on a large open source dataset to obtain the pre-trained weights of the model;

[0011] Combined with the annotation information of the training sample images, positive and negative click data are randomly generated and encoded to obtain positive and negative click encoding data;

[0012] Combining the positive and negative click coding data with sample images in the training data set, and inputting the data into the pre-trained backbone network to extract feature images;

[0013] The segmentation sub-network of the model is trained based on the feature image and the sample image annotation information in the training data set to obtain a model that has been trained once;

[0014] Processing the training data set based on the trained model and outputting a predicted mask;

[0015] Automatically generate simulated differential positive and negative click encoding data according to the real sample annotation information of the training data set and the predicted mask;

[0016] Based on the differential positive and negative click coding data, the original positive and negative click coding data are updated, and the model trained once is iteratively trained N times to obtain a trained target model; N is a positive integer;

[0017] The image data to be annotated is processed based on the target model and combined with interactive click coding to obtain high-precision annotated sample data.

[0018] Optionally, the preprocessing of the acquired high-resolution remote sensing images to obtain a training data set includes:

[0019] Acquire high-resolution remote sensing images;

[0020] Preprocessing the remote sensing image to generate RGB three-band image slices;

[0021] A real sample annotation mask is produced according to the image slices of the RGB three-bands to obtain the training data set.

[0022] Optionally, the step of producing a real sample annotation mask according to the RGB three-band image slices to obtain the training data set includes:

[0023] Draw the vector boundary of each ground object target in the RGB three-band image slice, perform category labeling and raster binarization processing on it, and use the obtained mask raster file as a training labeling sample.

[0024] Optionally, the randomly generating positive and negative click data and encoding them to obtain positive and negative click coded data includes:

[0025] Generate positive and negative click data based on random sampling;

[0026] The positive and negative click data are encoded based on a spatial encoding method to generate the positive and negative click encoding data.

[0027] Optionally, the processing of the image data to be annotated based on the target model and combining interactive click coding to obtain high-precision annotated sample data includes:

[0028] Inputting a preset number of image data to be annotated into the target model;

[0029] Processing the image data to be annotated based on the target model to obtain sample image data with mask annotation;

[0030] Based on an interactive click method, positive and negative click encoding is performed on the sample image with mask annotation, and regional mask correction is automatically performed to obtain a high-precision sample annotation data set.

[0031] In a second aspect, an embodiment of the present invention provides a sample preparation device for remote sensing image ground object elements, comprising:

[0032] The training data set acquisition module is used to pre-process the acquired high-resolution remote sensing images to obtain the training data set;

[0033] The pre-trained model weight acquisition module is used to pre-train the backbone network of the model based on a large open source data set to obtain the pre-trained weights of the model;

[0034] A positive and negative coding data acquisition module is used to randomly generate positive and negative click data and encode them in combination with the annotation information of the training sample image to obtain positive and negative click coding data;

[0035] A feature image extraction module is used to combine the positive and negative click coding data with the sample images in the training data set and input them into the pre-trained backbone network to extract feature images;

[0036] A single training model acquisition module is used to input the segmentation sub-network of the model for training based on the feature image and the sample image annotation information in the training data set to obtain a model that has been trained once;

[0037] A predicted mask output module, used to process the training data set based on the trained model and output a predicted mask;

[0038] A differential coding data generation module, used to automatically generate simulated differential positive and negative click coding data according to the real sample annotation information of the training data set and the predicted mask;

[0039] A target model acquisition module, used to update the original positive and negative click coding data based on the differential positive and negative click coding data, iteratively train the model trained once N times, and obtain a trained target model; N is a positive integer;

[0040] The sample annotation data acquisition module is used to process the image data to be annotated based on the target model and obtain high-precision annotated sample data in combination with interactive click coding.

