Image processing method and device, computer device, storage medium and program product

Abstract

The application provides an image processing method and device, computer equipment, a storage medium and a program product, and relates to the field of remote sensing image processing. The image processing method comprises the following steps: acquiring a plurality of target domain ground feature images; the target domain ground feature image is a ground feature image of a target field to be identified; performing style migration on the target domain ground feature image to a source field to obtain a migrated ground feature image; the source domain sample ground feature image is a ground feature image of the source field, which is labeled with a standard ground class mark; performing classification on the migrated ground feature image based on a ground feature classification model to obtain a ground feature classification result of the target domain ground feature image; and the ground feature classification model is obtained by training based on the source domain sample ground feature image. The accuracy of ground feature classification for the target domain can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and relates to an image processing method and device, a computer device, a storage medium and a program product. BACKGROUND

[0002] Remote sensing technology is currently widely used in geosciences, such as forest resource planning, crop yield estimation, environmental assessment, disaster monitoring, etc. Remote sensing image classification technology is a key step for converting remote sensing images from data to information. Through a semantic segmentation model, remote sensing images can be classified in detail.

[0003] Currently, when classifying a new land cover classification task, a land cover classification model is usually transferred from an existing land cover classification task, but it is difficult to guarantee the generalization ability of the transferred model, which may result in low accuracy of the new land cover classification. SUMMARY

[0004] The present application provides an image processing method, device, computer device, storage medium and program product, which can solve the problem of low accuracy of new land cover classification in related technologies. The technical solution is as follows:

[0005] In one aspect, an image processing method is provided, comprising:

[0006] Obtaining a plurality of target domain land cover images; the target domain land cover image is a land cover image of a target field to be identified;

[0007] Performing style transfer on the target domain land cover image to the source field to obtain a transferred land cover image;

[0008] The source domain sample land cover image is a land cover image of the source field labeled with a standard land class label.

[0009] Classifying the transferred land cover image based on a land cover classification model to obtain a land cover classification result of the target domain land cover image;

[0010] The land cover classification model is trained based on the source domain sample land cover image.

[0011] In one possible implementation, performing style transfer on the target domain land cover image to the source field based on the source domain sample land cover image to obtain a transferred land cover image, comprises:

[0012] Obtaining a land cover style transfer model based on the source domain sample land cover image and the target domain land cover image;

[0013] Performing style transfer on the target domain land cover image to the source field through the land cover style transfer model to obtain a transferred land cover image.

[0014] In a possible implementation, the feature style transfer model is trained in the following manner:

[0015] Obtain a plurality of source domain sample feature images;

[0016] For each source domain sample feature image, at least one source domain single-class feature image is extracted from the source domain sample feature image based on a standard class label;

[0017] For each target domain feature image, the target domain feature image is segmented to obtain at least one sub-region;

[0018] The sub-regions of each target domain feature image are classified based on a feature index feature to obtain at least one target domain single-class feature image;

[0019] The initial style transfer model is trained based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model.

[0020] In a possible implementation, the standard class label includes a class label of at least one type of feature; and the at least one source domain single-class feature image is extracted from the source domain sample feature image based on the standard class label, including:

[0021] For each class label in the source domain sample feature image, the class label is selected to face a corresponding type of feature for extraction;

[0022] Regions other than the selected features in the source domain sample feature image are set to a preset color to obtain a source domain single-class feature image for the selected type of feature.

[0023] In a possible implementation, the target domain feature image is segmented to obtain at least one sub-region, including:

[0024] The target domain feature image is mapped to a corresponding undirected graph;

[0025] The undirected graph is segmented into two unconnected first vertex sets and second vertex sets;

[0026] The target domain feature image is segmented based on the first vertex set and the second vertex set to obtain at least one sub-region.

[0027] In a possible implementation, the sub-regions of each target domain feature image are classified based on a feature index feature to obtain at least one target domain single-class feature image, including:

[0028] Based on the feature index feature, the sub-regions of each target domain feature image are classified to obtain at least one sub-region corresponding to each type of feature;

[0029] For any one type of feature in the target domain feature image, the color channel information of the sub-region corresponding to the one type of feature is maintained;

[0030] The other sub-regions in the target domain feature image except the sub-region corresponding to the selected one type of feature are set to a preset color to obtain a target domain single-feature label for the selected one type of feature.

[0031] In one possible implementation, the land class index feature includes at least one of a normalized vegetation index and a soil-adjusted vegetation index.

[0032] In one possible implementation, the initial style transfer model is trained based on the at least one source domain single-feature image and the at least one target domain single-feature image to obtain the feature style transfer model, including:

[0033] determining at least one type of feature corresponding to the at least one source domain single-feature image;

[0034] selecting, from the at least one target domain single-feature image, target domain single-feature images respectively corresponding to the determined at least one type of feature;

[0035] training the initial style transfer model based on the at least one source domain single-feature image and the selected target domain single-feature images to obtain the feature style transfer model.

[0036] In one possible implementation, the initial style transfer model is trained based on the at least one source domain single-feature image and the at least one target domain single-feature image to obtain the feature style transfer model, including:

[0037] adjusting the at least one source domain single-feature image and the at least one target domain single-feature image to a preset size;

[0038] training the initial style transfer model based on the source domain single-feature image and the target domain single-feature image adjusted to the preset size to obtain the feature style transfer model.

[0039] In another aspect, an image processing device is provided, and the device includes:

[0040] an acquisition module configured to acquire a plurality of target domain feature images; the target domain feature images are feature images of a target field to be identified;

[0041] a transfer module configured to perform style transfer on the target domain feature images to a source field to obtain transferred feature images;

[0042] The feature style transfer model is trained based on the target domain feature images and source domain sample feature images; the source domain sample feature images are feature images of the source field labeled with standard land class labels.

[0043] a classification module configured to classify the target domain land object image based on a land object classification model to obtain a land object classification result of the target domain land object image;

[0044] The land object classification model is obtained by training based on the source domain sample land object image.

[0045] In one possible implementation, the migration module is specifically configured to:

[0046] obtain a land object style migration model based on the source domain sample land object image and the target domain land object image;

[0047] migrate the target domain land object image to the source domain based on the land object style migration model to obtain the migrated land object image.

