Target labeling method and system based on multi-dimensional space feature model optimal source domain
By determining the optimal source domain through a multidimensional spatial feature model and constructing a cross-category commonality description model, the problems of low image annotation efficiency and high manual cost are solved. This achieves automatic annotation of target datasets with high generalization and strong domain adaptability, thereby improving the training efficiency of image detection models.
Patent Information
- Application Number
- CN202211697161.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-28
- Publication Date
- 2026-02-27
- Estimated Expiration
- 2042-12-28
AI Technical Summary
Current technologies suffer from low efficiency in image annotation, high cost of manual annotation, and poor generalization performance of deep learning models in different scenarios and environments, requiring extensive independent training, which is time-consuming and labor-intensive.
The optimal source domain is determined by a multidimensional spatial feature model. A cross-category common description model is constructed using multidimensional feature quantitative analysis and a cross-category common description model to automatically label target images, reduce manual labeling, and improve model generalization.
It achieves automatic annotation of target datasets with high generalization and strong domain adaptability, reducing the cost and time of manual annotation and improving the training efficiency of image detection models.
Smart Images
Figure CN116168392B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing and intelligent information extraction, and in particular to a target labeling method and system based on a multi-dimensional space feature model optimal source domain. BACKGROUND
[0002] With the continuous development of computer technology, the demand for image processing is increasing. Among them, image labeling is a technology that divides an image into several specific regions and labels each region with a specific label. For example, high-performance fruit detection technology in an orchard is an important basic technology in the practical application of modern smart orchards, which can be applied to fruit positioning, fruit sorting, fruit yield prediction, fruit automatic picking and many other intelligent work in smart orchards.
[0003] In related technologies, a large number of sample images are manually labeled to obtain sample labeling information corresponding to each sample image. Based on a large number of sample images and corresponding sample labeling information, an image labeling model is trained, and the trained image labeling model can be called to process any target image in the future to obtain a labeled image corresponding to the target image. However, in the above method, a large number of sample images need to be manually labeled in order to train the image labeling model, which consumes manpower and time, resulting in low image labeling efficiency and thus low training efficiency of the image detection model. Therefore, although the fruit detection technology based on deep learning has been widely applied at present, it needs to rely on a large number of labeled data sets to support the training and learning of the detection model, resulting in high labor cost of manual labeling. Moreover, due to the poor generalization performance of the current deep learning model, the model needs to be independently trained and learned to produce a new target data set and a new detection model for application in different scenes, different environments, different shooting methods and different target types, which is time-consuming and labor-intensive.
[0004] Therefore, there is an urgent need for an automatic labeling method for a target data set with higher generalization and stronger domain adaptability. SUMMARY
[0005] In order to solve the problems existing in the prior art, the present application provides a target labeling method and system based on a multi-dimensional space feature model optimal source domain. The optimal source domain is determined based on a multi-dimensional space feature model, and a target data set of multiple target domains can be corresponded by a target data set of one source domain, so that the conversion generalization of the target image conversion model for different targets of different target domains is higher.
[0006] The first aspect of the present application provides a target labeling method based on a multi-dimensional space feature model optimal source domain, comprising:
[0007] S1, acquiring target domain foreground images of different categories;
[0008] S2, performing multi-dimensional spatial feature quantitative analysis based on the different categories of target domain foreground images and constructing a cross-category common description model based on the multi-dimensional spatial features after quantitative analysis; and obtaining the best source domain of the target based on the cross-category common description model;
[0009] S3, performing target labeling based on the best source domain of the target.
[0010] Preferably, the different categories of target domain foreground images can be images pre-stored by the computer device, images downloaded by the computer device from other devices, images uploaded by other devices to the computer device, or images currently collected by the computer device.
[0011] Preferably, the S2 comprises:
[0012] S21, extracting appearance features of the target from the different categories of target domain foreground images;
[0013] S22, abstracting the appearance features into specific shapes, colors and textures, and calculating relative distances of the specific shapes, colors and textures as analysis description sets of the appearance features of different target individuals based on a multi-dimensional feature quantitative analysis method for different target features;
[0014] S23, constructing a cross-category common description model based on multi-dimensional feature space reconstruction and feature difference division of the analysis description sets;
[0015] S24, obtaining the best source domain of the target based on the cross-category common description model.