[0041] Optionally, the training data set acquisition module includes:

[0042] A remote sensing image acquisition unit, used for acquiring high-resolution remote sensing images;

[0043] An image slice generating unit is used to pre-process the remote sensing image to generate RGB three-band image slices;

[0044] The training data set acquisition unit is used to produce a real sample annotation mask according to the image slices of the RGB three-bands to obtain the training data set.

[0045] Optionally, the training data set acquisition unit includes:

[0046] The training annotation sample acquisition subunit is used to draw the vector boundary of each ground object target in the RGB three-band image slice, perform category annotation and raster binarization processing on it, and use the obtained mask raster file as the training annotation sample.

[0047] Optionally, the positive and negative coding data acquisition module includes:

[0048] A positive and negative click data generating unit, used for generating positive and negative click data based on a random sampling method;

[0049] The positive and negative click coding data generating unit is used to encode the positive and negative click data based on a spatial encoding method to generate the positive and negative click coding data.

[0050] Optionally, the model training sample acquisition module includes:

[0051] A sample image data input unit, used to input a preset number of sample image data to be labeled into the target model;

[0052] A labelled image data preliminary acquisition unit, used for processing the sample image data to be labeled based on the target model to obtain sample image data with mask labelling;

[0053] The sample annotation information optimization unit is used to perform positive and negative click encoding on the sample image with mask annotation based on an interactive click method, and automatically perform regional mask correction to obtain a high-precision sample annotation data set.

[0054] The advantages of the present invention compared with the prior art are:

[0055] The sample preparation scheme for remote sensing image ground object elements provided by the embodiment of the present invention can reduce the cost of deep learning data production, effectively improve the problem of lack of sample label data for semantic segmentation of remote sensing images, and can be applied to various types of ground objects, and has good adaptability to domestic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart of the steps of a method for preparing samples of ground object elements in remote sensing images provided by an embodiment of the present invention;

[0057] Figure 2 A flow chart of a method for preparing samples of remote sensing image ground object elements based on iterative training and interactive click segmentation model provided by an embodiment of the present invention;

[0058] Figure 3 A schematic diagram of the structure of a device for preparing samples of ground features in remote sensing images provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0059] Embodiment 1

[0060] Reference Figure 1 , which shows a flow chart of the steps of a method for preparing samples of remote sensing image ground object elements provided by an embodiment of the present invention, such as Figure 1 As shown, the sample preparation method may include the following steps:

[0061] Step 101: Preprocess the acquired high-resolution remote sensing image to obtain a training data set.

[0062] For the sample production process, it can be combined Figure 2 In an embodiment of the present invention, when obtaining batch model training samples, high-resolution remote sensing images can be first obtained, and a training data set can be produced based on preprocessed image slices. Specifically, it can be described in detail in combination with the following specific implementation method.

[0063] In a specific implementation of the embodiment of the present invention, the above step 101 may include:

[0064] Sub-step A1: Acquire high-resolution remote sensing images;

[0065] Sub-step A2: pre-processing the remote sensing image to generate RGB three-band image slices;

[0066] Sub-step A3: Create a real sample annotation mask based on the RGB three-band image slices to obtain the training data set.

[0067] In an embodiment of the present invention, a high-resolution remote sensing image refers to a remote sensing image with a resolution higher than a set resolution threshold, and the specific value of the set resolution threshold can be determined according to business needs, and this embodiment does not limit this. To obtain a high-resolution remote sensing image, optical satellite remote sensing data with visible light-near infrared sensors, such as the United States' WorldView-3, the domestic Gaofen-2, etc., can be used, or aerial remote sensing data collected by a general optical camera, such as drone data, can be used. Afterwards, these data are preprocessed, including geometric correction, fusion of full-color and multi-spectral images, band selection, etc., to meet the model's requirements for space and spectrum. In addition, in order to improve the training efficiency of the model, it is also necessary to select an area of ​​interest with a typical target in the image and crop it according to a uniform size and format. Finally, a series of RGB three-band image slices that can be directly input into the model are generated. Then, ArcGIS or other tools are used to make samples of deep learning semantic segmentation, draw the vector boundaries of each ground object target, and label it by category. The generated vector label data needs to be rasterized, and the corresponding raster mask file can be used as a training annotation sample.