[0048] In one possible implementation, the image processing apparatus further comprises a training module configured to:

[0049] obtain a plurality of source domain sample land object images;

[0050] for each source domain sample land object image, extract at least one source domain single-class land object image from the source domain sample land object image based on a standard land class label;

[0051] for each target domain land object image, segment the target domain land object image to obtain at least one sub-region;

[0052] classify the sub-regions of each target domain land object image based on a land object index feature to obtain at least one target domain single-class land object image;

[0053] train an initial style migration model based on the at least one source domain single-class land object image and the at least one target domain single-class land object image to obtain the land object style migration model.

[0054] In one possible implementation, the standard land class label comprises a land class label of at least one type of land object; and the training module is specifically configured to:

[0055] for each land class label in the source domain sample land object image, select the land class label to extract a corresponding type of land object;

[0056] set other regions in the source domain sample land object image except the selected land object to a preset color to obtain a source domain single-class land object image for the selected type of land object.

[0057] In a possible implementation, when the training module segments the target domain feature image to obtain at least one sub-region, the training module is specifically configured to:

[0058] map the target domain feature image to a corresponding undirected graph;

[0059] segment the undirected graph into two unconnected first vertex sets and second vertex sets;

[0060] segment the target domain feature image based on the first vertex sets and the second vertex sets to obtain at least one sub-region.

[0061] In a possible implementation, when the training module classifies the sub-regions of each target domain feature image based on the feature index of the ground object to obtain at least one target domain single-class feature image, the training module is specifically configured to:

[0062] classify the sub-regions of each target domain feature image based on the feature index of the ground object to obtain at least one sub-region corresponding to each class of ground object;

[0063] for any class of ground object in the target domain feature image, maintain the color channel information of the sub-region corresponding to the class of ground object;

[0064] set other sub-regions in the target domain feature image except the sub-region corresponding to the selected class of ground object to a preset color to obtain a target domain single-class feature label for the selected class of ground object.

[0065] In a possible implementation, the feature index of the ground object includes at least one of a normalized vegetation index and a soil-adjusted vegetation index.

[0066] In a possible implementation, when the training module trains an initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, the training module is specifically configured to:

[0067] determine at least one class of ground object corresponding to the at least one source domain single-class feature image;

[0068] select, from the at least one target domain single-class feature image, target domain single-class feature images respectively corresponding to the determined at least one class of ground object;

[0069] train the initial style transfer model based on the at least one source domain single-class feature image and the selected target domain single-class feature images to obtain the feature style transfer model

[0070] In a possible implementation, when the training module trains an initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, the training module is specifically configured to:

[0071] adjust the at least one source domain single-class feature image and the at least one target domain single-class feature image to a preset size;

[0072] train the initial style transfer model based on the source domain single-class feature image and the target domain single-class feature image adjusted to the preset size to obtain the feature style transfer model.

[0073] In one possible implementation, when the training module trains the initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, the training module is specifically configured to:

[0074] for any one type of feature, taking the target domain single-class feature image corresponding to the one type of feature as input and taking the source domain single-class feature image corresponding to the one type of feature as output, the initial style transfer model is trained until a preset end condition of a training loss function for the one type of feature is reached, and an intermediate model corresponding to the one type of feature is obtained;

[0075] for each type of feature, the intermediate model obtained by the last training is repeatedly set as a new initial style transfer model, the target domain single-class feature image corresponding to the one type of feature is taken as input, and the source domain single-class feature image corresponding to the one type of feature is taken as output, the initial style transfer model is trained until the preset end condition of the training loss function for the one type of feature is reached;

[0076] the intermediate model obtained by the training for the last type of feature is set as the feature style transfer model.

[0077] In another aspect, a computer device is provided, including a memory, a processor, and a computer program stored in the memory, the processor executes the computer program to implement the image processing method described above.

[0078] In another aspect, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, the computer program is executed by a processor to implement the image processing method described above.

[0079] In another aspect, a computer program product is provided, and the computer program product includes a computer program, the computer program is executed by a processor to implement the image processing method described above.

[0080] The technical scheme provided in the present application has the beneficial effects that:

[0081] The feature style transfer model is trained through the target domain feature image and the source domain sample feature image, and then the target domain feature image is transferred to the source domain through the feature style transfer model to obtain a transferred feature image similar to the source domain, and then the transferred feature image is classified through the feature classification model suitable for the source domain, so that the classification result of the target domain feature image is obtained, and the accuracy of the feature classification of the target domain is effectively improved.

[0082] Further, at least one source domain single feature image is extracted from the source domain sample feature image, at least one target domain single feature image is extracted from the target domain feature image, and the initial style transfer model is trained based on the at least one source domain single feature image and the at least one target domain single feature image, so that the obtained style transfer model can perform style transfer on the selected feature characteristics, thereby improving the accuracy and generalization of the style transfer model, and the feature classification model suitable for the source domain has better classification effect on the transferred feature image and higher classification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0083] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the description of the embodiments of the present application will be briefly introduced.

[0084] Figure 1 An implementation environment schematic diagram of an image processing method provided by the embodiments of the present application;

[0085] Figure 2 A flowchart of an image processing method provided by the embodiments of the present application;

[0086] Figure 3 An acquisition scheme schematic diagram of a feature style transfer model provided by the embodiments of the present application;

[0087] Figure 4 A scheme schematic diagram for processing the source domain sample feature image provided by the embodiments of the present application;

[0088] Figure 5 A scheme schematic diagram for processing the target domain feature image provided by the embodiments of the present application;

[0089] Figure 6 A scheme schematic diagram of an image processing scheme provided by the embodiments of the present application;

[0090] Figure 7 A structure schematic diagram of an image processing device provided by the embodiments of the present application;

[0091] Figure 8 A structure schematic diagram of a computer device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0092] The embodiments of the present application will be described below in conjunction with the accompanying drawings. It should be understood that the embodiments described below in conjunction with the accompanying drawings are exemplary descriptions of the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0093] Those skilled in the art can understand that the singular forms "a", "an" and "the" used herein include plural forms unless specifically stated otherwise. It should be further understood that the terms "include" and "contain" used in the embodiments of the present application mean that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude other features, information, data, steps, operations, elements, components and / or their combinations supported by the present technology. It should be understood that when we say that an element is "connected" or "coupled" to another element, the element can be directly connected or coupled to the other element, or it can mean that the element and the other element are connected through an intermediate element. In addition, "connected" or "coupled" used herein can include wireless connection or wireless coupling. The term "and / or" used herein indicates at least one of the items defined by the term, for example, "A and / or B" indicates implementation as "A", or implementation as "A", or implementation as "A and B".

