A remote sensing image recognition method and system based on deep learning

Abstract

The present invention is applicable to the technical field of remote sensing image recognition, and provides a remote sensing image recognition method and system based on deep learning. The present invention screens multiple valid sample images; standardizes and stages the images; creates multiple stage generators and multiple stage discriminators, performs dynamic adversarial training, generates multiple extended sample images; performs training and optimization, generates a target recognition model; and performs image recognition. It is capable of effectively screening, standardizing and stage-by-stage processing of multiple remote sensing sample images, creating multiple stage generators and multiple stage discriminators, performing dynamic adversarial training, generating multiple extended sample images, and then training and optimizing the deep learning model to generate the target recognition model, thereby achieving effective expansion of samples, so that the deep learning model has a large number of samples for training and optimization, effectively improving the remote sensing image recognition performance of the target recognition model, and being able to meet the high requirements of accuracy and reliability in application scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image recognition, and in particular relates to a remote sensing image recognition method and system based on deep learning. Background Art

[0002] Remote sensing image recognition is the process of using computer vision and image processing technology to analyze and interpret image data obtained by remote sensing. These images usually contain spectral information of the earth's surface or other targets. By analyzing this information, we can identify the type of land objects, monitor environmental changes, and manage resources. It has a wide range of applications in agriculture, forestry, urban planning, disaster monitoring and other fields.

[0003] In the existing technology, in some cases of sample scarcity, deep learning algorithms find it difficult to fully capture and learn the complex features and patterns in remote sensing images, which restricts the learning efficiency and generalization ability of the model. Therefore, the deep learning models constructed under such circumstances often lead to a significant decline in remote sensing image recognition performance, making it difficult to meet the high requirements for accuracy and reliability in some remote sensing image recognition application scenarios. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a remote sensing image recognition method and system based on deep learning, aiming to solve the technical problems existing in the prior art mentioned in the background technology.

[0005] The embodiment of the present invention is implemented as follows:

[0006] A remote sensing image recognition method based on deep learning, the method specifically comprises the following steps:

[0007] Acquire a plurality of remote sensing sample images, perform source identification and sensitivity identification on the plurality of remote sensing sample images, and screen a plurality of valid sample images from the plurality of remote sensing sample images;

[0008] According to the preset staged requirements, the plurality of valid sample images are subjected to standardization processing and staged processing to generate a plurality of stage sample images;

[0009] According to the preset stage requirements, multiple stage generators and multiple stage discriminators are created, and dynamic adversarial training is performed on multiple stage sample images to generate multiple expanded sample images;

[0010] Select a deep learning model, train and optimize it through the multiple valid sample images and the multiple expanded sample images, and generate and deploy a target recognition model;

[0011] Obtain a target recognition image, and perform image recognition processing using the target recognition model.

[0012] As a further limitation of the technical solution of the embodiment of the present invention, the acquiring of multiple remote sensing sample images, performing source identification and sensitivity identification on the multiple remote sensing sample images, and screening multiple valid sample images from the multiple remote sensing sample images specifically include the following steps:

[0013] Acquire multiple remote sensing sample images;

[0014] Acquiring image source information and image position data of a plurality of remote sensing sample images;

[0015] According to a preset standard source address, source matching is performed on the plurality of image source information, and a plurality of standard source images are matched from the plurality of remote sensing sample images;

[0016] According to the preset sensitive position data, sensitive identification is performed on the plurality of image position data, a plurality of sensitive sample images are eliminated from the plurality of standard source images, and a plurality of valid sample images are screened.

[0017] As a further limitation of the technical solution of the embodiment of the present invention, the step of performing standardization and stage processing on the plurality of valid sample images according to the preset stage requirements to generate a plurality of stage sample images specifically comprises the following steps:

[0018] Determine standardized sizes and multiple stage resolutions according to preset stage requirements;

[0019] According to the standardized size, a plurality of the valid sample images are subjected to standardized cropping processing to generate a plurality of standard sample images;

[0020] According to the plurality of stage resolutions, the plurality of standard sample images are subjected to stage resolution adjustment to generate a plurality of stage sample images.

[0021] As a further limitation of the technical solution of the embodiment of the present invention, the step of creating multiple stage generators and multiple stage discriminators according to the preset stage requirements, performing dynamic adversarial training on multiple stage sample images, and generating multiple expanded sample images specifically includes the following steps:

[0022] According to the preset phased requirements, determine the number of phases, phase relationships and phase constraints;

[0023] According to the number of stages, the relationship between the stages and the stage constraints, a plurality of stage generators and stage discriminators corresponding to the combinations are created;

[0024] Through multiple stage generators and multiple stage discriminators, dynamic adversarial training is performed on multiple stage sample images to generate multiple expanded sample images.