[0068] After the training data set is obtained, step 102 is executed.

[0069] Step 102: Pre-train the backbone network of the model based on a large open source dataset to obtain the pre-training weights of the model.

[0070] In this embodiment, the backbone network of the model can be pre-trained based on a large open source data set to obtain the pre-trained weights of the model. Specifically, the neural network structure of the present invention is basically similar to the semantic segmentation network, including a backbone network, a segmentation subnetwork and a detection head. Among them, the backbone network is a deep convolutional neural network, mainly composed of a convolution layer, a pooling layer and a fully connected layer. The number of output channels of the convolution layer is consistent with the number of channels of its convolution kernel. Therefore, the convolution operation can not only be used to obtain the feature information of the image, but also to perform data space dimension transformation, such as transforming 3-band RGB image data into 64 channels. In semantic segmentation, common backbone networks include ResNet series, EfficientNet series, etc. Before training, it is usually necessary to pre-train the backbone network to accelerate the training process. The common pre-training data set in natural image target detection is ImageNet. In the present invention, we use the open source data set UC MercedLand-Use Data Set, which contains a total of 21 types of remote sensing images of scenes, which can be used for pre-training of the backbone network.

[0071] Step 103: Combined with the annotation information of the training sample image, randomly generate positive and negative click data and encode them to obtain positive and negative click encoded data.

[0072] In the initial training, the present invention combines the sample annotation mask information of the image and uses random sampling to automatically generate a certain number of positive and negative clicks. The positive and negative clicks are represented in the image in the form of coordinate points and need to meet certain conditions:

[0073] a、N pos positive points are generated in the mask area of ​​the target object and satisfy: 1) The distance between any two positive points is greater than the preset threshold d step , 2) The minimum distance between any positive point and the mask boundary point is greater than the preset threshold d margin .

[0074] b、N neg The negative points surround and cover the outer boundary of the mask of the target object, and the distance between each negative click is as large as possible.

[0075] In order to input positive and negative clicks into the convolutional network, they need to be encoded in spatial form first. Encoding clicks by distance transformation is the most common way. However, when adding or moving a point, the impact of this method is global, and the huge change of the encoding map will cause confusion in the network. Therefore, the present invention uses a disk with a certain radius to represent positive and negative clicks, so that when adding or moving a point, the change of the encoding map is very small.

[0076] Step 104: Combine the positive and negative click coding data with the sample images in the training data set and input them into the pre-trained backbone network to extract feature images.

[0077] After obtaining the positive and negative click coding data, the positive and negative click coding data can be combined with the sample images in the training data set and input into the pre-trained backbone network to extract the feature images.

[0078] In practical applications, the input of the semantic segmentation model is usually RGB three-band images. In order to introduce positive and negative click encoding data, it is necessary to solve the problem of the data size of the additional input in the backbone network.

[0079] In the present invention, a convolutional block is introduced, and the output tensor size is exactly the same as the first convolutional block in the backbone network. Then, this output tensor is summed pixel by pixel with the output data of the sample image in the first convolutional layer of the backbone network. Usually, these output tensor data are 64 channels.

[0080] Step 105: input the segmentation sub-network of the model for training based on the feature image and the sample image annotation information in the training data set to obtain a trained model.

[0081] After the feature image is extracted, the segmentation subnetwork of the model can be trained based on the feature image and the annotation information of the sample image in the training data set to obtain a trained model. Specifically, a series of hyperparameters for model training can be set according to the structure of the segmentation subnetwork, including learning rate, learning rate update coefficient and update step, number of iterations, batch size, weight decay coefficient, etc. In addition, it is necessary to select Adam or SGD as the optimizer for training. After the acquired feature image is input into the segmentation subnetwork, a predicted mask can be initially generated at the head of the network. After that, the real sample annotation mask is compared with the predicted mask, and the prediction optimization is performed based on the loss function. When the loss function reaches a convergence state, the model training process is completed.