[0094] In order to make the purposes, technical solutions and advantages of the present application clearer, the embodiments of the present application will be described in further detail below in conjunction with the accompanying drawings.

[0095] Remote sensing technology is currently widely used in geosciences applications, such as forest resource planning, crop yield estimation, environmental assessment, disaster monitoring, etc. Remote sensing image classification technology is a key step to convert remote sensing images from data to information. Among them, through the semantic segmentation model, the remote sensing image can be classified in detail. In the model construction process, in order to improve the generalization ability of the ground object classification model on images of different satellites, different times and different regions, the method of transfer learning is often used, which uses the rules learned from an existing ground object classification task (containing remote sensing images and corresponding label data of a certain satellite, a certain time and a certain region, usually referred to as the source domain) to classify a brand new ground object classification task (only containing remote sensing images of another satellite, another time and another region, without or with less label data, not enough to train an effective classifier, usually referred to as the target domain).

[0096] Currently, a data subset with the maximum similarity in the source domain and the target domain can be found to assist the target domain in feature classification; or feature extraction can be performed to reduce the dimension to a feature subspace with the maximum similarity in the source domain and the target domain, and the feature subspace is used for training of a feature classification model of the target domain. The two methods have achieved certain results in theoretical research, but there is still a certain distance from actual production application.

[0097] In the related art, principal component transformation is performed on the source domain image and the target domain image, the first three principal component components are extracted to form three new band images, then the new images are input into a convolutional neural network to obtain corresponding deep features, and finally, based on the connection between the deep features, transfer learning from the source domain image to the target domain image is realized. In this way, the model generalization ability is difficult to guarantee.

[0098] In some related art, features are extracted for the whole image to realize transformation of the overall effect of the target domain. If unavoidable noise such as cloud layer, shadow and other interference occurs in the source domain, the noise feature extracted by the transfer learning will indirectly affect the feature classification effect of the target domain.

[0099] The present application faces the feature, and trains a feature style transfer model based on deep convolutional network for the source domain and the target domain image. Through the present application, the transfer learning between the source domain and the target domain is realized, the influence of the source domain noise on the target domain data in the transfer learning process is avoided, and the generalization ability of the feature classification model can be effectively improved.

[0100] The embodiments of the present application can be applied to various scenes, including but not limited to artificial intelligence and the like. For example, the image processing method provided by the present application can be applied to the following scene: in a plurality of similar remote sensing feature classification projects, there are multiple source domain remote sensing images and corresponding label data and an initial feature classification algorithm, and it is necessary to classify features of a target domain remote sensing image.

[0101] Artificial intelligence (AI) is the use of digital computers or digital computer-controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making.

[0102] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software level technology. Artificial intelligence basic technology generally includes, such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning, automatic driving, intelligent transportation and several other major directions.

[0103] Machine learning (ML) is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a specialized study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning.

[0104] Figure 1 is a schematic diagram of an implementation environment of an image processing method provided by an embodiment of the application, see Figure 1 , specifically, the ground feature style transfer model is trained based on the target domain ground feature image and the source domain sample ground feature image, the target domain ground feature image is style transferred to the source domain through the ground feature style transfer model, a migrated ground feature image is obtained, and then the migrated ground feature image is classified based on the ground feature classification model suitable for the source domain to obtain the ground feature classification result of the target domain ground feature image.

[0105] It can be understood that, Figure 1 represents an application scenario in an example, and does not limit the application scenario of the image processing method of the application.

[0106] Figure 2 is a flowchart of an image processing method provided by an embodiment of the application. The execution subject of the method can be a computer device. As shown in Figure 2 , the method can include the following steps:

[0107] Step 201, obtaining a plurality of target domain ground feature images.

[0108] Among them, the target domain ground feature image is a ground feature image to be identified in a target field.

[0109] Among them, the target field can refer to a specific spatial resolution, a specific satellite, a specific time, and a specific region, and is not limited to a specific region.

[0110] Step 202, style migration of the target domain feature image to the source domain is performed to obtain a migrated feature image.

[0111] The source domain sample feature image is a feature image of the source domain labeled with a standard feature class label.

[0112] Specifically, the target domain feature image and the source domain sample feature image can be transformed by a feature style migration model, or the target domain feature image can be directly style migrated to the source domain based on image processing or other mathematical calculations.

[0113] Specifically, each target domain single-class feature image can be extracted from the target domain feature image, and each source domain single-class feature image can be extracted from the source domain feature image. The style migration transformation is obtained according to the target domain feature image and the corresponding feature images of the source domain.

[0114] Step 203, the migrated feature image is classified based on a feature classification model to obtain a feature classification result of the target domain feature image.

[0115] The feature classification model is obtained by training based on the source domain sample feature image.

[0116] The feature classification model can be a semantic segmentation network, for example, an HRNet network (High-Resolution Net).

[0117] In the specific implementation process, the feature classification model is suitable for the source domain. After the target domain feature image is style migrated to the source domain, the migrated feature image is similar to the source domain, and the migrated feature image can be classified by the feature classification model suitable for the source domain.

[0118] In the above embodiment, the feature style migration model is obtained by training based on the target domain feature image and the source domain sample feature image. Then, the target domain feature image is style migrated to the source domain by the feature style migration model to obtain a migrated feature image similar to the source domain. The migrated feature image is classified by the feature classification model suitable for the source domain, and the classification result of the target domain feature image can be obtained, which effectively improves the accuracy of the feature classification of the target domain.

[0119] In one possible implementation, step S202 of style migration of the target domain feature image to the source domain based on the source domain sample feature image to obtain a migrated feature image can include:

[0120] (1) obtaining a feature style migration model based on the source domain sample feature image and the target domain feature image;

[0121] (2) performing style transfer on the target domain feature image to the source domain feature image by using the feature style transfer model to obtain a transferred feature image.