[0016] Preferably, the S22 comprises:
[0017] S221, performing target shape extraction based on Fourier descriptors and discretizing the Fourier descriptors;
[0018] S222, extracting spatial distribution and proportion of Lab colors in the target foreground and drawing a CIELab spatial color distribution histogram;
[0019] S223, extracting pixel value gradient and direction derivative information of the target foreground to obtain texture information description based on LBP algorithm;
[0020] S224, performing single appearance feature relative distance calculation based on correlation and spatial distribution based on Fourier descriptor discretization, the drawn CIELab spatial color distribution histogram and the texture information description based on LBP algorithm;
[0021] S225, constructing a relative distance matrix based on the calculated single appearance feature relative distance values.
[0022] Preferably, the S23, constructing a cross-category commonality description model based on multi-dimensional feature space reconstruction and feature difference division of the analysis description set comprises:
[0023] S231, multi-dimensional feature space reconstruction: constructing a multi-dimensional feature space through the relative distance between each pair of target features, thereby converting the relative distance between different target features into absolute distance in the same feature space; the multi-dimensional feature space reconstruction adopts the MDS algorithm;
[0024] S232, feature difference division: dividing the feature points in the space based on a clustering algorithm, clustering according to the similarity of the target features, thereby obtaining the best source domain target in different clusters; the clustering algorithm used in the feature difference division is the DBSCAN algorithm.
[0025] Preferably, the S24, obtaining the best source domain of the target based on the cross-category commonality description model comprises:
[0026] S241, classifying different targets according to the appearance features based on the cross-category commonality description model;
[0027] S242, selecting the best source domain category from the classification according to the actual demand target domain category, inputting it into the generation model for target conversion, and optimizing the generation model to solve the problem that single category source domain background-free target image is difficult to convert into multi-category target domain background-free image, the optimization of the generation model comprises obtaining multi-modal images of multi-category target domain background-free targets through feature map extraction based on latent space and visual feature map extraction based on guided gradient information.
[0028] Preferably, the S3 comprises:
[0029] S31, constructing a target domain synthesis data set based on the converted images, comprising:
[0030] performing image preprocessing and image conversion on the target sample data to obtain target domain simulation targets and background to form a target domain image synthesis component;
[0031] establishing a knowledge graph system based on hierarchical component synthesis rules based on the target domain image synthesis component;
[0032] constructing a synthesized image based on the knowledge graph system based on hierarchical component synthesis rules;
[0033] recording the position information of the target, forming label data information together with the target domain synthesized image, and inputting them into a detection model for training and learning;
[0034] A synthetic dataset for the target domain is formed based on the trained and learned detection model;
[0035] S32, Training the detection model based on the target domain synthetic dataset and label data information;
[0036] S33 performs automatic target labeling based on a trained detection model.
[0037] A second aspect of the invention provides a target annotation system based on an optimal source domain using a multidimensional spatial feature model, comprising:
[0038] The image acquisition module is used to acquire foreground images of different categories of target domains;
[0039] The optimal source domain selection module is used to perform multi-dimensional spatial feature quantitative analysis based on the target domain foreground images of different categories and to construct a cross-category common description model based on the multi-dimensional spatial features after quantitative analysis; and to obtain the optimal source domain of the target based on the cross-category common description model.
[0040] The target annotation module is used to annotate targets based on their best source domain.
[0041] A third aspect of the present invention provides an electronic device including a processor and a memory, the memory storing a plurality of instructions, the processor being configured to read the instructions and execute the method as described in the first aspect.
[0042] A fourth aspect of the present invention provides a computer-readable storage medium storing a plurality of instructions which can be read by a processor and executed as described in the first aspect.
[0043] The target annotation method, system, electronic device, and computer-readable storage medium based on the optimal source domain of a multidimensional spatial feature model provided by this invention have the following beneficial technical effects:
[0044] To establish an automatic labeling method with higher generalization and domain adaptability, capable of meeting the needs of different types of fruit datasets; capable of automatically obtaining the labels of targets in the target domain, thus applying it to downstream smart agriculture projects; and significantly reducing the monetary and time costs incurred in manually labeling target boxes (compared to the existing technology where the average cost per labeling box for a single scene dataset is 0.2 yuan, with an average of 30 fruits per image, an average of 3 minutes of labeling time per image, and each dataset contains at least 10,000 images). Attached Figure Description
[0045] Figure 1 This is a flowchart of the target labeling method based on the optimal source domain of the multidimensional spatial feature model described in this invention.