[0025] As a further limitation of the technical solution of the embodiment of the present invention, the loss function of the multiple stage generators is:

[0026] ;

[0027] in, Representative stage, For the The loss function of the stage generator of the stage is, For the The prior noise distribution A random variable sampled from is a random variable The corresponding generated samples are To generate samples The true judgment probability of For the preset The stage constraints of each stage, Constraints for the stage Next pair generation sample The probability of true judgment;

[0028] The loss function of the multiple stage discriminators is:

[0029] ;

[0030] in, For the The loss function of the stage discriminator of the stage is, For the The real data distribution The real samples sampled from For real samples The true judgment probability of Constraints for the stage Next, the real sample The probability of true judgment.

[0031] As a further limitation of the technical solution of the embodiment of the present invention, the selecting of the deep learning model, training and optimizing through the multiple valid sample images and the multiple expanded sample images, and generating and deploying the target recognition model specifically include the following steps:

[0032] Select a deep learning model;

[0033] According to a preset sample division ratio, the plurality of valid sample images and the plurality of expanded sample images are divided into a training set and a test set;

[0034] The deep learning model is trained and optimized by using the training set and the test set to generate a target recognition model;

[0035] Deploy the object recognition model.

[0036] As a further limitation of the technical solution of the embodiment of the present invention, the acquiring of the target recognition image and performing image recognition processing through the target recognition model specifically include the following steps:

[0037] Acquire target recognition images;

[0038] Determining a target resolution from among the plurality of stage resolutions;

[0039] According to the standardized size and the target resolution, the target recognition image is adjusted to generate a plurality of adjusted recognition images;

[0040] The target recognition model is used to perform image recognition processing on the plurality of adjusted recognition images.

[0041] A remote sensing image recognition system based on deep learning, the system comprises an effective sample screening module, an image stage processing module, a sample expansion processing module, a model training optimization module and an image recognition processing module, wherein:

[0042] An effective sample screening module is used to obtain a plurality of remote sensing sample images, perform source identification and sensitivity identification on the plurality of remote sensing sample images, and screen a plurality of effective sample images from the plurality of remote sensing sample images;

[0043] An image stage processing module, used to perform standardization and stage processing on the plurality of valid sample images according to preset stage requirements, and generate a plurality of stage sample images;

[0044] A sample expansion processing module is used to create multiple stage generators and multiple stage discriminators according to preset stage requirements, perform dynamic adversarial training on multiple stage sample images, and generate multiple expanded sample images;

[0045] A model training and optimization module, used to select a deep learning model, train and optimize it through a plurality of the valid sample images and a plurality of the expanded sample images, and generate and deploy a target recognition model;

[0046] The image recognition processing module is used to obtain the target recognition image and perform image recognition processing through the target recognition model.

[0047] As a further limitation of the technical solution of the embodiment of the present invention, the effective sample screening module specifically includes:

[0048] A sample acquisition unit, used for acquiring a plurality of remote sensing sample images;

[0049] An information acquisition unit, used to acquire image source information and image position data of a plurality of remote sensing sample images;

[0050] A source matching unit, used for performing source matching on the plurality of image source information according to a preset standard source address, and matching a plurality of standard source images from the plurality of remote sensing sample images;

[0051] The sensitive elimination unit is used to perform sensitive identification on the plurality of image position data according to preset sensitive position data, eliminate a plurality of sensitive sample images from the plurality of standard source images, and screen a plurality of valid sample images.

[0052] As a further limitation of the technical solution of the embodiment of the present invention, the image stage processing module specifically includes:

[0053] A demand analysis unit, used to determine standardized sizes and multiple stage resolutions according to preset staged requirements;

[0054] A standardized cutting unit, used for performing standardized cutting processing on the plurality of valid sample images according to the standardized size to generate a plurality of standard sample images;

[0055] The resolution stage adjustment unit is used to perform resolution stage adjustment on the plurality of standard sample images according to the plurality of stage resolutions to generate a plurality of stage sample images.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The embodiment of the present invention screens multiple valid sample images; standardizes and stages the images; creates multiple stage generators and multiple stage discriminators, performs dynamic adversarial training, generates multiple extended sample images; performs training and optimization, generates a target recognition model; and performs image recognition. It is capable of effectively screening, standardizing and stage-by-stage processing of multiple remote sensing sample images, creating multiple stage generators and multiple stage discriminators, performing dynamic adversarial training, generating multiple extended sample images, and then training and optimizing a deep learning model to generate a target recognition model, thereby achieving effective expansion of samples, enabling the deep learning model to have a large number of samples for training and optimization, effectively improving the remote sensing image recognition performance of the target recognition model, and being able to meet the high requirements of accuracy and reliability in application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flowchart of a remote sensing image recognition method based on deep learning provided by an embodiment of the present invention is shown;