[0122] Specifically, the feature style transfer model can be trained by using the source domain feature image and the target domain feature image to obtain the feature style transfer model.

[0123] In the above embodiment, the style conversion between the source domain feature image and the target domain feature image is realized by using the feature style transfer model. In other embodiments, the style conversion can also be realized by using an image transfer algorithm, histogram transformation, etc.

[0124] The process of obtaining the feature style transfer model will be further described below with reference to an embodiment.

[0125] In one possible implementation, as shown in FIG. 3, the feature style transfer model can be trained in the following manner: Figure 3

[0126] Step S301: Obtain a plurality of source domain feature images.

[0127] Each source domain feature image is labeled with a standard feature label, that is, each feature in the source domain feature image is provided with a corresponding feature label.

[0128] Step S302: For each source domain feature image, at least one source domain single-feature image is extracted from the source domain feature image based on the standard feature label.

[0129] Each source domain feature image contains at least one feature, and each feature is provided with a corresponding feature label.

[0130] Specifically, the step S302 of extracting at least one source domain single-feature image from the source domain feature image based on the standard feature label can include:

[0131] (1) For each feature label in the source domain feature image, the feature label is selected to face the corresponding feature for extraction;

[0132] (2) The other regions in the source domain feature image except the selected feature are set to a preset color to obtain a source domain single-feature image for the selected feature.

[0133] In the specific implementation process, the source domain feature image contains standard feature labels of multiple features. The standard feature labels of the multiple features are first mapped back to the original image. For each feature label, the corresponding feature is extracted, and the remaining features are set to a background color to generate a source domain single-feature image for the feature.

[0134] ​That is, the region belonging to the object in the original image keeps the original RGB (color system) channel information, and the region belonging to other categories of objects is set to black to highlight each object belonging to the category on the image. Each category of object is extracted in turn, and the region corresponding to other categories of objects is set to black to obtain a source domain single-category image corresponding to each category of object.

[0135] As shown in the figure, the source domain data set image includes a plurality of source domain sample object images, and the source domain sample object image is provided with a corresponding standard category label, that is, the source domain data set multi-category label shown in the figure. Single-category extraction is performed on the source domain sample object image to obtain a source domain single-category object image of each category of object, that is, the a-class object image...n-class object image shown in the figure. Figure 4 Figure 4

[0136] Step S303, for each target domain object image, the target domain object image is segmented to obtain at least one sub-region.

[0137] Specifically, a normalization method based on graph theory can be used to segment the target domain object image to obtain at least one sub-region.

[0138] In the specific implementation process, the target domain object image can be first mapped to a corresponding undirected graph, and then the undirected graph is divided into two unconnected vertex sets, and then the segmented sub-regions are determined based on the two unconnected vertex sets. The segmentation process of the target domain object image will be described in detail below.

[0139] Step S304, based on the object index feature, the sub-regions of each target domain object image are classified to obtain at least one target domain single-category object image.

[0140] The object index feature can include normalized difference vegetation index (NDVI), soil-adjusted vegetation index (SAVI), etc.

[0141] Specifically, the object index feature can be combined to utilize the different reflection or absorption characteristics of different objects to different bands for preliminary classification of the sub-regions to obtain a target domain single-category image.

[0142] Specifically, the categories corresponding to the source domain single-category images and the categories corresponding to the target single-category images are respectively one-to-one corresponding. For example, the categories corresponding to the plurality of source domain single-category images are A, B, C, and D categories, respectively, and the categories corresponding to the plurality of target single-category images are also A, B, C, and D categories, respectively.

[0143] In the specific implementation process, step S304 can include:

[0144] (1) Based on the land type index characteristics, the sub-regions of the land feature images of each target domain are classified to obtain at least one sub-region corresponding to each type of land feature;

[0145] (2) For any type of land cover in the target domain land cover image, maintain the color channel information of the sub-region corresponding to the type of land cover;

[0146] (3) Set the sub-regions in the target domain land cover image other than the sub-regions corresponding to the selected land cover type to preset colors to obtain the target domain single land cover label for the selected land cover type.

[0147] Specifically, in the target domain land cover image, the sub-regions belonging to one type of land cover can retain their original RGB channel information, while the sub-regions belonging to other types of land cover can be set to black. This can highlight the selected type of land cover in the image. Repeat this operation for each type of land cover to obtain the target domain single-type land cover label corresponding to each type of land cover.

[0148] like Figure 5 As shown, the target domain dataset includes multiple target domain land cover images. Based on the land cover index features, the target domain land cover images are extracted, i.e., the index feature extraction in the figure, to obtain the target domain single-class land cover image for each type of land cover, i.e., the land cover image of class a in the figure...the land cover image of class n.

[0149] Step S305: Train the initial style transfer model based on at least one source domain single-class land cover image and at least one target domain single-class land cover image to obtain the land cover style transfer model.

[0150] Specifically, the initial style transfer model can be trained using source domain single-class land cover images and target domain single-class land cover images corresponding to each type of land cover, so that the style transfer model obtained after training can perform style transfer for each type of land cover feature.

[0151] In the specific implementation process, step S305 trains the initial style transfer model based on at least one source domain single-class land cover image and at least one target domain single-class land cover image to obtain the land cover style transfer model, which may include:

[0152] Adjust at least one source domain single-class land cover image and at least one target domain single-class land cover image to the preset size;

[0153] An initial style transfer model is trained based on source domain single-class land cover images and target domain single-class land cover images adjusted to a preset size, resulting in a land cover style transfer model.

[0154] Specifically, the size of the source domain single-class feature image and the target domain single-class feature image can be unified before training, for example, all are cropped to 512*512 size, so as to facilitate subsequent training of the initial style transfer model.

[0155] In the implementation process, the step S305 trains the initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, which can include:

[0156] Determining at least one type of feature corresponding to the at least one source domain single-class feature image;

[0157] Selecting target domain single-class feature images corresponding to the at least one type of feature from the at least one target domain single-class feature image;

[0158] Training the initial style transfer model based on the at least one source domain single-class feature image and the selected target domain single-class feature image to obtain the feature style transfer model.

