Image registration method and device based on deep learning, equipment and medium

Abstract

The invention relates to the related technical field of image registration, in particular to an image registration method and device based on deep learning, equipment and a medium. The method comprises the following steps: acquiring an original image and a target image; registering the original image and the target image by using a preset deep learning model to obtain a target registered image; wherein the deep learning model carries out registration of the original image and the target image based on a bidirectional cross-correlation operation algorithm. According to the scheme provided by the invention, image registration is carried out based on the deep learning model, so that the registration efficiency can be improved. And the influence of manual work on the registration precision is avoided, and the registration precision is improved.

Description

Technical Field

[0001] This application relates to the field of image registration technology, specifically to an image registration method, apparatus, device, and medium based on deep learning. Background Technology

[0002] With my country's rapid economic development, the scale and speed of power grid construction are among the world's leading levels. Currently, the total length of transmission lines in my country's six major power grids—South China, Northwest China, East China, Central China, North China, and Northeast China—exceeds 1.15 million kilometers, with 500kV and above transmission lines forming the mainstay of power transmission in each region. However, most of these transmission lines are located in remote mountainous areas, characterized by numerous points of distribution and vast areas. These lines are exposed to harsh natural environments for extended periods, making them prone to safety hazards such as breakage, wear, and corrosion. Therefore, regular inspections and surveys of transmission lines to understand their on-site conditions in real time have become crucial.

[0003] Traditional power line inspections primarily rely on manual surveys, with power workers inspecting and maintaining high-voltage electrical components such as power towers, transmission lines, and insulators on-site to ensure the normal operation of the power supply system. However, when conducting inspections in remote mountainous areas, inspectors can only observe distant objects through binoculars or climb into the mountainous towers for inspection. This manual inspection method is inefficient, labor-intensive, and unsafe, and also suffers from blind spots that make inspection difficult.

[0004] However, during the inspection process, the quality of the inspection images is often affected by factors such as bad weather, relative motion, and imaging equipment shaking, resulting in low image quality, low resolution, and image target occlusion. Therefore, it is necessary to take images from multiple angles and then perform multi-modal image multi-angle registration.

[0005] However, existing registration methods have low registration efficiency and require manually extracted features or local descriptors for registration. The actual registration effect is greatly affected by the accuracy of manually extracted features, resulting in low registration precision. Summary of the Invention

[0006] In view of this, embodiments of this application aim to provide an image registration method, apparatus, device, and medium based on deep learning to solve or at least alleviate the above-mentioned problems to some extent.

[0007] The first aspect of this application provides a deep learning-based image registration method, including:

[0008] Acquire the original image and the target image;

[0009] Using a pre-defined deep learning model, the original image and the target image are registered to obtain a target registered image;

[0010] The deep learning model is used to register the original image and the target image using an algorithm based on bidirectional cross-correlation.

[0011] In some embodiments, a preset deep learning model is used to register the original image and the target image to obtain a target registered image;

[0012] Based on a first preset scaling factor, the original image and the target image are downsampled to obtain two single-sampled images;

[0013] The sampled image is input into a preset deep learning model to obtain a registration result, and a registered image is obtained based on the registration result, the original image, and the target image.

[0014] Based on the second preset scale factor, the primary registration image and the target image are downsampled to obtain two secondary sampled images;

[0015] The secondary sampled image is input into a preset deep learning model to obtain a secondary registration result, and a secondary registered image is obtained based on the secondary registration result, the original image, and the target image;

[0016] Based on three preset scale factors, the secondary registration image and the target image are downsampled to obtain two cubically sampled images;

[0017] The subsampled image is input into a preset deep learning model to obtain a three-stage registration result. Based on the three-stage registration result, the original image, and the target image, a target registered image is obtained.

[0018] In some embodiments, the deep learning model is used for:

[0019] Feature extraction is performed on the two input images to obtain the features;

[0020] Based on the aforementioned features, bidirectional registration is performed to obtain the registration result.