[0059] Figure 2 A flow chart of source identification and sensitive identification in the method provided by an embodiment of the present invention is shown;

[0060] Figure 3 A flowchart of the standardization and phased processing in the method provided by the embodiment of the present invention is shown;

[0061] Figure 4 A flow chart showing the generation of multiple expanded sample images in the method provided by an embodiment of the present invention;

[0062] Figure 5 A flow chart of generating a target recognition model in a method provided by an embodiment of the present invention is shown;

[0063] Figure 6 A flowchart of image recognition processing in the method provided by an embodiment of the present invention is shown;

[0064] Figure 7 An application architecture diagram of a remote sensing image recognition system based on deep learning provided by an embodiment of the present invention is shown;

[0065] Figure 8 It shows a structural block diagram of a valid sample screening module in a system provided by an embodiment of the present invention;

[0066] Fig. 9 The structure block diagram of the image stage processing module in the system provided by the embodiment of the present invention is shown. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0068] It is understandable that in the existing technology, in some cases of sample scarcity, it is difficult for deep learning algorithms to fully capture and learn the complex features and patterns in remote sensing images, which restricts the learning efficiency and generalization ability of the model. Therefore, the deep learning model constructed under such circumstances often leads to a significant decline in remote sensing image recognition performance, which makes it difficult to meet the high requirements for accuracy and reliability in some remote sensing image recognition application scenarios.

[0069] To solve the above problems, an embodiment of the present invention discloses a remote sensing image recognition method and system based on deep learning, which obtains multiple remote sensing sample images, performs source identification and sensitivity identification on the multiple remote sensing sample images, and screens multiple valid sample images from the multiple remote sensing sample images; according to preset staged requirements, the multiple valid sample images are standardized and staged to generate multiple stage sample images; according to preset staged requirements, multiple stage generators and multiple stage discriminators are created, and dynamic adversarial training is performed on the multiple stage sample images to generate multiple extended sample images; a deep learning model is selected, and training and optimization are performed through multiple valid sample images and multiple extended sample images to generate and deploy a target recognition model; a target recognition image is obtained, and image recognition processing is performed through the target recognition model. It can effectively screen, standardize and stage multiple remote sensing sample images, create multiple stage generators and multiple stage discriminators, perform dynamic adversarial training, generate multiple expanded sample images, and then train and optimize the deep learning model to generate a target recognition model, thereby achieving effective expansion of samples. This allows the deep learning model to have a large number of samples for training and optimization, effectively improving the remote sensing image recognition performance of the target recognition model and meeting the high requirements for accuracy and reliability in application scenarios.

[0070] Specifically, Figure 1 A flowchart of a remote sensing image recognition method based on deep learning provided by an embodiment of the present invention is shown.

[0071] In a preferred embodiment of the present invention, a remote sensing image recognition method based on deep learning comprises the following steps:

[0072] Step S101: Acquire a plurality of remote sensing sample images, perform source identification and sensitivity identification on the plurality of remote sensing sample images, and select a plurality of valid sample images from the plurality of remote sensing sample images.

[0073] In an embodiment of the present invention, by acquiring multiple remote sensing sample images, and acquiring image source information and image position data corresponding to the multiple remote sensing sample images, source matching is performed on the multiple image source information according to a preset standard source address, remote sensing sample images that do not belong to the standard source address are eliminated, and the remote sensing sample images that belong to the standard source address are marked as standard source images, so that multiple standard source images can be matched from multiple remote sensing sample images, and then sensitive identification is performed on the multiple image position data according to preset sensitive position data, standard source images belonging to sensitive positions are eliminated, and standard source images that do not belong to sensitive positions are marked as valid sample images, so that multiple valid sample images can be screened from multiple standard source images.

[0074] It can be understood that by performing source matching, the remote sensing sample images can be screened for the first time, so that the multiple standard source images screened out meet the standard source addresses, which can ensure that the standard source images are obtained through official and authoritative channels. Specifically, the official and authoritative channels include: Digital Earth Open Platform, Resource and Environmental Science and Data Center of the Chinese Academy of Sciences, Natural Resources Satellite Remote Sensing Cloud Service Platform, China Remote Sensing Data Sharing Network, China Resources Satellite Center, etc.

