Multi-source heterogeneous data fusion processing method
By extracting the texture and edge features of optical images and synthesized aperture radar images, using the channel attention mechanism and image fusion model, the problem of low efficiency and accuracy of multi-source heterogeneous data fusion processing is solved, and high-quality image fusion is achieved and enhanced Feature representation and noise immunity.
Patent Information
- Application Number
- CN202411249136.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-12-13
AI Technical Summary
In the prior art, the efficiency and accuracy of multi-source heterogeneous data fusion processing are poor, are susceptible to noise, and it is difficult to ensure the quality of image fusion.
By determining the multi-source heterogeneous data of the target object, including optical images and synthetic aperture radar images, extracting texture features and edge feature maps, using the channel attention mechanism for feature processing, and fusing through the trained image fusion model to generate the target Fusion images.
The fusion efficiency and accuracy of optical images and synthetic aperture radar images are improved, the feature integrity and feature representation capabilities of the fusion image are ensured, and the anti-noise performance is enhanced.
Smart Images
Figure CN119131540A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer application technology, and in particular to a multi-source heterogeneous data fusion processing method. Background Art
[0002] With the development of image technology, various images are widely used in scenes such as target detection and identification and positioning. Single modality images have their own advantages and limitations. For example, optical images have high resolution and rich spectral information, but are easily disturbed by external factors; synthetic aperture radar images are not easily disturbed by external factors, but image interpretation is difficult and lacks spectral information. At present, multi-source heterogeneous images are usually fused to take advantage of the complementary advantages of data to provide a basis for subsequent tasks.
[0003] In related technologies, multi-source heterogeneous data are usually fused in pixel dimension through weighted averaging and wavelet transform, but the fusion process is easily affected by noise. Usually, it is difficult to ensure the efficiency and accuracy of multi-source heterogeneous data fusion processing. Summary of the invention
[0004] The present invention provides a multi-source heterogeneous data fusion processing method to solve the technical problem of poor efficiency and accuracy of multi-source heterogeneous data fusion processing in the prior art.
[0005] According to one aspect of the present invention, a method for fusion processing of multi-source heterogeneous data is provided, wherein the method comprises:
[0006] Determining multi-source heterogeneous data to be fused corresponding to the target object, wherein the multi-source heterogeneous data includes a target optical image and a synthetic aperture radar image;
[0007] Determine a first texture feature map and a first edge feature map corresponding to the target optical image, and determine a second texture feature map and a second edge feature map corresponding to the synthetic aperture radar image;
[0008] Determine a texture feature fusion map according to the first texture feature map and the second texture feature map, and determine an edge feature fusion map according to the first edge feature map and the second edge feature map;
[0009] The texture feature fusion map and the edge feature fusion map are respectively processed by a channel attention mechanism to obtain a target texture fusion map and a target edge fusion map in turn;
[0010] The input target texture fusion image and the target edge fusion image are fused through the trained image fusion model to obtain a target fusion image corresponding to the target object.
[0011] The technical solution of the embodiment of the present invention is to determine the multi-source heterogeneous data to be fused corresponding to the target object, wherein the multi-source heterogeneous data includes the target optical image and the synthetic aperture radar image; determine the first texture feature map and the first edge feature map corresponding to the target optical image, determine the second texture feature map and the second edge feature map corresponding to the synthetic aperture radar image; determine the texture feature fusion map according to the first texture feature map and the second texture feature map, and determine the edge feature fusion map according to the first edge feature map and the second edge feature map; perform feature processing on the texture feature fusion map and the edge feature fusion map respectively through the channel attention mechanism to obtain the target texture fusion map and the target edge fusion map in turn; fuse the input target texture fusion map and the target edge fusion map through the trained image fusion model to obtain the target fusion image corresponding to the target object. The technical solution of the present invention can improve the fusion efficiency and accuracy of optical images and synthetic aperture radar images, ensure the feature integrity of the determined fused image, and improve the feature representation capability of the determined fused image.