[0082] Step 106: Process the training data set based on the trained model and output a predicted mask.

[0083] After obtaining a trained model, the training data set can be processed based on this model to output the predicted mask. Specifically, the image in the training data set can be used as the data to be inferred and input into the model trained in the previous step, which can output the predicted mask and category confidence corresponding to the image.

[0084] Step 107: Automatically generate simulated differential positive and negative click encoding data based on the real sample annotation information of the training data set and the predicted mask.

[0085] In order to evaluate and analyze the predicted mask, it is compared with the real sample annotation information of the training data set. Obviously, the initial predicted mask will have missed detection and false detection. Based on this, the positive and negative click encoding data can be generated by iterative sampling. The main methods are as follows:

[0086] a. Compare the predicted mask with the mask annotated by the real sample to obtain the erroneous part of the prediction result (including missed detection and false detection).

[0087] b. Taking pixels as units, use connected components to divide the error parts into different clusters.

[0088] c. Apply morphological erosion to each cluster to obtain a new difference region. Determine a largest cluster based on the number of difference pixels, and generate a click inside the new difference region of this cluster. If a click has occurred in this cluster before, in order to avoid multiple clicks being too close to each other, the point collected this time needs to have the largest Euclidean distance with the boundary of the cluster after erosion and the previous click; if there was no click before, then the center point of the difference region after erosion of this cluster can be determined as the click position collected this time.

[0089] Step 108: Based on the differential positive and negative click coding data, the original positive and negative click coding data is updated, and the model that has been trained once is iteratively trained N times to obtain a trained target model, where N is a positive integer.

[0090] During the iterative training process, the output segmentation mask and click coding data can provide additional prior information to improve the quality of prediction. In the present invention, the two-channel point data of positive and negative click coding and the predicted binary mask data are first stacked. Here, when no inference is performed, the mask data can be empty. Afterwards, a convolution block is introduced so that the size of its output tensor is exactly the same as the first convolution block in the backbone network. Then the pixel-by-pixel summation is performed and input into the backbone network. At this point, the backbone network no longer needs to be pre-trained, and training can continue directly on the basis of the completion of the previous iteration. The model training process is executed in a loop, and the predicted mask and positive and negative click coding are continuously updated until the iterative training reaches a preset number of times N, and the obtained model can be output and saved.

[0091] Step 109: Process the image data to be annotated based on the target model and combine it with interactive click coding to obtain high-precision annotated sample data.

[0092] According to the above sample image production method, the image to be annotated is processed to generate image slices of appropriate size. The image slices are input into the segmentation model after iterative training to obtain the mask data with target annotation.

[0093] In order to improve the quality of the sample, the interactive tool can be used to post-process the mask data, that is, by setting positive and negative clicks, the model can automatically determine the error mode of the result and make corrections. For example, if a positive click is set in a predetermined way in the target area that the model missed, the model will automatically correct it to the target and update the mask. If a negative click is set in the area that the model misdetected, the mask can also be deleted in the area. Compared with the traditional sample production method that relies entirely on manual delineation, this method has the advantages of being fast, accurate, and efficient.

[0094] Embodiment 2

[0095] Reference Figure 3 , which shows a schematic diagram of the structure of a sample preparation device for remote sensing image ground object elements provided by an embodiment of the present invention, such as Figure 3 As shown, the sample preparation device may include the following modules:

[0096] The training data set acquisition module 310 is used to pre-process the acquired high-resolution remote sensing images to obtain a training data set;

[0097] A pre-trained model weight acquisition module 320 is used to pre-train the backbone network of the model based on a large open source data set to obtain the pre-trained weights of the model;

[0098] The positive and negative coding data acquisition module 330 is used to randomly generate positive and negative click data and encode them in combination with the annotation information of the training sample image to obtain positive and negative click coding data;