[0159] Specifically, the type of feature corresponding to the source domain single-class feature image and the type of feature corresponding to the target single-class feature image are one-to-one corresponding.

[0160] In the implementation process, the step S305 trains the initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, which can include:

[0161] Inputting the at least one source domain single-class feature image into the initial style transfer model, taking the at least one target domain single-class feature image as the output, and training the initial style transfer model based on the at least one source domain single-class feature image corresponding to the at least one target domain single-class feature image to obtain the feature style transfer model.

[0162] Specifically, the preset end condition can be that the training loss function converges, or the training reaches a preset number of times, which is not limited here.

[0163] For example, there are three types of features in the target domain feature image, and there are also three types of corresponding features in the source domain sample feature image. Three target domain single-class feature images corresponding to the three types of features are taken as input, and the source domain single-class feature images corresponding to the three types of features are taken as output. The initial style transfer model is trained to obtain the feature style transfer model.

[0164] Specifically, the loss function in the training process can be determined based on the following manner:

[0165] For each type of feature, a first loss function between the source domain sample feature image corresponding to the type of feature and the target domain feature image is determined;

[0166] The sum of the first loss function corresponding to each type of ground object is set as the loss function in the training process.

[0167] In other embodiments, the average value between the first loss function corresponding to each type of ground object can also be determined to obtain the loss function in the training process.

[0168] The loss function in the training process is determined based on at least one source domain single-class ground object image and at least one target domain single-class ground object image, and the specific determination manner is not limited here.

[0169] In the above embodiments, at least one source domain single-class ground object image is extracted from the source domain sample ground object image, at least one target domain single-class ground object image is extracted from the target domain ground object image, and the initial style transfer model is trained based on the at least one source domain single-class ground object image and the at least one target domain single-class ground object image, so that the obtained style transfer model can perform style transfer on the feature of each type of ground object, thereby improving the generalization of the style transfer model, so that the ground object classification model suitable for the source field has better classification effect on the transferred ground object image, and the classification accuracy is higher.

[0170] The above embodiments describe the specific training process of the style transfer model, and the segmentation process for the target domain ground object image will be further described below in combination with the drawings and embodiments.

[0171] Specifically, for the target domain ground object image, a normalized segmentation method based on graph theory can be used for segmentation. The target domain ground object image is mapped into an undirected graph, and the connection between the vertices of the graph is the weight.

[0172] For the target domain ground object image I, a weighted undirected graph G=(V, E) is constructed, V is the set of all pixel nodes in the graph, and E is the edge of the graph. By deleting some edges of the graph, it is divided into two unconnected vertex sets A and B. The sum of the deleted edges is the dissimilarity of A and B.

[0173] The segmentation formula based on graph theory is as follows:

[0174]

[0175] Where W(u, v) represents the similarity between pixels i and j in the image.

[0176] The calculation formula of the normalized graph theory segmentation is as follows:

[0177]

[0178] Where assoc(A, V) is the similarity between the points in A and the points in the graph; and assoc(B, V) is the similarity between the points in B and the points in the graph.

[0179] The minimum problem of solving Ncut(A, B) is converted into a problem of solving the eigenvalue and eigenvector of a matrix. The obtained eigenvector is clustered by using a clustering algorithm, and then the segmentation effect is obtained.

[0180] In the above embodiment, the target domain feature image is extracted by using the index feature. In other embodiments, the target domain feature image can also be extracted by using a threshold-based segmentation method, a region-based segmentation method, an edge-based segmentation method, etc., which are not limited herein. In the specific implementation process, the segmented region is mapped back to the original image, combined with the index feature, and the different features have different reflection or absorption characteristics for different wave bands, so as to preliminarily classify the segmented sub-regions and obtain the target domain single feature image. That is, the sub-regions belonging to the feature in the original image retain the original RGB channel information, and the sub-regions belonging to other features are set to black, so as to highlight each feature on the image. The above operation is repeated to obtain the target domain single feature image corresponding to each feature.

[0181] In the above embodiment, the target domain feature image is extracted by using the index feature. In other embodiments, the target domain feature image can also be extracted by using a threshold-based segmentation method, a region-based segmentation method, an edge-based segmentation method, etc., which are not limited herein. In the specific implementation process, the segmented region is mapped back to the original image, combined with the index feature, and the different features have different reflection or absorption characteristics for different wave bands, so as to preliminarily classify the segmented sub-regions and obtain the target domain single feature image. That is, the sub-regions belonging to the feature in the original image retain the original RGB channel information, and the sub-regions belonging to other features are set to black, so as to highlight each feature on the image. The above operation is repeated to obtain the target domain single feature image corresponding to each feature.

[0182] In order to more clearly illustrate the image processing method of the present application, the image processing method of the present application will be further described below in conjunction with examples.

[0183] As shown in Figure 6 In one example, the image processing method of the present application can include:

[0184] Obtaining a plurality of source domain sample feature images, i.e., the source domain dataset images shown in the figure; the source domain sample feature images are labeled with standard feature labels, i.e., the multi-class labels of the source domain dataset shown in the figure;

[0185] Training a feature classification model based on the source domain sample feature images and the standard feature labels;

[0186] Extracting single features of the source domain sample feature images based on the standard feature labels, to obtain source domain single feature images of each type of feature of the source domain, i.e., the a-type feature images of the source domain, …, n-type feature images of the source domain shown in the figure;

[0187] Obtaining a plurality of target domain feature images to be identified, i.e., the target domain dataset images shown in the figure;

[0188] Exponential feature extraction is performed on the target domain feature image to obtain a target domain single-feature image of each type of feature in the target domain, i.e., the a-type feature image to the n-type feature image of the target domain shown in the figure.

[0189] The initial style transfer model is trained by using the source domain single-feature image and the target domain single-feature image of each type of feature to obtain a feature style transfer model.

[0190] The target domain feature image is subjected to style transfer to the source domain by using the feature style transfer model to obtain a transferred feature image, i.e., the image after style transfer shown in the figure.

[0191] The transferred feature image is classified based on the feature classification model to obtain a feature classification result of the target domain feature image.