[0046] Figure 2 The flow chart of the optimal source domain selection method for the corresponding multi-class target domain fruit growth of the application.
[0047] Figure 3 The flow chart of the multi-feature quantitative analysis in the optimal source domain selection of the application.
[0048] Figure 4 The relative distance matrix diagram of fruit features in the multi-dimensional feature quantitative analysis of the application.
[0049] Figure 5 The output result diagram of the cross-category commonality description model construction method of the application.
[0050] Figure 6 The flow chart of the target domain fruit data collection of the application.
[0051] Figure 7 The knowledge graph flow chart of the target domain fruit data collection of the application.
[0052] Figure 8 The target labeling system architecture diagram based on the optimal source domain of the multi-dimensional space feature model of the application.
[0053] Figure 9 The electronic device structure diagram of the application. DETAILED DESCRIPTION
[0054] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings in the specification and specific embodiments.
[0055] The method provided by the application can be implemented in a terminal environment, which can include one or more of the following components: a processor, a memory, and a display screen. The memory stores at least one instruction, which is loaded and executed by the processor to implement the method described in the embodiments below.
[0056] The processor can include one or more processing cores. The processor connects various parts of the entire terminal through various interfaces and lines, executes various functions of the terminal and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory, and calling data stored in the memory.
[0057] The memory can include random access memory (RAM) and read-only memory (ROM). The memory can be used to store instructions, programs, codes, code sets or instructions.
[0058] The display screen is used to display the user interface of each application program.
[0059] In addition, those skilled in the art can understand that the structure of the terminal described above does not constitute a limitation on the terminal, and the terminal can include more or fewer components, or combine certain components, or different component arrangements. For example, the terminal also includes radio frequency circuit, input unit, sensor, audio circuit, power supply and other components, which are not described here.
[0060] Embodiment one
[0061] Referring to Figure 1 The embodiment provides a target labeling method based on a multi-dimensional space feature model optimal source domain, which comprises the following steps:
[0062] S1, obtaining target domain foreground images of different categories;
[0063] In the embodiment, the target domain foreground images of different categories can be images pre-stored by the computer device, or images downloaded by the computer device from other devices, or images uploaded to the computer device by other devices, or the target domain foreground images of different categories can be images currently collected by the computer device, and the embodiment of the application does not limit this. For example, in the embodiment, the fruit labeling in the orchard is taken as a specific application scenario, and a high-definition camera device is used to assist the aerial photography of the unmanned aerial vehicle to obtain wide-area orchard images as target images. And the target image and the finally formed labeling image have the same size, for example, both are 96px*96px.
[0064] S2, performing multi-dimensional space feature quantitative analysis based on the target domain foreground images of different categories, and constructing a cross-category common description model based on the multi-dimensional space features after quantitative analysis; and obtaining the optimal source domain of the target based on the cross-category common description model.
[0065] As a preferred embodiment, S2 comprises:
[0066] S21, extracting appearance features of the target from the target domain foreground images of different categories, wherein the appearance features include but are not limited to edge contour, global color and local details, etc.
[0067] S22, abstracting the appearance features into specific shapes, colors and textures, and calculating the relative distances of the specific shapes, colors and textures as the analysis description set of the appearance features of different target individuals based on the multi-dimensional feature quantitative analysis method;
[0068] S23, constructing a cross-category common description model based on multi-dimensional feature space reconstruction and feature difference division on the analysis description set;
[0069] S24, obtaining the best source domain of the target based on the cross-category commonality description model.
[0070] In this embodiment, the best source domain selection module is used to design and analyze the phenotypic characteristics of different categories of fruits. The specific process of this module is shown in Figure 2 This module mainly calculates the commonality between its characteristics as prior knowledge of the deep learning algorithm, providing guidelines for data set selection and training parameter setting of deep learning. This module mainly includes two parts: first, a multi-dimensional feature quantitative analysis method is proposed to analyze and describe the appearance characteristics of different fruit individuals; second, a cross-category commonality description model is constructed to classify different fruits according to their phenotypic characteristics, and the best source domain fruit category is selected.