[0075] It can be understood that by performing sensitive identification, remote sensing images related to sensitive areas can be eliminated from multiple standard source images, mainly involving the elimination of remote sensing images in areas such as national security and privacy protection, such as: military management areas, important infrastructure, personal privacy areas, etc.

[0076] Specifically, Figure 2 A flow chart of source identification and sensitive identification in the method provided by an embodiment of the present invention is shown.

[0077] Among them, in the preferred embodiment provided by the present invention, the obtaining of multiple remote sensing sample images, performing source identification and sensitivity identification on the multiple remote sensing sample images, and screening multiple valid sample images from the multiple remote sensing sample images specifically include the following steps:

[0078] Step S1011, obtaining a plurality of remote sensing sample images.

[0079] Step S1012: Obtain image source information and image position data of a plurality of remote sensing sample images.

[0080] Step S1013: performing source matching on the plurality of image source information according to a preset standard source address, and matching a plurality of standard source images from the plurality of remote sensing sample images.

[0081] Step S1014: According to the preset sensitive position data, sensitive identification is performed on the plurality of image position data, and a plurality of sensitive sample images are eliminated from the plurality of standard source images to screen a plurality of valid sample images.

[0082] Furthermore, the remote sensing image recognition method based on deep learning also includes the following steps:

[0083] Step S102: performing standardization processing and stage processing on the plurality of valid sample images according to preset stage requirements to generate a plurality of stage sample images.

[0084] In an embodiment of the present invention, a standardized size and multiple stage resolutions are determined according to preset stage requirements, and then multiple valid sample images are standardized and cropped according to the standardized size to generate multiple standard sample images that meet the standardized size. Then, according to the multiple stage resolutions, the multiple standard sample images are adjusted in resolution stage to generate multiple stage sample images.

[0085] It can be understood that, in the standardized cropping process of multiple valid sample images, different standardized cropping processes are performed according to the original image sizes of different valid sample images. Specifically: if the original image size of the valid sample image is an integer multiple of the standardized size, the valid sample image is cropped without overlapping, and an integer multiple of standard sample images are obtained by cropping, and there are no identical image areas between the integer multiple of standard sample images; if the original image size of the valid sample image is not an integer multiple of the standardized size, the valid sample image is cropped with partial overlap, and some image areas are identical between the multiple standard sample images obtained by cropping.

[0086] It can be understood that performing stage-by-stage resolution adjustment on multiple standard sample images means performing multiple stage resolution adjustments on each standard sample image, and the process of stage-by-stage resolution adjustment is a process of gradually adjusting from high resolution to low resolution. For example, when the multiple stage resolutions include “50×50”, “100×100”, “200×200”, and “400×400”, the resolution of each standard sample image is gradually adjusted to “400×400”, “200×200”, “100×100”, and “50×50”.

[0087] Specifically, Figure 3 A flow chart of the standardization and phased processing in the method provided by the embodiment of the present invention is shown.

[0088] Among them, in the preferred embodiment provided by the present invention, the standardization and stage processing of the plurality of valid sample images according to the preset stage requirements to generate a plurality of stage sample images specifically includes the following steps:

[0089] Step S1021: Determine the standardized size and multiple stage resolutions according to the preset stage requirements.

[0090] Step S1022: Perform standardized cropping processing on the multiple valid sample images according to the standardized size to generate multiple standard sample images.

[0091] Step S1023: According to the multiple stage resolutions, the multiple standard sample images are subjected to stage resolution adjustment to generate multiple stage sample images.

[0092] Furthermore, the remote sensing image recognition method based on deep learning also includes the following steps:

[0093] Step S103: according to the preset stage requirements, create multiple stage generators and multiple stage discriminators, perform dynamic adversarial training on multiple stage sample images, and generate multiple expanded sample images.

[0094] In an embodiment of the present invention, according to the preset stage requirements, the number of stages, the relationship between the stages and the constraints of the stages are determined, and then according to the number of stages, the relationship between the stages and the constraints of the stages, a plurality of stage generators and stage discriminators corresponding to the combinations are created, and then through the plurality of stage generators and the plurality of stage discriminators, in the order of the stage resolution from low to high, the plurality of stage sample images are dynamically adversarially trained, so that the generated images can be similar to the features of the plurality of stage sample images, and the resolution of the generated images is gradually improved, the details of the images are supplemented, and a plurality of expanded sample images are generated. Specifically, the loss function of the plurality of stage generators is:

[0095] ;

[0096] in, Representative stage, For the The loss function of the stage generator of the stage is, For the The prior noise distribution A random variable sampled from is a random variable The corresponding generated samples are To generate samples The true judgment probability of For the preset The stage constraints of each stage, Constraints for the stage Next pair generation sample The probability of true judgment;

[0097] The loss function of the multiple stage discriminators is:

[0098] ;

[0099] in, For the The loss function of the stage discriminator of the stage is, For the The real data distribution The real samples sampled from For real samples The true judgment probability of Constraints for the stage Next, the real sample The probability of true judgment.