[0012] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0014] Figure 1 is a flowchart of a multi-source heterogeneous data fusion processing method provided according to the first embodiment of the present invention;
[0015] Figure 2 is a flow chart of performing feature enhancement on a feature map through an efficient channel attention network according to an embodiment of the present invention;
[0016] Figure 3 is a flow chart of performing image fusion through an image fusion model provided according to an embodiment of the present invention;
[0017] Figure 4 A flowchart of a multi-source heterogeneous data fusion processing method provided in the second embodiment of the present invention;
[0018] Figure 5 An overall flow chart of a multi-source heterogeneous data fusion processing method provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0019] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0020] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0021] Embodiment 1
[0022] Figure 1 A flowchart of a multi-source heterogeneous data fusion processing method is provided for the first embodiment of the present invention. This embodiment is applicable to the case where feature-level fusion of optical images and synthetic aperture radar images is performed, such as Figure 1 As shown, the method includes:
[0023] S110 , determining multi-source heterogeneous data to be fused corresponding to the target object, wherein the multi-source heterogeneous data includes a target optical image and a synthetic aperture radar image.
[0024] The target object may be understood as an object to be analyzed based on the fused image. In the embodiment of the present invention, the target image may be preset according to the scene requirements, and is not specifically limited here. Optionally, the target object may include a target region or a ground building.
[0025] The multi-source heterogeneous data can be understood as data from multiple sources and with different structures.
[0026] The target optical image can be understood as an image obtained by an optical method. Optionally, the target optical image can include infrared imaging, laser interference image, holographic image, speckle image, etc.
[0027] The synthetic aperture radar image may be understood as an image generated by synthetic aperture radar imaging (Synthetic Aperture Radar, SAR) technology.
[0028] S120: Determine a first texture feature map and a first edge feature map corresponding to the target optical image, and determine a second texture feature map and a second edge feature map corresponding to the synthetic aperture radar image.
[0029] The first texture feature map may represent the texture features of the target optical image.
[0030] The first edge feature map may characterize edge features of the target optical image.
[0031] The second texture feature map may represent texture features of the synthetic aperture radar image.
[0032] The second edge feature map may be understood as representing edge features of the synthetic aperture radar image.
[0033] Specifically, texture feature extraction is performed on the target optical image to obtain a first texture feature, and the first texture feature map corresponding to the first texture feature is determined; edge feature extraction is performed on the target optical image to obtain a first edge feature, and the first edge feature map corresponding to the first edge feature is determined.
[0034] Correspondingly, texture feature extraction is performed on the synthetic aperture radar image to obtain a second texture feature, and the second texture feature map corresponding to the second texture feature is determined; edge feature extraction is performed on the synthetic aperture radar image to obtain a second edge feature, and the second edge feature map corresponding to the second edge feature is determined.
[0035] S130: Determine a texture feature fusion map according to the first texture feature map and the second texture feature map, and determine an edge feature fusion map according to the first edge feature map and the second edge feature map.
[0036] The texture feature fusion map may be understood as a feature map that fuses the texture features of the target optical image and the texture features of the synthetic aperture radar image.
[0037] The edge feature fusion map can be understood as a feature map that fuses the edge features of the target optical image and the edge features of the synthetic aperture radar image.
[0038] Specifically, the first texture feature map and the second texture feature map are integrated to obtain the texture feature fusion map. More specifically, a first texture feature matrix corresponding to the first texture feature map and a second texture feature matrix corresponding to the second texture feature map are matrix-superimposed to determine the texture feature fusion map. Correspondingly, the specific determination method of the edge feature fusion map is the same as that of the texture feature fusion map, which is not described in detail here.
[0039] S140. Perform feature processing on the texture feature fusion map and the edge feature fusion map respectively through a channel attention mechanism to obtain a target texture fusion map and a target edge fusion map in turn.
[0040] The channel attention mechanism can be understood as a technology that can automatically learn and focus on important features in an image, and can emphasize important features in an image and suppress minor features. Optionally, the channel attention mechanism can include an efficient channel attention network (ECA-Net).
[0041] The target texture fusion map can be understood as the texture feature fusion map with enhanced feature representation capability. Correspondingly, the target edge fusion map is the edge feature fusion map with enhanced feature representation capability.