[0192] The above image processing method trains a feature style transfer model by using a target domain feature image and a source domain sample feature image, and then subjects the target domain feature image to style transfer to the source domain by using the feature style transfer model to obtain a transferred feature image similar to the source domain. The transferred feature image is classified by using a feature classification model suitable for the source domain, so that a classification result for the target domain feature image can be obtained, and the accuracy of feature classification for the target domain is effectively improved.

[0193] Further, at least one source domain single-feature image is extracted from a source domain sample feature image, and at least one target domain single-feature image is extracted from a target domain feature image. The initial style transfer model is trained based on the at least one source domain single-feature image and the at least one target domain single-feature image, so that the obtained style transfer model can perform style transfer for each type of feature, thereby improving the generalization of the style transfer model, and enabling the feature classification model suitable for the source domain to have better classification effect on the transferred feature image and have higher classification accuracy.

[0194] Figure 7 A structure schematic diagram of an image processing device provided by an embodiment of the present application is shown in FIG. 1. Figure 7 As shown in the figure, the device comprises:

[0195] The acquisition module 701 is configured to acquire a plurality of target domain feature images; the target domain feature image is a feature image of a target domain to be identified;

[0196] The transfer module 702 is configured to subject the target domain feature image to style transfer to the source domain to obtain a transferred feature image.

[0197] The feature style transfer model is trained based on the target domain feature image and the source domain sample feature image; the source domain sample feature image is a feature image of the source domain labeled with a standard feature type label.

[0198] The classification module 703 is configured to classify the target domain land object image based on a land object classification model to obtain a land object classification result of the target domain land object image.

[0199] The land object classification model is obtained by training based on the source domain sample land object image.

[0200] In one possible implementation, the migration module 702 is configured to, when performing style migration of the target domain land object image to the source domain based on the source domain sample land object image to obtain the migrated land object image, specifically:

[0201] obtain a land object style migration model based on the source domain sample land object image and the target domain land object image;

[0202] perform style migration of the target domain land object image to the source domain through the land object style migration model to obtain the migrated land object image.

[0203] In one possible implementation, the image processing apparatus further includes a training module configured to:

[0204] obtain a plurality of source domain sample land object images;

[0205] for each source domain sample land object image, extract at least one source domain single-class land object image from the source domain sample land object image based on a standard land class label;

[0206] for each target domain land object image, perform segmentation on the target domain land object image to obtain at least one sub-region;

[0207] classify the sub-regions of each target domain land object image based on a land object index feature to obtain at least one target domain single-class land object image;

[0208] train an initial style migration model based on the at least one source domain single-class land object image and the at least one target domain single-class land object image to obtain a land object style migration model.

[0209] In one possible implementation, the standard land class label includes a land class label of at least one type of land object, and the training module, when extracting the at least one source domain single-class land object image from the source domain sample land object image based on the standard land class label, is specifically configured to:

[0210] for each land class label in the source domain sample land object image, select the land class label to extract a corresponding type of land object;

[0211] set other regions in the source domain sample land object image except the selected land object to a preset color to obtain a source domain single-class land object image for the selected type of land object.

[0212] In a possible implementation, when the training module performs segmentation on the target domain feature image to obtain at least one sub-region, the training module is specifically configured to:

[0213] map the target domain feature image into a corresponding undirected graph;

[0214] segment the undirected graph into a first vertex set and a second vertex set that are not connected;

[0215] segment the target domain feature image based on the first vertex set and the second vertex set to obtain at least one sub-region.

[0216] In a possible implementation, when the training module classifies the sub-regions of each target domain feature image based on the feature index of the ground object to obtain at least one target domain single-class feature image, the training module is specifically configured to:

[0217] classify the sub-regions of each target domain feature image based on the feature index of the ground object to obtain at least one sub-region corresponding to each type of ground object;

[0218] for any type of ground object in the target domain feature image, maintain the color channel information of the sub-region corresponding to the type of ground object;

[0219] set other sub-regions in the target domain feature image except the sub-region corresponding to the selected type of ground object to a preset color to obtain a target domain single-class feature label for the selected type of ground object.

[0220] In a possible implementation, the feature index of the ground object includes at least one of a normalized vegetation index and a soil-adjusted vegetation index.

[0221] In a possible implementation, when the training module trains the initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, the training module is specifically configured to:

[0222] determine at least one type of ground object corresponding to the at least one source domain single-class feature image;

[0223] select, from the at least one target domain single-class feature image, target domain single-class feature images respectively corresponding to the at least one type of ground object determined;

[0224] train the initial style transfer model based on the at least one source domain single-class feature image and the selected target domain single-class feature images to obtain the feature style transfer model. In a possible implementation, when the training module trains the initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, the training module is specifically configured to:

[0225] adjust the at least one source domain single-class feature image and the at least one target domain single-class feature image to a preset size;

[0226] train the initial style transfer model based on the source domain single-class feature image and the target domain single-class feature image adjusted to the preset size to obtain the feature style transfer model.

[0227] In one possible implementation, when the training module trains the initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style transfer model, the training module is specifically configured to:

[0228] For any one type of feature, the target domain single-class feature image corresponding to the one type of feature is taken as input, the source domain single-class feature image corresponding to the one type of feature is taken as output, the initial style transfer model is trained until the training loss function for the one type of feature reaches a preset end condition, and the intermediate model corresponding to the one type of feature is obtained.

[0229] For each type of feature, the intermediate model obtained by the last training is set as a new initial style transfer model, the target domain single-class feature image corresponding to the one type of feature is taken as input, the source domain single-class feature image corresponding to the one type of feature is taken as output, the initial style transfer model is trained until the training loss function for the one type of feature reaches a preset end condition.

[0230] The intermediate model obtained by training for the last type of feature is set as the feature style transfer model.

[0231] The image processing apparatus described above trains the feature style transfer model through the target domain feature image and the source domain sample feature image, performs style transfer on the target domain feature image to the source domain through the feature style transfer model, obtains a transferred feature image similar to the source domain, and classifies the transferred feature image through the feature classification model suitable for the source domain, so that the classification result for the target domain feature image can be obtained, and the accuracy of feature classification for the target domain is effectively improved.