[0021] In some embodiments, feature extraction is performed on the two input images to obtain features, including:

[0022] Feature extraction is performed on the two input images to obtain the first feature and the second feature.

[0023] In some embodiments, bidirectional registration based on the features is used to obtain a registration result, including:

[0024] The first mapping relationship from the first feature to the second feature is determined based on a pre-defined neural network.

[0025] The second mapping relationship from the second feature to the first feature is determined based on a pre-defined neural network.

[0026] Based on a preset parameter regression network, the first transformation parameter corresponding to the first mapping relationship is determined, and the second transformation parameter corresponding to the second mapping relationship is determined.

[0027] The final parameters are obtained by weighted summation of the first and second transformation parameters.

[0028] Based on the final parameters, the registration result is determined.

[0029] In some embodiments, feature extraction is performed on the two input images to obtain features, including:

[0030] Based on the channel attention mechanism, features are extracted from the two images, namely the third and fourth features.

[0031] In some embodiments, bidirectional registration based on the features is used to obtain a registration result, including:

[0032] The third mapping relationship between the third feature and the fourth feature is determined based on a pre-defined neural network.

[0033] The fourth mapping relationship between the fourth feature and the third feature is determined based on a pre-defined neural network.

[0034] Based on a preset parameter regression network, the third transformation parameter corresponding to the third mapping relationship is determined, and the fourth transformation parameter corresponding to the fourth mapping relationship is determined.

[0035] The final parameters are obtained by weighted summation of the third and fourth transformation parameters.

[0036] Based on the final parameters, the registration result is determined.

[0037] A second aspect of this application provides a deep learning-based image registration apparatus, comprising:

[0038] The acquisition module is used to acquire the original image and the target image;

[0039] The registration module is used to register the original image and the target image using a preset deep learning model to obtain a target registered image;

[0040] The deep learning model is used to register the original image and the target image using an algorithm based on bidirectional cross-correlation.

[0041] A third aspect of this application provides an electronic device, comprising:

[0042] A processor, and a memory for storing a processor-executable program;

[0043] The processor is configured to implement the deep learning-based image registration method described above by running a program in the memory.

[0044] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when run by a processor, causes the processor to perform the deep learning-based image registration method described above.

[0045] This application provides a deep learning-based image registration method, comprising: acquiring an original image and a target image; and registering the original image and the target image using a preset deep learning model to obtain a target registered image; wherein the deep learning model is an algorithm based on bidirectional cross-correlation operation for registering the original image and the target image. Thus, in the solution provided by this application, image registration based on a deep learning model not only improves registration efficiency but also eliminates the influence of manual intervention on registration accuracy, thereby improving registration precision. Attached Figure Description

[0046] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0047] Figure 1a This is a flowchart illustrating an image registration method based on deep learning provided in one embodiment of this application.

[0048] Figure 1b This is a flowchart provided in one embodiment of this application.

[0049] Figure 2 This is a schematic diagram of the structure of a model provided in one embodiment of this application.

[0050] Figure 3 This is a schematic diagram of the structure of a model provided in one embodiment of this application.

[0051] Figure 4 This is a schematic diagram of the structure of an image registration device based on deep learning provided in one embodiment of this application.

[0052] Figure 5 This is a schematic diagram of an electronic device structure provided in one embodiment of this application. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Application Overview

[0055] With my country's rapid economic development, the scale and speed of power grid construction are among the world's leading levels. Currently, the total length of transmission lines in my country's six major power grids—South China, Northwest China, East China, Central China, North China, and Northeast China—exceeds 1.15 million kilometers, with 500kV and above transmission lines forming the mainstay of power transmission in each region. However, most of these transmission lines are located in remote mountainous areas, characterized by numerous points of distribution and vast areas. These lines are exposed to harsh natural environments for extended periods, making them prone to safety hazards such as breakage, wear, and corrosion. Therefore, regular inspections and surveys of transmission lines to understand their on-site conditions in real time have become crucial.