[0042] Optionally, the performing feature processing on the texture feature fusion map through a channel attention mechanism to obtain a target texture fusion map includes:
[0043] Determine a plurality of feature matrix channels corresponding to the texture feature fusion map, and for each of the feature matrix channels, perform global average pooling on the feature matrix channel to obtain a global feature;
[0044] Performing local cross-channel interaction on the multiple global features to obtain the attention weight of each feature matrix channel;
[0045] A target texture fusion map corresponding to the texture feature fusion map is determined according to the attention weight and the feature matrix channel.
[0046] The feature matrix channel can be understood as the channel of the feature matrix corresponding to the texture feature fusion map. In the embodiment of the present invention, the number of the feature matrix channels is related to the dimension of the feature matrix corresponding to the texture feature fusion map, which is not specifically limited here.
[0047] The global feature can be understood as a feature that characterizes the overall attribute of the image. Optionally, the global feature can include color distribution or shape contour, etc.
[0048] The attention weight can be understood as the weight of the feature matrix channel determined based on the attention mechanism.
[0049] Figure 2 1 is a flow chart of performing feature enhancement on a feature map through an efficient channel attention network according to an embodiment of the present invention. Figure 2 As shown, feature enhancement of feature maps through efficient channel attention networks can be: feature extraction of texture feature fusion maps and edge feature fusion maps through ECA-Net respectively. Specifically, first, global average pooling is performed on each channel of the input feature fusion map to obtain the global features of each channel; further, a one-dimensional convolutional layer is used to perform local cross-channel interaction on multiple global features to generate channel attention weights; finally, the weights are multiplied by the feature fusion map of the original input to obtain a feature map enhanced by the channel attention mechanism. In an embodiment of the present invention, ECA-Net can flexibly capture the dependencies between channels, enhance the feature representation capability of images, and has the advantages of being lightweight and efficient.
[0050] S150, fusing the input target texture fusion image and the target edge fusion image through the trained image fusion model to obtain a target fusion image corresponding to the target object.
[0051] The image fusion model can be understood as a model with image fusion function. In an embodiment of the present invention, the input data of the image fusion model can be the target texture fusion map and the target edge fusion map. The output data can be the target fusion image completed by fusing the texture features in the target texture fusion map and the edge features in the target edge fusion map.
[0052] The target fusion image can be understood as a fusion image corresponding to the target optical image and the synthetic aperture radar image. In an embodiment of the present invention, the target fusion image includes both the texture features and edge features of the target optical image and the texture features and edge features of the synthetic aperture radar image.
[0053] Optionally, the trained image fusion model fuses the input target texture fusion image and the target edge fusion image to obtain a target fusion image corresponding to the target object, including:
[0054] Performing data division on the target texture fusion map and the target edge fusion map respectively, to obtain a plurality of first data sub-maps and a plurality of second data sub-maps in sequence;
[0055] Determine a first vector sequence corresponding to a plurality of the first data subgraphs, and determine a second vector sequence corresponding to a plurality of the second data subgraphs;
[0056] The first vector sequence and the second vector sequence are interactively processed through a multi-head self-attention mechanism to obtain a target fused image corresponding to the target object.
[0057] The first data sub-image can be understood as a partial data image of the target texture fusion image (refer to Figure 3 ). The first vector sequence can be understood as a vector sequence corresponding to the target texture fusion map. Correspondingly, the second data sub-map can be understood as a partial data map of the target edge fusion map (refer to Figure 3 ). The second vector sequence can be understood as a vector sequence corresponding to the target edge fusion graph.
[0058] Optionally, the determining of the first vector sequences corresponding to the plurality of first data subgraphs includes:
[0059] For each of the first data subgraphs, determining a first preliminary vector corresponding to the first data subgraph, and determining spatial information of the first data subgraph, and determining a first spatial vector of the first data subgraph according to the first preliminary vector and the spatial information;
[0060] A first vector sequence corresponding to a plurality of the first data subgraphs is determined based on the first space vector.