[0100] It can be understood that the stage constraints are constraints on image resolution at different stages.

[0101] Understandably, It is The phases are from the preset prior noise distribution The random variable sampled from is The input of the stage discriminator of the stage.

[0102] Understandably, It is The real data distribution In the embodiment of the present invention, the real data distribution , are sample images of multiple stages.

[0103] Understandably, is a random variable The corresponding generated samples, when the dynamic adversarial training is completed, the generated samples with the highest resolution are the expanded sample images.

[0104] Specifically, Figure 4 A flow chart of generating multiple expanded sample images in the method provided by an embodiment of the present invention is shown.

[0105] Among them, in the preferred embodiment provided by the present invention, the creation of multiple stage generators and multiple stage discriminators according to the preset stage requirements, dynamic adversarial training of multiple stage sample images, and generation of multiple expanded sample images specifically include the following steps:

[0106] Step S1031: Determine the number of stages, stage relationships and stage constraints according to the preset stage requirements.

[0107] Step S1032: Create a plurality of combinations of corresponding stage generators and stage discriminators according to the number of stages, the stage relationships and the stage constraints.

[0108] Step S1033: Perform dynamic adversarial training on the multiple stage sample images through the multiple stage generators and the multiple stage discriminators to generate multiple expanded sample images.

[0109] Furthermore, the remote sensing image recognition method based on deep learning also includes the following steps:

[0110] Step S104: select a deep learning model, perform training and optimization through multiple valid sample images and multiple expanded sample images, and generate and deploy a target recognition model.

[0111] In an embodiment of the present invention, by selecting a deep learning model, multiple valid sample images and multiple expanded sample images are divided and processed according to a preset sample division ratio, and the multiple valid sample images and multiple expanded sample images are randomly divided into a training set and a test set. The deep learning model is trained through the training set, and then the performance of the model is evaluated and tuned through the test set. The target recognition model is generated and exported to a deployable format. In the application scenario of remote sensing image recognition, the target recognition model is deployed to the corresponding hardware platform, such as a server, embedded device, etc.

[0112] It is understandable that the deep learning models that can be selected include convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), and Transformer models, etc.

[0113] It is understandable that in performance evaluation and tuning, evaluation indicators include accuracy, recall, F1 score, etc.

[0114] Specifically, Figure 5 A flow chart of generating a target recognition model in the method provided in an embodiment of the present invention is shown.

[0115] Among them, in the preferred embodiment provided by the present invention, the selection of the deep learning model, training and optimization through the multiple valid sample images and the multiple expanded sample images, and generating and deploying the target recognition model specifically include the following steps:

[0116] Step S1041: Select a deep learning model.

[0117] Step S1042: Divide the plurality of valid sample images and the plurality of expanded sample images into a training set and a test set according to a preset sample division ratio.

[0118] Step S1043: train and optimize the deep learning model using the training set and the test set to generate a target recognition model.

[0119] Step S1044: deploy the target recognition model.

[0120] Furthermore, the remote sensing image recognition method based on deep learning also includes the following steps:

[0121] Step S105: Acquire a target recognition image, and perform image recognition processing using the target recognition model.

[0122] In an embodiment of the present invention, a target recognition image is acquired, and then the largest stage resolution is marked as the target resolution from multiple stage resolutions, and then the target recognition image is cropped according to a standardized size to obtain multiple standard recognition images that all meet the standardized size, and then the resolution of the multiple standard recognition images is adjusted according to the target resolution to generate multiple adjusted recognition images, which are imported into the deployed target recognition model, and image recognition processing is performed on the multiple adjusted recognition images through the target recognition model.

[0123] It can be understood that the image resolution of multiple standard recognition images is consistent with the target recognition image and is not lower than the target resolution. During the resolution adjustment process, if the image resolution is the same as the target resolution, no resolution adjustment is required; if the image resolution is higher than the target resolution, the resolution of multiple standard recognition images is reduced according to the target resolution.

[0124] Specifically, Figure 6 A flow chart of image recognition processing in the method provided by an embodiment of the present invention is shown.

[0125] Among them, in the preferred embodiment provided by the present invention, the acquiring of the target recognition image and performing image recognition processing through the target recognition model specifically include the following steps:

[0126] Step S1051, obtaining a target recognition image.

[0127] Step S1052: Determine a target resolution from the multiple stage resolutions.