[0056] Traditional power line inspections primarily rely on manual surveys, with power workers inspecting and maintaining high-voltage electrical components such as power towers, transmission lines, and insulators on-site to ensure the normal operation of the power supply system. However, when conducting inspections in remote mountainous areas, inspectors can only observe distant objects through binoculars or climb into the mountainous towers for inspection. This manual inspection method is inefficient, labor-intensive, and unsafe, and also suffers from blind spots that make inspection difficult.

[0057] However, during the inspection process, the quality of the inspection images is often affected by factors such as bad weather, relative motion, and imaging equipment shaking, resulting in low image quality, low resolution, and image target occlusion. Therefore, it is necessary to take images from multiple angles and then perform multi-modal image multi-angle registration.

[0058] However, existing registration methods have low registration efficiency and require manually extracted features or local descriptors for registration. The actual registration effect is greatly affected by the accuracy of manually extracted features, resulting in low registration precision.

[0059] To address the aforementioned problems, this application provides a solution that first acquires an original image and a target image; then, using a pre-defined deep learning model, registers the original image and the target image to obtain a target registered image; wherein the deep learning model is an algorithm based on bidirectional cross-correlation operations for registering the original image and the target image. Thus, the solution provided in this application, which uses a deep learning model for image registration, not only improves registration efficiency but also eliminates the influence of manual intervention on registration accuracy, thereby improving registration precision.

[0060] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0061] Exemplary methods

[0062] Figure 1 is a flowchart illustrating an image registration method based on deep learning provided in an embodiment of this application. Figure 1a As shown, the method includes the following:

[0063] Step S110: Obtain the original image and the target image;

[0064] The main objective of this application is to investigate the influence of angle on the registration of infrared and visible light images, including the accuracy of registration fusion, and to propose requirements and strategies for infrared and visible light data acquisition to achieve automatic registration of infrared and visible light images, providing effective data support for subsequent data fusion analysis. Based on this, in the scheme provided in this application, the original image can be a visible light image; the target image is an infrared image.

[0065] Step S120: Using a preset deep learning model, register the original image and the target image to obtain a target registered image;

[0066] The deep learning model is used to register the original image and the target image using an algorithm based on bidirectional cross-correlation.

[0067] Therefore, the solution provided in this application, which uses a deep learning model for image registration, not only improves the efficiency of registration but also eliminates the influence of manual intervention on registration accuracy, thereby improving the precision of registration.

[0068] Furthermore, the step of registering the original image and the target image using a preset deep learning model to obtain a target registered image includes:

[0069] Based on a first preset scaling factor, the original image and the target image are downsampled to obtain two single-sampled images;

[0070] The sampled image is input into a preset deep learning model to obtain a registration result, and a registered image is obtained based on the registration result, the original image, and the target image.

[0071] Based on the second preset scale factor, the primary registration image and the target image are downsampled to obtain two secondary sampled images;

[0072] The secondary sampled image is input into a preset deep learning model to obtain a secondary registration result, and a secondary registered image is obtained based on the secondary registration result, the original image, and the target image;

[0073] Based on three preset scale factors, the secondary registration image and the target image are downsampled to obtain two cubically sampled images;

[0074] The subsampled image is input into a preset deep learning model to obtain a three-stage registration result. Based on the three-stage registration result, the original image, and the target image, a target registered image is obtained.

[0075] Specifically, this application employs a coarse-to-fine strategy. (See reference...) Figure 1b The three network-in-network (NIN) models are concatenated and stacked. Image pairs at multiple scales (such as different downsampling rates) are input into the NIN to directly generate registered images, such as... Figure 2 As shown. First, the scale 1 NIN model aims to perform initial and global alignment of the original image with the target image. Specifically, the original image is downsampled to a scale factor of 1 / 4 and input into the first NIN model for one registration to produce the first corrected perceptual image.