[0061] The first preliminary vector may be understood as a vector containing the fused texture features in the first data sub-image. The spatial information may represent the spatial position of the first data sub-image in the overall target texture fusion image. The first spatial vector may be understood as a vector obtained by adding the spatial information to the first preliminary vector.
[0062] Figure 3 FIG. 1 is a flow chart of performing image fusion through an image fusion model according to an embodiment of the present invention. Figure 3 As shown, specifically, the process of image fusion through the image fusion model can be: the feature map output by ECA-Net is further processed by a Transformer-based visual encoder (ViT) to complete the feature-level fusion of two image data (the target texture fusion map and the target edge fusion map), and the image data is processed by a decoder and an activation function to obtain the final target fusion image. The structure of the ViT encoder can refer to Figure 3. Further detailed specific steps may be, by dividing the input data (i.e., two-way images) into fixed-size data blocks (i.e., data subgraphs), flattening each block into a vector, and embedding through linear transformation, and then adding position encoding to retain spatial information; further, the obtained multiple embedded vector sequences are input into a standard Transformer encoder, and interactively processed through a multi-head self-attention mechanism to obtain a target fused image. In an embodiment of the present invention, the multi-head self-attention mechanism of the ViT encoder can better capture the global features of the image and the long-distance dependencies in the image, for example, the texture and structural information in the SAR image, which can improve the accuracy of image fusion.
[0063] The technical solution of the embodiment of the present invention is to determine the multi-source heterogeneous data to be fused corresponding to the target object, wherein the multi-source heterogeneous data includes the target optical image and the synthetic aperture radar image; determine the first texture feature map and the first edge feature map corresponding to the target optical image, determine the second texture feature map and the second edge feature map corresponding to the synthetic aperture radar image; determine the texture feature fusion map according to the first texture feature map and the second texture feature map, and determine the edge feature fusion map according to the first edge feature map and the second edge feature map; perform feature processing on the texture feature fusion map and the edge feature fusion map respectively through the channel attention mechanism to obtain the target texture fusion map and the target edge fusion map in turn; fuse the input target texture fusion map and the target edge fusion map through the trained image fusion model to obtain the target fusion image corresponding to the target object. The technical solution of the present invention can improve the fusion efficiency and accuracy of optical images and synthetic aperture radar images, ensure the feature integrity of the determined fused image, and improve the feature representation capability of the determined fused image.
[0064] Embodiment 2
[0065] Figure 4 This is a flowchart of a multi-source heterogeneous data fusion processing method provided in the second embodiment of the present invention. This embodiment is to fuse the input target texture fusion image and the target edge fusion image by the image fusion model completed through training described in the above embodiment. Figure 4 As shown, the method includes:
[0066] S210: Acquire a sample optical image and a sample radar image, and determine a sample texture fusion map corresponding to the sample optical image and a sample edge fusion map corresponding to the sample radar image.
[0067] In the embodiment of the present invention, the image fusion model may be obtained by training an original fusion model based on sample optical images and sample radar images.
[0068] The sample optical image and the sample radar image may be understood as sample images used to train the image fusion model.
[0069] Optionally, determining the sample texture fusion map corresponding to the sample optical image and the sample edge fusion map corresponding to the sample radar image may include:
[0070] Determine a first texture sample map and a first edge sample map corresponding to the sample optical image, and determine a second texture sample map and a second edge sample map corresponding to the sample radar image;
[0071] A sample texture fusion map is determined according to the first texture sample map and the second texture sample map, and a sample edge fusion map is determined according to the first edge sample map and the second edge sample map.
[0072] S220, determining an original fusion model, and fusing the input sample texture fusion map and the input sample edge fusion map through the original fusion model to obtain a sample fusion result.
[0073] The original fusion model may be understood as a fusion model with default model parameters, or may be understood as a fusion model without model parameter adjustment.
[0074] The sample fusion result can be understood as a result obtained by fusing the sample texture fusion map and the sample edge fusion map through the original fusion model. Optionally, the sample fusion result may include at least one of a sample fusion image, an optical reconstruction image, and a radar reconstruction image.