[0071] As shown in Figure 3 As a preferred embodiment, S22 abstracts the appearance characteristics into specific shapes, colors and textures, and calculates the relative distances of specific shapes, colors and textures based on the multi-dimensional feature quantitative analysis method for different target characteristics as the analysis description set of the appearance characteristics of different target individuals, including:
[0072] S221, shape extraction of the target (fruit in this embodiment) based on Fourier descriptor, and Fourier descriptor discretization;
[0073] S222, extracting the spatial distribution and proportion of Lab color in the foreground of the target (fruit in this embodiment), and drawing a CIELab space color distribution histogram;
[0074] S223, extracting the pixel value gradient and direction derivative information of the foreground of the target (fruit in this embodiment) to obtain texture information description based on LBP algorithm;
[0075] S224, correlation and spatial distribution based single appearance characteristic relative distance calculation based on Fourier descriptor discretization, the drawn CIELab space color distribution histogram and the texture information description based on LBP algorithm;
[0076] S225, constructing a relative distance matrix based on the calculated single appearance characteristic relative distance value.
[0077] As shown in Figure 4 The relative distance matrix of fruit characteristics in the orchard application scenario of this embodiment is shown in the figure.
[0078] As a preferred embodiment, S23, constructing a cross-category commonality description model based on multi-dimensional feature space reconstruction and feature difference division of the analysis description set includes:
[0079] S231, multi-dimensional feature space reconstruction: a multi-dimensional feature space is constructed by the relative distance between the features of each pair of targets (in this embodiment, fruits), so as to convert the relative distance between the features of different targets (in this embodiment, fruits) into absolute distance in the same feature space, and facilitate the simple and accurate description of the phenotype of each target (in this embodiment, fruit) image by the position coordinates of a pair of two-dimensional planes.
[0080] In this embodiment, the multi-dimensional feature space reconstruction adopts the MDS algorithm, which includes: projecting the points in the high-dimensional coordinates into the low-dimensional coordinates according to the distance, keeping the relative distance between the points in the high-dimensional coordinates and the points in the low-dimensional coordinates unchanged, and projecting the points in the low-dimensional coordinates into the two-dimensional plane space, and converting the relative distance into absolute distance. Of course, other algorithms can also be used by those skilled in the art, as long as the relative distance can be converted into absolute distance through coordinate projection and relative distance relationship, which is within the protection scope of the field.
[0081] S232, feature difference division: based on a clustering algorithm, the feature points in the space are divided, and clustering is performed according to the similarity of the features of the targets (in this embodiment, fruits), so as to obtain the best source domain targets (in this embodiment, fruits) in different clusters.
[0082] In this embodiment, the clustering algorithm used in the feature difference division is the DBSCAN algorithm, which includes: clustering according to the closeness of the samples in the multi-dimensional feature space, automatically dividing and selecting the categories and the number of source domain targets (in this embodiment, fruits); automatically determining the number of clusters according to the distribution difference of the features of the targets (in this embodiment, fruits), and the target (in this embodiment, fruit) species at the geometric center of each cluster is selected as the best source domain target (in this embodiment, fruit) species.
[0083] As a preferred embodiment, the S24, obtaining the best source domain of the target based on the cross-category commonality description model includes:
[0084] S241, classifying different targets according to the appearance features based on the cross-category commonality description model;
[0085] S242, for the target domain species of actual demand, selecting the best source domain species from the classification and inputting into the generation model for target conversion, and optimizing the generation model.
[0086] As Figure 5 The output result of the cross-category commonality description model construction method is shown in the figure, and different dashed boxes are different clusters, and the fruit solid line box is the best source domain in each cluster.
[0087] As a preferred embodiment, for step S242, since the most suitable source domain data is selected, sometimes the most suitable source domain cannot be selected (only one target or fruit in some clusters), so the generation model needs to be optimized, and realistic conversion can be realized when the shape, color and texture change greatly, and the domain difference is reduced.
[0088] The optimization of the generation model includes obtaining multi-class target domain background-free target multi-modal images through latent space-based feature map extraction and guided gradient information-based visual feature map extraction, thereby solving the problem of single-class best source domain background-free target image.
[0089] S3, target labeling based on the best source domain of the target.