[0076] Secondly, the NIN model in scale 2 aims to perform residual and detailed alignment between the target image and the first-registered image. Specifically, the target image and the first-registered image are downsampled by a scale factor of 1 / 2 and input into the second NIN model for a second registration, resulting in the second-registered image.

[0077] Finally, the scale 3 NIN model performs residual and detailed alignment between the target image and the secondary registration image to generate the final registration image.

[0078] Furthermore, the deep learning model is used to: extract features from two input images to obtain features; and perform bidirectional registration based on the features to obtain a registration result.

[0079] Specifically, the deep learning model is mainly used to perform the following processes: extracting features from two input images to obtain features, including: extracting features from the two input images respectively to obtain a first feature and a second feature; determining a first mapping relationship from the first feature to the second feature based on a preset neural network; determining a second mapping relationship from the second feature to the first feature based on a preset neural network; determining a first transformation parameter corresponding to the first mapping relationship and a second transformation parameter corresponding to the second mapping relationship based on a preset parametric regression network; weighting and summing the first transformation parameter and the second transformation parameter to obtain the final parameter; and determining the registration result based on the final parameter.

[0080] Furthermore, feature extraction is performed on the two input images to obtain the features, including:

[0081] Based on a channel attention mechanism, features are extracted from two images to obtain third and fourth features. Bidirectional registration is then performed based on these features to obtain the registration result, including: determining a third mapping relationship from the third feature to the fourth feature using a preset neural network; determining a fourth mapping relationship from the fourth feature to the third feature using a preset neural network; determining a third transformation parameter corresponding to the third mapping relationship and a fourth transformation parameter corresponding to the fourth mapping relationship using a preset parametric regression network; weighting and summing the third and fourth transformation parameters to obtain the final parameters; and determining the registration result based on the final parameters.

[0082] Specifically, refer to Figure 2 The proposed solution utilizes a network within a network that integrates an attention mechanism as the backbone to extract features from both the reference and target images. A feature extraction structure combining pre-training and channel attention is constructed. By extracting source and target features from the third layer of the pre-trained network, along with the source and target features and channel attention mechanism after pre-training, a foundation is laid for subsequent operations. Multi-scale matching relationships are constructed using the extracted feature maps: bidirectional matching structures at different scales are built. First, Pearson cross-correlation is performed on the source and target features; target and source features are extracted from the third layer of the pre-trained network to obtain the first set of bidirectional matching relationships. Then, Pearson cross-correlation is performed on the source and target features, and the target and source features are extracted through the attention mechanism to obtain the second set of bidirectional matching relationships. The four matching relationships obtained from feature matching are input into convolutional layers with kernels of 5x5 and 7x7, respectively, and finally, the four required parameters are output through a fully connected layer. A loss function is designed using the obtained four parameters. By reducing network loss, the network model is iteratively optimized to output the optimal model and the most accurate transformation parameters. By combining the bidirectional parameters of the integrated channel focus, an affine transformation is performed on the source image to obtain the final registration result.

[0083] This application proposes an improved network to address the challenges of complex backgrounds, difficult feature extraction, and poor accuracy in multi-angle, high-resolution inspection images. The network incorporates an attention mechanism during feature extraction to suppress redundant information and enhance global capture capabilities. In feature matching, it constructs bidirectional matching structures at different scales to improve registration accuracy. Furthermore, a loss function is designed to reduce network loss, thus resolving issues of poor real-time performance and low instance matching accuracy. The algorithm exhibits relatively stable scene matching performance, and experiments demonstrate that this end-to-end network can be effectively applied to multi-angle inspection image registration.

[0084] In some embodiments, the deep learning module provided in this application mainly performs the following process:

[0085] 1. Feature Extraction

[0086] To address the challenge of multi-angle inspection image registration, network-in-network (NIN) employs micro-neural networks to extract not only linear features from inspection images of complex terrain but also high-dimensional nonlinear features, thereby enhancing the model's feature extraction capabilities.