[0075] The sample fusion image may be understood as a fusion image corresponding to the sample optical image and the sample radar image determined by the original fusion model.
[0076] The optical reconstructed image can be understood as a reconstructed optical model, and the optical reconstructed image corresponds to the sample optical image. The radar reconstructed image can be understood as a reconstructed radar model, and the radar reconstructed image corresponds to the sample radar image.
[0077] In the present invention, the output data of the original fusion model can be divided into three paths and passed through the Transformer to generate a sample fusion image of two paths of data, an optical reconstruction image, and a radar reconstruction image. The structure of the Transformer here is consistent with that of the ViT encoder.
[0078] S230: Adjust the model parameters of the original fusion model based on the sample fusion result, and determine the image fusion model based on the adjusted model parameters.
[0079] Optionally, adjusting the model parameters of the original fusion model based on the sample fusion result includes:
[0080] determining a first loss value based on the optically reconstructed image and the sample optical image, and determining a second loss value based on the radar reconstructed image and the sample radar image;
[0081] A target loss value is determined according to the first loss value and the second loss value, and model parameters of the original fusion model are adjusted based on the target loss value.
[0082] The first loss value may represent the loss between the optically reconstructed image and the sample optical image. The second loss value may represent the loss between the radar reconstructed image and the sample radar image. The target loss value may represent the loss of the sample fusion result.
[0083] In the embodiment of the present invention, the reconstruction of the optical image and the SAR image can ensure that the original fusion model fully learns the features of the two modal input images, so as to improve the accuracy of the trained image fusion model.
[0084] S240: Determine multi-source heterogeneous data to be fused corresponding to the target object, wherein the multi-source heterogeneous data includes a target optical image and a synthetic aperture radar image.
[0085] S250: Determine a first texture feature map and a first edge feature map corresponding to the target optical image, and determine a second texture feature map and a second edge feature map corresponding to the synthetic aperture radar image.
[0086] S260: Determine a texture feature fusion map according to the first texture feature map and the second texture feature map, and determine an edge feature fusion map according to the first edge feature map and the second edge feature map.
[0087] S270, performing feature processing on the texture feature fusion map and the edge feature fusion map respectively through a channel attention mechanism to obtain a target texture fusion map and a target edge fusion map in turn.
[0088] S280, fusing the input target texture fusion image and the target edge fusion image through the trained image fusion model to obtain a target fusion image corresponding to the target object.
[0089] The technical solution of the embodiment of the present invention obtains a sample optical image and a sample radar image, and determines a sample texture fusion map corresponding to the sample optical image and a sample edge fusion map corresponding to the sample radar image; determines an original fusion model, and fuses the input sample texture fusion map and the sample edge fusion map through the original fusion model to obtain a sample fusion result; adjusts the model parameters of the original fusion model based on the sample fusion result, and determines the image fusion model based on the adjusted model parameters, thereby improving the image fusion accuracy of the trained image fusion model.
[0090] Optional, Figure 5 The overall flow chart of a multi-source heterogeneous data fusion processing method provided by an embodiment of the present invention. Before elaborating on the specific steps, the terms are first uniformly defined, wherein texture information can represent a texture feature map, edge information can represent an edge feature map, fused texture information can represent a texture feature fusion map, fused edge information can represent an edge feature fusion map, channel attention module can represent a channel attention mechanism, fusion unit can represent a fusion model, and Transformer module can represent a Transformer model. Based on the above uniform terminology definition, specifically, the overall process of the multi-source heterogeneous data fusion processing method can be: first, the input optical image and SAR image are respectively split into edge information and texture information, and each is preliminarily merged to obtain fused texture information and fused edge information; the channel attention module is used to perform preliminary feature extraction; the image is fused through the ViT encoder (by capturing the long thread dependency of the image data for further feature processing), the decoder (data fusion through the Transformer Block) and the activation function of the fusion unit; the three-way output is determined through the Transformer model, the first way is the fusion output, the second way is the reconstruction result of the optical data, and the third way is the reconstruction result of the SAR data.