[0090] As shown in Figure 6 As a preferred embodiment, S3 includes:
[0091] S31, constructing a target domain synthesis dataset based on the converted image, including:
[0092] Image preprocessing and image conversion are performed on the target sample data to obtain target domain simulation targets (fruits), backgrounds and other (leaves) components to form target domain image synthesis components;
[0093] A knowledge graph system based on hierarchical component synthesis rules is established based on the target domain image synthesis components; in this embodiment, the knowledge graph refers to a knowledge graph system based on hierarchical component synthesis rules, which is constructed by setting growth rules according to natural semantic structure, growth semantic structure and target domain background features for each component;
[0094] The hierarchical component synthesis rule-based knowledge graph system constructs a synthesized image;
[0095] The position information of the target (fruit) is recorded to form label data information, which is input into the detection model together with the target domain synthesis image for training and learning;
[0096] The target domain synthesis dataset is formed based on the trained and learned detection model.
[0097] S32, training a detection model based on the target domain synthesis dataset and label data information;
[0098] S33, automatic target labeling based on the trained detection model.
[0099] As a preferred embodiment, the hierarchical component synthesis rule-based knowledge graph system based on the target domain synthesis image makes the construction of the target domain synthesis dataset follow certain rules, including:
[0100] The construction process from components to scenes is formed by the composition rule based on natural semantics, the construction rule based on growth semantics and the domain adaptation rule based on scene environment.
[0101] In the embodiment, since the orchard scene is complex and the environment is changeable, it is very difficult to realize automatic data collection by completely relying on the random placement method, therefore, the method classifies each component according to different situations based on the structural and regular relationship between components, forms a knowledge graph based on the hierarchical structure of the orchard scene, and reasonably divides the synthesis weight between different components.
[0102] As shown in Figure 7 The domain adaptation rule based on scene environment forms the basic components of the orchard scene distribution, including land, sky, skeleton, leaves and fruits.
[0103] The component rule based on growth semantics forms the basic construction components of the growth trend of fruit trees (including trees and hidden fruits) and the combined components of the growth trend of fruit trees (including trees with fruits), wherein the hidden fruits are formed by the fruit domain adaptation sub-rule based on scene environment, the trees are formed by the skeleton and leaf component sub-rule based on growth semantics, and the trees with fruits are formed by the tree and hidden fruit component sub-rule based on growth semantics.
[0104] The composition rule based on natural semantics forms the orchard scene of natural semantic structure, wherein the trees with fruits, the sky and the land are finally formed into the target domain synthesis image by the domain adaptation rule based on scene environment and the composition rule based on natural semantics.
[0105] Embodiment two
[0106] Referring to Figure 8 A target labeling system based on a multi-dimensional space feature model best source domain, comprising: an image acquisition module 101 configured to acquire target domain foreground images of different categories; a best source domain selection module 102 configured to perform multi-dimensional space feature quantitative analysis based on the target domain foreground images of different categories and construct a cross-category common description model based on the quantitative analysis of the multi-dimensional space features; obtain the best source domain of the target based on the cross-category common description model; and a target labeling module 103 configured to perform target labeling based on the best source domain of the target.
[0107] The application further provides a memory storing a plurality of instructions for implementing the method according to the first embodiment.
[0108] As shown in Figure 9As shown, the application further provides an electronic device, comprising a processor 501 and a memory 502 connected with the processor 501, wherein the memory 502 stores a plurality of instructions, and the instructions can be loaded and executed by the processor to enable the processor to perform the method according to the embodiment one.
[0109] Although preferred embodiments of the application have been described herein, additional changes and modifications can be suggested to one skilled in the art, particularly in light of the essential novel teachings herein. The disclosures herein are thus intended to be illustrative, but not limiting, of the scope of the application. It is intended that the scope of the application shall be limited only to the extent of the patent issuing hereon. Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.