[0087] Furthermore, the mean pooling used in the NIN structure effectively avoids overfitting caused by too many parameters, and can effectively handle the problem of inspected images. After extracting as many features as possible, in order to reduce the computational cost of the feature matching stage, a channel attention mechanism is added after the pre-trained feature extraction network to filter out the required useful feature information. Figure 3 A structural diagram of the channel attention mechanism is shown.

[0088] The channel attention mechanism first passes through parallel average pooling and max pooling layers, then each layer is fed into a three-layer perceptron. Finally, the results from the two layers are summed and sent to the ReLU activation function to obtain the channel attention mechanism and feature mapping. The process is shown in the following formula:

[0089] M C (F)=δ(MLP(AvgPool(F))+MLP(MaxPool(F))) (1)

[0090] In the formula M C(F) represents the feature map obtained through channel attention, σ is a parameter, multilayer perceptron (MLP) is a three-layer perceptron, AvgPool is the average pooling layer, MaxPool is the max pooling layer, and F is the input feature map. Considering that different layers of features have different representational capabilities for images—lower-level features have weak semantic information representation capabilities but strong geometric detail representation capabilities; higher-level features have strong semantic information representation capabilities but weak geometric detail representation capabilities—to compensate for the deficiencies of each layer of features, the third-layer features from the pre-trained network and the fourth-layer features after the channel attention mechanism are extracted, paving the way for subsequent operations.

[0091] 2. Feature matching

[0092] To address the issue that single-scale matching in the feature matching stage can lose many correct matching points due to the low resolution of inspection images, this paper employs multi-scale bidirectional cross-correlation computation in the feature matching stage. The bidirectional computation can calculate the multi-scale similarity between feature maps and utilize multi-scale features for registration, thus improving the efficiency of the matching stage.

[0093] A bidirectional cross-correlation relationship is established using feature information S, S', T, and T' at different scales obtained through a feature extraction network. Let i and j be coordinate points on the image coordinate system, k be the index on the channel slice, and C... ST (i,j,k) and C TS (i,j,k) represent bidirectional data obtained from the third layer of the pre-trained network. Matching relationships (mapping from source features to target features, and mapping from target features to source features); C S'T' (i,j,k) and C T'S' (i,j,k) represents the feature mapping relationship after the attention mechanism (the mapping from source features to target features, and the mapping from target features to source features). Then C ST (i,j,k), C TS (i,j,k), C S'T' (i,j,k) and C T'S' (i,j,k) is:

[0094] C ST (i,j,k)=f T (i,j) k f s (i k ,j k (2)

[0095] C TS (i,j,k)=f S (i,j) k f T (i k ,j k(3)

[0096] C S'T' (i,j,k)=f S' (i,j) k f T' (i k ,j k (4)

[0097] C T'S' (i,j,k)=f T' (i,j) k f S' (i k ,j k (5)

[0098] 3. Parametric Regression

[0099] To obtain the final transformation parameters, the mapping relationships between features obtained from feature matching are sent to a parameter regression network. After two 5x5 convolutional layers and a fully connected layer, the transformation parameters θ of the four relationships are obtained through regression. S→T θ T→S θ S'→T' θ T'→S'

[0100] 4. Design the loss function

[0101] The algorithm uses mesh loss as the baseline loss function, which measures the distance deviation between the coordinates after affine transformation and the manually labeled coordinates to guide network optimization. This can be expressed as:

[0102]

[0103] Where, θ gt The parameters are the actual values, and θ is the transformed output parameter. N is the number of coordinate points, (x i ,y i Let be the coordinate point with x-coordinate i and y-coordinate j. This invention obtains a total of four transformation parameters. The loss function is optimized using these transformation parameters at different scales to obtain the losses at different scales, as shown below:

[0104]

[0105]

[0106]

[0107] Where L_1 is the third-layer feature loss of the pre-trained network, L_2 is the fourth-layer feature loss after the channel attention mechanism, and the overall loss function is l. After training the optimal network, the transformation parameters obtained after the channel attention mechanism are weighted and combined to obtain the final transformation parameters. Then, an affine transformation is performed on the source image to obtain the registration result. The weighted combination result is as follows:

[0108]

[0109] Where θ S'→T' θ is the transformation parameter from source features to target features after channel focus. T'→S' To map the target features to source features after channel focus, the transformation parameters are averaged and synthesized to obtain the final parameter θ. en .