[0128] Step S1053: adjusting the target recognition image according to the standardized size and the target resolution to generate a plurality of adjusted recognition images.

[0129] Step S1054: performing image recognition processing on the plurality of adjusted recognition images through the target recognition model.

[0130] Furthermore, Figure 7 An application architecture diagram of a remote sensing image recognition system based on deep learning provided by an embodiment of the present invention is shown.

[0131] Among them, in another preferred embodiment provided by the present invention, a remote sensing image recognition system based on deep learning includes:

[0132] The effective sample screening module 101 is used to obtain a plurality of remote sensing sample images, perform source identification and sensitivity identification on the plurality of remote sensing sample images, and screen a plurality of effective sample images from the plurality of remote sensing sample images.

[0133] In an embodiment of the present invention, the effective sample screening module 101 obtains multiple remote sensing sample images, and obtains image source information and image position data corresponding to the multiple remote sensing sample images, and performs source matching on the multiple image source information according to a preset standard source address, eliminates remote sensing sample images that do not belong to the standard source address, and marks the remote sensing sample images that belong to the standard source address as standard source images, so that multiple standard source images can be matched from multiple remote sensing sample images, and then sensitive identification is performed on multiple image position data according to preset sensitive position data, standard source images belonging to sensitive positions are eliminated, and standard source images that do not belong to sensitive positions are marked as effective sample images, so that multiple effective sample images can be screened from multiple standard source images.

[0134] Specifically, Figure 8 It shows a structural block diagram of the effective sample screening module 101 in the system provided by an embodiment of the present invention.

[0135] Among them, in the preferred embodiment provided by the present invention, the effective sample screening module 101 specifically includes:

[0136] The sample acquisition unit 1011 is used to acquire a plurality of remote sensing sample images.

[0137] The information acquisition unit 1012 is used to acquire the image source information and image position data of the plurality of remote sensing sample images.

[0138] The source matching unit 1013 is used to perform source matching on the plurality of image source information according to a preset standard source address, and match a plurality of standard source images from the plurality of remote sensing sample images.

[0139] The sensitive elimination unit 1014 is used to perform sensitive identification on the plurality of image position data according to preset sensitive position data, eliminate a plurality of sensitive sample images from the plurality of standard source images, and screen a plurality of valid sample images.

[0140] Furthermore, the remote sensing image recognition system based on deep learning also includes:

[0141] The image stage processing module 102 is used to perform standardization processing and stage processing on the multiple valid sample images according to the preset stage requirements to generate multiple stage sample images.

[0142] In an embodiment of the present invention, the image stage processing module 102 determines the standardized size and multiple stage resolutions according to the preset stage requirements, and then performs standardized cropping processing on multiple valid sample images according to the standardized size to generate multiple standard sample images that meet the standardized size, and then performs resolution stage adjustment on the multiple standard sample images according to the multiple stage resolutions to generate multiple stage sample images.

[0143] Specifically, Fig. 9 It shows a structural block diagram of the image stage processing module 102 in the system provided by an embodiment of the present invention.

[0144] Among them, in the preferred embodiment provided by the present invention, the image stage processing module 102 specifically includes:

[0145] The demand analysis unit 1021 is used to determine the standardized size and multiple stage resolutions according to the preset staged demand.

[0146] The standardized cropping unit 1022 is used to perform standardized cropping processing on the multiple valid sample images according to the standardized size to generate multiple standard sample images.

[0147] The resolution stage adjustment unit 1023 is used to perform resolution stage adjustment on the plurality of standard sample images according to the plurality of stage resolutions to generate a plurality of stage sample images.

[0148] Furthermore, the remote sensing image recognition system based on deep learning also includes:

[0149] The sample expansion processing module 103 is used to create multiple stage generators and multiple stage discriminators according to preset stage requirements, perform dynamic adversarial training on multiple stage sample images, and generate multiple expanded sample images.

[0150] In the embodiment of the present invention, the sample expansion processing module 103 determines the number of stages, the relationship between the stages and the constraints of the stages according to the preset stage requirements, and then creates a plurality of stage generators and stage discriminators corresponding to the combinations according to the number of stages, the relationship between the stages and the constraints of the stages, and then performs dynamic adversarial training on the plurality of stage sample images in the order of low to high stage resolution through the plurality of stage generators and the plurality of stage discriminators, so that the generated images can be similar to the features of the plurality of stage sample images, and gradually improve the resolution of the generated images, supplement the details of the images, and generate a plurality of expanded sample images. Specifically, the loss function of the plurality of stage generators is:

[0151] ;

[0152] in, Representative stage, For the The loss function of the stage generator of the stage is, For the The prior noise distribution A random variable sampled from is a random variable The corresponding generated samples are To generate samples The true judgment probability of For the preset The stage constraints of each stage, Constraints for the stage Next pair generation sample The probability of true judgment;

[0153] The loss function of the multiple stage discriminators is:

[0154] ;

[0155] in, For the The loss function of the stage discriminator of the stage is, For the The real data distribution The real samples sampled from For real samples The true judgment probability of Constraints for the stage Next, the real sample The probability of true judgment.