[0232] Further, at least one source domain single-class feature image is extracted from the source domain sample feature image, at least one target domain single-class feature image is extracted from the target domain feature image, and the initial style transfer model is trained based on the at least one source domain single-class feature image and the at least one target domain single-class feature image, so that the obtained style transfer model can perform style transfer on the feature of each type, thereby improving the generalization of the style transfer model, and the feature classification model suitable for the source domain has better classification effect on the transferred feature image and higher classification accuracy.

[0233] Figure 8 is a structural schematic diagram of a computer device provided in an embodiment of the present application. As shown inFigure 8 As shown in the figure, the computer device comprises a memory and a processor; at least one program stored in the memory for being executed by the processor to realize, compared with the prior art:

[0234] The feature style transfer model is trained through the target domain feature image and the source domain sample feature image, and then the target domain feature image is transferred to the source domain through the feature style transfer model to obtain a transferred feature image similar to the source domain. Then, the transferred feature image is classified through the feature classification model suitable for the source domain, so that the classification result of the target domain feature image can be obtained, and the accuracy of the feature classification of the target domain is effectively improved.

[0235] Further, at least one source domain single-class feature image is extracted from the source domain sample feature image, and at least one target domain single-class feature image is extracted from the target domain feature image. The initial style transfer model is trained based on the at least one source domain single-class feature image and the at least one target domain single-class feature image, so that the obtained style transfer model can perform style transfer on each class of feature characteristics, thereby improving the generalization of the style transfer model, and making the feature classification model suitable for the source domain have better classification effect on the transferred feature image and higher classification accuracy.

[0236] In an optional embodiment, a computer device is provided, as shown in the figure, Figure 8 As shown in the figure, Figure 8 The computer device 800 shown in the figure comprises a processor 801 and a memory 803. The processor 801 and the memory 803 are connected, such as through a bus 802. Optionally, the computer device 800 can further comprise a transceiver 804, which can be used for data interaction between the computer device and other computer devices, such as data sending and / or data receiving, etc. It should be noted that in actual application, the transceiver 804 is not limited to one, and the structure of the computer device 800 does not constitute a limitation on the embodiments of the present application.

[0237] The processor 801 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It can implement or execute the various exemplary logical blocks, modules and circuits described in connection with the disclosure. The processor 801 can also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0238] The bus 802 can include a path for transmitting information between the above-mentioned components. The bus 802 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, or the like. The bus 802 can be divided into an address bus, a data bus, a control bus, and the like. For convenience of representation, Figure 8 In the figure, only one thick line is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0239] The memory 803 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0240] The memory 803 is configured to store application code (computer program) for implementing the scheme of the present application, and the processor 801 is configured to control the execution. The processor 801 is configured to execute the application code stored in the memory 803 to implement the content shown in the foregoing method embodiments.

[0241] The computer device includes, but is not limited to, a virtualized computer device, a virtual machine, a server, a service cluster, a terminal of a user, and the like.

[0242] The computer readable storage medium stores the computer program, and when the computer program is run on the computer, the computer can execute the corresponding content of the image processing method in the foregoing method embodiments.

[0243] The computer program product or the computer program includes computer instructions stored in the computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the image processing method.

[0244] It should be understood that, although each step in the flowchart of the accompanying drawings is displayed in sequence according to the indication of the arrow, these steps are not necessarily executed in sequence according to the indication of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other sequences. Moreover, at least part of the steps in the flowchart of the accompanying drawings can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0245] It should be noted that the computer-readable medium in the above disclosure can be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. The computer-readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples of the computer-readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or apparatus. In the present disclosure, the computer-readable signal medium can include a data signal carried in a baseband or as a part of a carrier wave, which carries computer-readable program code. Such a propagated data signal can take many forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable storage medium, which can send, propagate or transmit a program for use by or in conjunction with an instruction execution system, device or apparatus. The program code contained in the computer-readable medium can be transmitted by any suitable medium, including but not limited to a wire, a cable, an RF (radio frequency) or the like, or any suitable combination of the above.

[0246] The computer-readable medium described above can be contained in the electronic device described above; or can exist separately and not be assembled into the electronic device.

[0247] The computer-readable medium described above carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods shown in the above embodiments.

[0248] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0249] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0250] The modules described in the embodiments of this disclosure can be implemented in software or in hardware. The name of a module does not necessarily limit the module itself; for example, a migration module can also be described as "a module for migrating the style of geographical features".

[0251] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. An image processing method, characterized by, The method comprises: obtaining a plurality of target domain feature images; the target domain feature image is a feature image of a target field to be identified; performing style migration of the target domain feature image to a source field based on a source domain sample feature image, to obtain a migrated feature image; wherein the source domain sample feature image is a feature image of a source field labeled with a standard land class label; classifying the migrated feature image based on a feature classification model to obtain a feature classification result of the target domain feature image; wherein the feature classification model is obtained by training based on the source domain sample feature image; the style migration of the target domain feature image to the source field based on the source domain sample feature image comprises: performing style migration of the target domain feature image to the source field by the feature style migration model; the feature style migration model is obtained by training in the following manner: obtaining a plurality of source domain sample feature images; for each source domain sample feature image, extracting at least one source domain single-class feature image from the source domain sample feature image based on the standard land class label; for each target domain feature image, segmenting the target domain feature image to obtain at least one sub-region; classifying the sub-regions of each target domain feature image based on feature index characteristics to obtain at least one target domain single-class feature image, wherein the corresponding land class of the source domain single-class feature image and the corresponding land class of the target domain single-class feature image are one-to-one corresponding respectively; training an initial style migration model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the feature style migration model; the loss function in the training process of the feature style migration model can be determined in the following manner: for each land class, determining a first loss function between the source domain sample feature image corresponding to the land class and the target domain feature image; summing the first loss function corresponding to each land class as the loss function in the training process.

2. The image processing method of claim 1, wherein, The standard land class label comprises a land class label of at least one land class; extracting at least one source domain single-class feature image from the source domain sample feature image based on the standard land class label comprises: for each land class label in the source domain sample feature image, selecting the land class label to extract the corresponding land class; setting other regions in the source domain sample feature image except the selected land class to a preset color to obtain a source domain single-class feature image for the selected land class.