[0099] A feature image extraction module 340 is used to combine the positive and negative click coding data with the sample images in the training data set and input them into the pre-trained backbone network to extract feature images;

[0100] A single training model acquisition module 350 is used to input the segmentation sub-network of the model for training based on the feature image and the sample image annotation information in the training data set to obtain a model that has been trained once;

[0101] A predicted mask output module 360, configured to process the training data set based on the trained model and output a predicted mask;

[0102] A differential encoding data generation module 370, for automatically generating simulated differential positive and negative click encoding data according to the real sample annotation information of the training data set and the predicted mask;

[0103] The target model acquisition module 380 is used to update the original positive and negative click coding data based on the differential positive and negative click coding data, and iteratively train the model trained once N times to obtain a trained target model; N is a positive integer;

[0104] The sample annotation data acquisition module 390 is used to process the image data to be annotated based on the target model and obtain high-precision annotated sample data in combination with interactive click coding.

[0105] Optionally, the training data set acquisition module includes:

[0106] A remote sensing image acquisition unit, used for acquiring high-resolution remote sensing images;

[0107] An image slice generating unit is used to pre-process the remote sensing image to generate RGB three-band image slices;

[0108] The training data set acquisition unit is used to produce a real sample annotation mask according to the image slices of the RGB three-bands to obtain the training data set.

[0109] Optionally, the training data set acquisition unit includes:

[0110] The training annotation sample acquisition subunit is used to draw the vector boundary of each ground object target in the RGB three-band image slice, perform category annotation and raster binarization processing on it, and use the obtained mask raster file as the training annotation sample.

[0111] Optionally, the positive and negative coding data acquisition module includes:

[0112] A positive and negative click data generating unit, used for generating positive and negative click data based on a random sampling method;

[0113] The positive and negative click coding data generating unit is used to encode the positive and negative click data based on a spatial encoding method to generate the positive and negative click coding data.

[0114] Optionally, the model training sample acquisition module includes:

[0115] A sample image data input unit, used to input a preset number of sample image data to be labeled into the target model;

[0116] A labelled image data preliminary acquisition unit, used for processing the sample image data to be labeled based on the target model to obtain sample image data with mask labelling;

[0117] The sample annotation information optimization unit is used to perform positive and negative click encoding on the sample image with mask annotation based on an interactive click method, and automatically perform regional mask correction to obtain a high-precision sample annotation data set.

[0118] The specific implementation methods described in this application can enable those skilled in the art to understand this application more comprehensively, but do not limit this application in any way. Therefore, those skilled in the art should understand that this application can still be modified or replaced by equivalents; and all technical solutions and improvements that do not deviate from the spirit and technical essence of this application should be included in the protection scope of the patent of this application.

[0119] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.

Claims

1. A method for preparing samples of ground features of remote sensing images, characterized in that: include: Preprocess the acquired high-resolution remote sensing images to obtain a training data set; Pre-train the backbone network of the model based on a large open source dataset to obtain the pre-trained weights of the model; Combined with the annotation information of the training sample images, positive and negative click data are randomly generated and encoded to obtain positive and negative click encoded data; Combining the positive and negative click coding data with sample images in the training data set, and inputting the data into the pre-trained backbone network to extract feature images; The segmentation sub-network of the model is trained based on the feature image and the sample image annotation information in the training data set to obtain a model that has been trained once; Processing the training data set based on the trained model and outputting a predicted mask; Automatically generate simulated differential positive and negative click encoding data according to the real sample annotation information of the training data set and the predicted mask; Based on the differential positive and negative click coding data, the original positive and negative click coding data are updated, and the model trained once is iteratively trained N times to obtain a trained target model; N is a positive integer; The image data to be annotated is processed based on the target model and combined with interactive click coding to obtain high-precision annotated sample data.