[0156] The model training optimization module 104 is used to select a deep learning model, perform training and optimization through multiple valid sample images and multiple expanded sample images, and generate and deploy a target recognition model.

[0157] In an embodiment of the present invention, the model training optimization module 104 selects a deep learning model, divides and processes multiple valid sample images and multiple expanded sample images according to a preset sample division ratio, and randomly divides the multiple valid sample images and multiple expanded sample images into a training set and a test set. The deep learning model is trained through the training set, and the performance of the model is evaluated and tuned through the test set. The target recognition model is generated and exported to a deployable format. In the application scenario of remote sensing image recognition, the target recognition model is deployed to the corresponding hardware platform, such as a server, an embedded device, etc.

[0158] The image recognition processing module 105 is used to obtain a target recognition image and perform image recognition processing through the target recognition model.

[0159] In an embodiment of the present invention, the image recognition processing module 105 obtains a target recognition image, and then marks the largest stage resolution as the target resolution from multiple stage resolutions, and then crops the target recognition image according to a standardized size to obtain multiple standard recognition images that all meet the standardized size, and then adjusts the resolution of the multiple standard recognition images according to the target resolution to generate multiple adjusted recognition images, which are imported into the deployed target recognition model, and image recognition processing is performed on the multiple adjusted recognition images through the target recognition model.

[0160] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0161] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0162] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A remote sensing image recognition method based on deep learning, characterized in that: The method specifically comprises the following steps: Acquire a plurality of remote sensing sample images, perform source identification and sensitivity identification on the plurality of remote sensing sample images, and screen a plurality of valid sample images from the plurality of remote sensing sample images; According to the preset staged requirements, the plurality of valid sample images are subjected to standardization processing and staged processing to generate a plurality of stage sample images; According to the preset stage requirements, multiple stage generators and multiple stage discriminators are created, and dynamic adversarial training is performed on multiple stage sample images to generate multiple expanded sample images; Select a deep learning model, train and optimize it through the multiple valid sample images and the multiple expanded sample images, and generate and deploy a target recognition model; Acquire a target recognition image, and perform image recognition processing through the target recognition model; The method of creating multiple stage generators and multiple stage discriminators according to the preset stage requirements, performing dynamic adversarial training on multiple stage sample images, and generating multiple expanded sample images specifically includes the following steps: According to the preset phased requirements, determine the number of phases, phase relationships and phase constraints; According to the number of stages, the relationship between the stages and the stage constraints, a plurality of stage generators and stage discriminators corresponding to the combinations are created; Performing dynamic adversarial training on the plurality of stage sample images through the plurality of stage generators and the plurality of stage discriminators to generate a plurality of expanded sample images; The loss function of the generators in the above stages is: ； in, Representative stage, For the The loss function of the stage generator of the stage is, For the The prior noise distribution A random variable sampled from is a random variable The corresponding generated samples are To generate samples The true judgment probability of For the preset The stage constraints of each stage, Constraints for the stage Next pair generation sample The probability of true judgment; The loss function of the multiple stage discriminators is: ； in, For the The loss function of the stage discriminator of the stage is, For the The real data distribution The real samples sampled from For real samples The true judgment probability of Constraints for the stage Next, the real sample The probability of true judgment.

2. The remote sensing image recognition method based on deep learning according to claim 1, characterized in that: The step of acquiring a plurality of remote sensing sample images, performing source identification and sensitivity identification on the plurality of remote sensing sample images, and screening a plurality of valid sample images from the plurality of remote sensing sample images specifically comprises the following steps: Acquire multiple remote sensing sample images; Acquiring image source information and image position data of a plurality of remote sensing sample images; According to a preset standard source address, source matching is performed on the plurality of image source information, and a plurality of standard source images are matched from the plurality of remote sensing sample images; According to the preset sensitive position data, sensitive identification is performed on the plurality of image position data, a plurality of sensitive sample images are eliminated from the plurality of standard source images, and a plurality of valid sample images are screened.