[0091] On the other hand, the present invention designs a set of loss functions based on image reconstruction. Specifically, the image reconstruction loss of two-way data is used to constrain the learning of the fusion model on the features of multi-source data, that is, through the training of the model, the loss value of image reconstruction is continuously reduced, so that the reconstructed optical image is as close as possible to the sample optical image, and the reconstructed SAR image is as close as possible to the original SAR image, thereby ensuring that the fusion model learns sufficient feature information and improving the image fusion accuracy of the obtained fusion model.
[0092] The technical solution of the present invention has better anti-noise performance, lower computational complexity and higher flexibility, and can enable the determined target fusion image to have stronger information expression capability.
Claims
1. A multi-source heterogeneous data fusion processing method, characterized in that: include: Determining multi-source heterogeneous data to be fused corresponding to the target object, wherein the multi-source heterogeneous data includes a target optical image and a synthetic aperture radar image; Determine a first texture feature map and a first edge feature map corresponding to the target optical image, and determine a second texture feature map and a second edge feature map corresponding to the synthetic aperture radar image; Determine a texture feature fusion map according to the first texture feature map and the second texture feature map, and determine an edge feature fusion map according to the first edge feature map and the second edge feature map; The texture feature fusion map and the edge feature fusion map are respectively processed by a channel attention mechanism to obtain a target texture fusion map and a target edge fusion map in turn; The input target texture fusion image and the target edge fusion image are fused through the trained image fusion model to obtain a target fusion image corresponding to the target object.
2. The method according to claim 1, characterized in that Before the trained image fusion model fuses the input target texture fusion image and the target edge fusion image, the method further includes: Acquire a sample optical image and a sample radar image, and determine a sample texture fusion map corresponding to the sample optical image and a sample edge fusion map corresponding to the sample radar image; Determine an original fusion model, and fuse the input sample texture fusion map and the sample edge fusion map through the original fusion model to obtain a sample fusion result; The model parameters of the original fusion model are adjusted based on the sample fusion result, and the image fusion model is determined based on the adjusted model parameters.
3. The method according to claim 2, characterized in that The sample fusion result includes at least one of a sample fusion image, an optical reconstruction image, and a radar reconstruction image.
4. The method according to claim 3, characterized in that The adjusting the model parameters of the original fusion model based on the sample fusion result includes: determining a first loss value based on the optically reconstructed image and the sample optical image, and determining a second loss value based on the radar reconstructed image and the sample radar image; A target loss value is determined according to the first loss value and the second loss value, and model parameters of the original fusion model are adjusted based on the target loss value.
5. The method according to claim 1, characterized in that The step of performing feature processing on the texture feature fusion map through a channel attention mechanism to obtain a target texture fusion map includes: Determine a plurality of feature matrix channels corresponding to the texture feature fusion map, and for each of the feature matrix channels, perform global average pooling on the feature matrix channel to obtain a global feature; Performing local cross-channel interaction on the multiple global features to obtain the attention weight of each feature matrix channel; A target texture fusion map corresponding to the texture feature fusion map is determined according to the attention weight and the feature matrix channel.
6. The method according to claim 1, characterized in that The trained image fusion model fuses the input target texture fusion image and the target edge fusion image to obtain a target fusion image corresponding to the target object, including: Performing data division on the target texture fusion map and the target edge fusion map respectively, to obtain a plurality of first data sub-maps and a plurality of second data sub-maps in sequence; Determine a first vector sequence corresponding to a plurality of the first data subgraphs, and determine a second vector sequence corresponding to a plurality of the second data subgraphs; The first vector sequence and the second vector sequence are interactively processed through a multi-head self-attention mechanism to obtain a target fused image corresponding to the target object.
7. The method according to claim 6, characterized in that The determining of the first vector sequences corresponding to the plurality of the first data subgraphs comprises: For each of the first data subgraphs, determining a first preliminary vector corresponding to the first data subgraph, and determining spatial information of the first data subgraph, and determining a first spatial vector of the first data subgraph according to the first preliminary vector and the spatial information; A first vector sequence corresponding to a plurality of the first data subgraphs is determined based on the first space vector.