Claims
1. A target annotation method based on a multidimensional spatial feature model of the optimal source domain, characterized in that, include: S1, acquire foreground images of different categories of target domains; S2, perform multi-dimensional spatial feature quantitative analysis on the target domain foreground images of different categories, and construct a cross-category commonality description model based on the multi-dimensional spatial features after quantitative analysis; The optimal source domain for the target is obtained based on a cross-category commonality description model; S3, target annotation is performed based on the optimal source domain of the target; Wherein, S2 includes: S21, Extract the appearance features of the target from foreground images of different categories of target domains; S22, the appearance features are abstracted into specific shapes, colors and textures, and the relative distances of specific shapes, colors and textures are calculated for different target features based on the multidimensional feature quantitative analysis method as an analysis description set of appearance features of different target individuals; S23, Construct a cross-category common description model based on multi-dimensional feature space reconstruction and feature difference segmentation of the analysis description set; S24, Obtain the optimal source domain of the target based on the cross-category common description model; S23, which involves constructing a cross-category common description model based on multi-dimensional feature space reconstruction and feature difference segmentation of the analysis description set, includes: S231, Multidimensional Feature Space Reconstruction: A multidimensional feature space is constructed by using the relative distance between each pair of target features, thereby converting the relative distance between different target features into the absolute distance in the same feature space; the multidimensional feature space reconstruction adopts the MDS algorithm; S232, Feature Difference Division: Based on the clustering algorithm, feature points in the space are divided and clustered according to the similarity of target features to obtain the best source domain target in different clusters; the clustering algorithm used for the feature difference division is the DBSCAN algorithm.
2. The target annotation method based on the optimal source domain of a multidimensional spatial feature model according to claim 1, characterized in that, The different categories of target domain foreground images can be images pre-stored by the computer device, images downloaded by the computer device from other devices, images uploaded to the computer device from other devices, or images currently acquired by the computer device.
3. The target annotation method based on the optimal source domain of a multidimensional spatial feature model according to claim 1, characterized in that, In step S22, the appearance features are abstracted into specific shapes, colors, and textures. Based on a multidimensional feature quantitative analysis method, the relative distances of specific shapes, colors, and textures are calculated for different target features. This serves as an analysis and description set of appearance features for different target individuals, including: S221, target shape extraction is performed based on Fourier descriptors, and the Fourier descriptors are discretized; S222, extract the spatial distribution and proportion of Lab color in the target foreground, and draw a CIELab space color distribution histogram; S223, extract the pixel value gradient and directional derivative information of the target foreground to obtain texture information description based on LBP algorithm; S224, based on Fourier descriptor discretization, the plotted CIELab space color distribution histogram, and texture information description based on LBP algorithm, calculates the relative distance of a single appearance feature based on correlation and spatial distribution. S225, construct a relative distance matrix based on the calculated relative distance values of a single appearance feature.
4. The target annotation method based on the optimal source domain of a multidimensional spatial feature model according to claim 1, characterized in that, In step S24, the optimal source domain for obtaining the target based on the cross-category commonality description model includes: S241, classify different targets according to the appearance features based on the cross-category commonality description model; S242, for the target domain types required in practice, select the best source domain type from the classification, input it into the generation model for target transformation, and optimize the generation model. The optimization of the generation model includes obtaining multimodal images of multi-class target domains without background targets by extracting feature maps based on latent space and extracting visual feature maps based on guided gradient information.
5. The target annotation method based on the optimal source domain of a multidimensional spatial feature model according to claim 1, characterized in that, S3 includes: S31, Construct a target domain synthetic dataset based on the transformed images, including: The target sample data is preprocessed and transformed to obtain the target domain simulation target and background to form a target domain image synthesis component. A knowledge graph system based on hierarchical component synthesis rules is established based on target domain image synthesis components; Constructing synthesized images based on a knowledge graph system of hierarchical component synthesis rules; Record the location information of the target, and input it together with the target domain synthetic image into the detection model for training and learning; A synthetic dataset for the target domain is formed based on the trained and learned detection model; S32, Training the detection model based on the target domain synthetic dataset and label data information; S33 performs automatic target labeling based on a trained detection model.
6. A target labeling system based on an optimal source domain using a multidimensional spatial feature model, used to implement the method described in any one of claims 1-5, characterized in that, include: The image acquisition module is used to acquire foreground images of different categories of target domains; The optimal source domain selection module is used to perform multi-dimensional spatial feature quantitative analysis based on the target domain foreground images of different categories and to construct a cross-category common description model based on the multi-dimensional spatial features after quantitative analysis; and to obtain the optimal source domain of the target based on the cross-category common description model. The target annotation module is used to annotate targets based on their best source domain.
7. An electronic device, characterized in that, It includes a processor and a memory, the memory storing multiple instructions, and the processor being used to read the instructions and execute the method as described in any one of claims 1-5.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions, which can be read by a processor and executed as described in any one of claims 1-5.