3. The image processing method of claim 1, wherein, The segmentation of the target domain feature image to obtain at least one sub-region comprises: mapping the target domain feature image into a corresponding undirected graph; segmenting the undirected graph into two unconnected first vertex sets and second vertex sets; segmenting the target domain feature image based on the first vertex set and the second vertex set to obtain at least one sub-region.

4. The image processing method of claim 3, wherein, The classification of the sub-regions of each target domain feature image based on feature index characteristics to obtain at least one target domain single-class feature image comprises: The sub-regions of each target domain feature image are classified based on the land type index features, and at least one sub-region corresponding to each type of feature is obtained. For any type of feature in the target domain feature image, the color channel information of the sub-region corresponding to the type of feature is maintained. The sub-regions other than the sub-region corresponding to the selected type of feature in the target domain feature image are set to a preset color to obtain a target domain single-feature label for the selected type of feature.

5. The image processing method of claim 4, wherein, The land type index features include at least one of a normalized vegetation index and a soil-adjusted vegetation index.

6. The image processing method of claim 1, wherein, The training of the initial style transfer model based on the at least one source domain single-feature image and the at least one target domain single-feature image to obtain the feature style transfer model includes: determining at least one type of feature corresponding to at least one source domain single-feature image; selecting a target domain single-feature image corresponding to the determined at least one type of feature from the at least one target domain single-feature image; training the initial style transfer model based on the at least one source domain single-feature image and the selected target domain single-feature image to obtain the feature style transfer model.

7. The image processing method of claim 1, wherein, The training of the initial style transfer model based on the at least one source domain single-feature image and the at least one target domain single-feature image to obtain the feature style transfer model includes: adjusting the at least one source domain single-feature image and the at least one target domain single-feature image to a preset size; training the initial style transfer model based on the source domain single-feature image and the target domain single-feature image adjusted to the preset size to obtain the feature style transfer model.

8. An image processing apparatus characterized by comprising: The device includes: an acquisition module configured to acquire a plurality of target domain feature images; the target domain feature images are feature images of a target field to be identified; a migration module configured to perform style migration of the target domain feature images to a source field based on source domain sample feature images, to obtain migrated feature images; the source domain sample feature images are feature images of the source field labeled with standard feature labels; a classification module configured to classify the migrated feature images based on a feature classification model to obtain a feature classification result of the target domain feature images; the feature classification model is obtained by training based on the source domain sample feature images; The migration module performs style migration of the target domain feature images to the source field based on the source domain sample feature images, and is specifically configured to: perform style migration of the target domain feature images to the source field through the feature style transfer model; The device further includes a training module configured to train the feature style transfer model in the following manner: acquire a plurality of source domain sample feature images; for each source domain sample feature image, extract at least one source domain single-feature image from the source domain sample feature image based on the standard feature labels; for each target domain feature image, segment the target domain feature image to obtain at least one sub-region; The sub-regions of the target domain feature image are classified based on feature indexes of ground objects, and at least one target domain single-class feature image is obtained, wherein the ground class corresponding to the source domain single-class feature image and the ground class corresponding to the target domain single-class feature image are one-to-one corresponding respectively; The initial style transfer model is trained based on the at least one source domain single-class feature image and the at least one target domain single-class feature image, and the ground object style transfer model is obtained. The loss function in the training process of the ground object style transfer model can be determined based on the following manner: For each type of ground object, a first loss function between the source domain sample feature image and the target domain feature image corresponding to the type of ground object is determined. The sum of the first loss functions corresponding to each type of ground object is used as the loss function in the training process.

9. The image processing apparatus according to claim 8, characterized by, The standard ground class label includes a ground class label of at least one type of ground object; when the training module extracts at least one source domain single-class feature image from the source domain sample feature image based on the standard ground class label, it is specifically used for: For each ground class label in the source domain sample feature image, the type of ground object corresponding to the ground class label is extracted; The regions other than the selected ground objects in the source domain sample feature image are set to a preset color to obtain a source domain single-class feature image for the selected type of ground object.

10. The image processing apparatus according to claim 8, characterized by, When the training module segments the target domain feature image to obtain at least one sub-region, it is specifically used for: The target domain feature image is mapped to a corresponding undirected graph; The undirected graph is segmented into two unconnected first vertex sets and second vertex sets; The target domain feature image is segmented based on the first vertex set and the second vertex set to obtain at least one sub-region.

11. The image processing apparatus according to claim 8, characterized by, When the training module classifies each sub-region of the target domain feature image based on feature indexes of ground objects to obtain at least one target domain single-class feature image, it is specifically used for: Each sub-region of the target domain feature image is classified based on ground class index features to obtain at least one sub-region corresponding to each type of ground object; For any type of ground object in the target domain feature image, the color channel information of the sub-region corresponding to the type of ground object is maintained; The other sub-regions in the target domain feature image other than the sub-region corresponding to the selected type of ground object are set to a preset color to obtain a target domain single-class feature label for the selected type of ground object.

12. The image processing apparatus according to claim 11, characterized by, The ground class index features include at least one of a normalized vegetation index and a soil-adjusted vegetation index.

13. The image processing apparatus according to claim 8, characterized by, When the training module trains an initial style transfer model based on the at least one source domain single-class feature image and the at least one target domain single-class feature image to obtain the ground object style transfer model, it is specifically used for: At least one type of ground object corresponding to at least one source domain single-class feature image is determined; Target domain single-class feature images corresponding to the determined at least one type of ground object are selected from the at least one target domain single-class feature image; The initial style transfer model is trained based on the at least one source domain single-class feature image and the selected target domain single-class feature image, and the ground object style transfer model is obtained.

14. The image processing apparatus according to claim 8, characterized by, The training module is specifically configured to: adjust the at least one source domain single-class feature image and the at least one target domain single-class feature image to a preset size; train the initial style transfer model based on the source domain single-class feature image and the target domain single-class feature image adjusted to the preset size, to obtain the feature style transfer model.

15. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-14. The processor executes the computer program to implement the image processing method of any one of claims 1 to 7.

16. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the image processing method of any one of claims 1 to 7.

17. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the image processing method of any one of claims 1 to 7.

Images