3. The remote sensing image recognition method based on deep learning according to claim 1, characterized in that: The step of performing standardization and staged processing on the plurality of valid sample images according to the preset staged requirements to generate a plurality of staged sample images specifically comprises the following steps: Determine standardized sizes and multiple stage resolutions according to preset stage requirements; According to the standardized size, a plurality of the valid sample images are subjected to standardized cropping processing to generate a plurality of standard sample images; According to the plurality of stage resolutions, the plurality of standard sample images are subjected to stage resolution adjustment to generate a plurality of stage sample images.

4. The remote sensing image recognition method based on deep learning according to claim 1, characterized in that: The selecting of the deep learning model, training and optimizing through the multiple valid sample images and the multiple expanded sample images, and generating and deploying the target recognition model specifically include the following steps: Select a deep learning model; According to a preset sample division ratio, the plurality of valid sample images and the plurality of expanded sample images are divided into a training set and a test set; The deep learning model is trained and optimized by using the training set and the test set to generate a target recognition model; Deploy the object recognition model.

5. The remote sensing image recognition method based on deep learning according to claim 3, characterized in that: The acquiring of the target recognition image and performing image recognition processing through the target recognition model specifically comprises the following steps: Acquire target recognition images; Determining a target resolution from among the plurality of stage resolutions; According to the standardized size and the target resolution, the target recognition image is adjusted to generate a plurality of adjusted recognition images; The target recognition model is used to perform image recognition processing on the plurality of adjusted recognition images.

6. A remote sensing image recognition system based on deep learning, characterized in that: The system includes an effective sample screening module, an image stage processing module, a sample expansion processing module, a model training optimization module and an image recognition processing module, wherein: An effective sample screening module is used to obtain a plurality of remote sensing sample images, perform source identification and sensitivity identification on the plurality of remote sensing sample images, and screen a plurality of effective sample images from the plurality of remote sensing sample images; An image stage processing module, used to perform standardization and stage processing on the plurality of valid sample images according to preset stage requirements, and generate a plurality of stage sample images; A sample expansion processing module is used to create multiple stage generators and multiple stage discriminators according to preset stage requirements, perform dynamic adversarial training on multiple stage sample images, and generate multiple expanded sample images; A model training and optimization module, used to select a deep learning model, train and optimize it through a plurality of the valid sample images and a plurality of the expanded sample images, and generate and deploy a target recognition model; An image recognition processing module is used to obtain a target recognition image and perform image recognition processing through the target recognition model; The method of creating multiple stage generators and multiple stage discriminators according to the preset stage requirements, performing dynamic adversarial training on multiple stage sample images, and generating multiple expanded sample images is specifically as follows: According to the preset phased requirements, determine the number of phases, phase relationships and phase constraints; According to the number of stages, the relationship between the stages and the stage constraints, a plurality of stage generators and stage discriminators corresponding to the combinations are created; Performing dynamic adversarial training on the plurality of stage sample images through the plurality of stage generators and the plurality of stage discriminators to generate a plurality of expanded sample images; The loss function of the generators in the above stages is: ； in, Representative stage, For the The loss function of the stage generator of the stage is, For the The prior noise distribution A random variable sampled from is a random variable The corresponding generated samples are To generate samples The true judgment probability of For the preset The stage constraints of each stage, Constraints for the stage Next pair generation sample The probability of true judgment; The loss function of the multiple stage discriminators is: ； in, For the The loss function of the stage discriminator of the stage is, For the The real data distribution The real samples sampled from For real samples The true judgment probability of Constraints for the stage Next, the real sample The probability of true judgment.

7. The remote sensing image recognition system based on deep learning according to claim 6, characterized in that: The effective sample screening module specifically includes: A sample acquisition unit, used for acquiring a plurality of remote sensing sample images; An information acquisition unit, used to acquire image source information and image position data of a plurality of remote sensing sample images; A source matching unit, used for performing source matching on the plurality of image source information according to a preset standard source address, and matching a plurality of standard source images from the plurality of remote sensing sample images; The sensitive elimination unit is used to perform sensitive identification on the plurality of image position data according to preset sensitive position data, eliminate a plurality of sensitive sample images from the plurality of standard source images, and screen a plurality of valid sample images.

8. The remote sensing image recognition system based on deep learning according to claim 6, characterized in that: The image stage processing module specifically includes: A demand analysis unit, used to determine standardized sizes and multiple stage resolutions according to preset staged requirements; A standardized cutting unit, used for performing standardized cutting processing on the plurality of valid sample images according to the standardized size to generate a plurality of standard sample images; The resolution stage adjustment unit is used to perform resolution stage adjustment on the plurality of standard sample images according to the plurality of stage resolutions to generate a plurality of stage sample images.

Images