2. The method according to claim 1, characterized in that The obtained high-resolution remote sensing images are preprocessed to obtain a training data set, including: Acquire high-resolution remote sensing images; Preprocessing the remote sensing image to generate RGB three-band image slices; A real sample annotation mask is produced according to the image slices of the RGB three-bands to obtain the training data set.

3. The method according to claim 2, characterized in that The step of producing a real sample annotation mask according to the image slices of the RGB three-bands to obtain the training data set includes: Draw the vector boundary of each ground object target in the RGB three-band image slice, perform category labeling and raster binarization processing on it, and use the obtained mask raster file as a training labeling sample.

4. The method according to claim 1, characterized in that: The randomly generated positive and negative click data and encoded thereof to obtain positive and negative click encoded data include: Generate positive and negative click data based on random sampling; The positive and negative click data are encoded based on a spatial encoding method to generate the positive and negative click encoding data.

5. The method according to claim 1, characterized in that: The image data to be annotated is processed based on the target model, and combined with interactive click coding to obtain high-precision annotated sample data, including: Inputting a preset number of image data to be annotated into the target model; Processing the image data to be annotated based on the target model to obtain sample image data with mask annotation; Based on an interactive click method, positive and negative click encoding is performed on the sample image with mask annotation, and regional mask correction is automatically performed to obtain a high-precision sample annotation data set.

6. A device for preparing samples of ground features in remote sensing images, characterized in that: include: The training data set acquisition module is used to pre-process the acquired high-resolution remote sensing images to obtain the training data set; The pre-trained model weight acquisition module is used to pre-train the backbone network of the model based on a large open source data set to obtain the pre-trained weights of the model; A positive and negative coding data acquisition module is used to randomly generate positive and negative click data and encode them in combination with the annotation information of the training sample image to obtain positive and negative click coding data; A feature image extraction module is used to combine the positive and negative click coding data with the sample images in the training data set and input them into the pre-trained backbone network to extract feature images; A single training model acquisition module is used to input the segmentation sub-network of the model for training based on the feature image and the sample image annotation information in the training data set to obtain a model that has been trained once; A predicted mask output module, used to process the training data set based on the trained model and output a predicted mask; A differential coding data generation module, used to automatically generate simulated differential positive and negative click coding data according to the real sample annotation information of the training data set and the predicted mask; A target model acquisition module, used to update the original positive and negative click coding data based on the differential positive and negative click coding data, iteratively train the model trained once N times, and obtain a trained target model; N is a positive integer; The sample annotation data acquisition module is used to process the image data to be annotated based on the target model and obtain high-precision annotated sample data in combination with interactive click coding.

7. The device according to claim 6, characterized in that The training data set acquisition module includes: A remote sensing image acquisition unit, used for acquiring high-resolution remote sensing images; An image slice generating unit is used to pre-process the remote sensing image to generate RGB three-band image slices; The training data set acquisition unit is used to produce a real sample annotation mask according to the image slices of the RGB three-bands to obtain the training data set.

8. The device according to claim 7, characterized in that The training data set acquisition unit comprises: The training annotation sample acquisition subunit is used to draw the vector boundary of each ground object target in the RGB three-band image slice, perform category annotation and raster binarization processing on it, and use the obtained mask raster file as the training annotation sample.

9. The device according to claim 6, characterized in that The positive and negative coding data acquisition module includes: A positive and negative click data generating unit, used for generating positive and negative click data based on a random sampling method; The positive and negative click coding data generating unit is used to encode the positive and negative click data based on a spatial encoding method to generate the positive and negative click coding data.

10. The device according to claim 6, characterized in that The sample annotation data acquisition module includes: A sample image data input unit, used to input a preset number of sample image data to be labeled into the target model; A labelled image data preliminary acquisition unit, used for processing the sample image data to be labeled based on the target model to obtain sample image data with mask labelling; The sample annotation information optimization unit is used to perform positive and negative click encoding on the sample image with mask annotation based on an interactive click method, and automatically perform regional mask correction to obtain a high-precision sample annotation data set.

Images