[0110] Exemplary device

[0111] The apparatus embodiments of this application can be used to execute the method embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments of this application.

[0112] Figure 4 The diagram shown is a block diagram of an image registration device based on deep learning according to an embodiment of this application. Figure 4 As shown, the device includes:

[0113] Acquisition module 41 is used to acquire the original image and the target image;

[0114] The registration module 42 is used to register the original image and the target image using a preset deep learning model to obtain a target registered image;

[0115] The deep learning model is used to register the original image and the target image using an algorithm based on bidirectional cross-correlation.

[0116] Exemplary electronic devices

[0117] Below, for reference Figure 5 This describes an electronic device according to embodiments of the present application. Figure 5 A block diagram of an electronic device according to an embodiment of this application is illustrated.

[0118] like Figure 5 As shown, the electronic device 500 includes one or more processors 510 and memory 520.

[0119] The processor 510 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 500 to perform desired functions.

[0120] The memory 520 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 510 may execute the program instructions to implement the deep learning-based image registration methods of the various embodiments of this application described above, and / or other desired functions. Various contents, such as category correspondences, may also be stored in the computer-readable storage medium.

[0121] In one example, the electronic device 500 may also include an input device 530 and an output device 540, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0122] In addition, the input device 530 may also include, for example, a keyboard, mouse, interface, etc. The output device 540 can output various information to the outside, including analysis results, etc. The output device 540 may include, for example, a monitor, speaker, printer, and communication network and its connected remote output devices, etc.

[0123] Of course, for the sake of simplicity, Figure 5 Only some of the components of the electronic device relevant to this application are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device may include any other suitable components depending on the specific application.

[0124] Exemplary computer program products and computer-readable storage media

[0125] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products comprising computer program instructions that, when executed by a processor, cause the processor to perform the steps in the deep learning-based image registration method described in the "Exemplary Methods" section of this specification according to various embodiments of this application.

[0126] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0127] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform the steps in the deep learning-based image registration methods according to various embodiments of this application described in the "Exemplary Methods" section of this specification.

[0128] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0129] The above description has been given for illustrative and descriptive purposes. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A deep learning-based image registration method, characterized in that, include: Acquire the original image and the target image; Using a pre-defined deep learning model, the original image and the target image are registered to obtain a target registered image; The deep learning model is used to register the original image and the target image using an algorithm based on bidirectional cross-correlation.

2. The image registration method based on deep learning according to claim 1, characterized in that, The step of registering the original image and the target image using a preset deep learning model to obtain a target registered image includes: Based on a first preset scaling factor, the original image and the target image are downsampled to obtain two single-sampled images; The sampled image is input into a preset deep learning model to obtain a registration result, and a registered image is obtained based on the registration result, the original image, and the target image. Based on the second preset scale factor, the primary registration image and the target image are downsampled to obtain two secondary sampled images; The secondary sampled image is input into a preset deep learning model to obtain a secondary registration result, and a secondary registered image is obtained based on the secondary registration result, the original image, and the target image; Based on three preset scale factors, the secondary registration image and the target image are downsampled to obtain two cubically sampled images; The subsampled image is input into a preset deep learning model to obtain a three-stage registration result. Based on the three-stage registration result, the original image, and the target image, a target registered image is obtained.

3. The image registration method based on deep learning according to claim 2, characterized in that, The deep learning model is used for: Feature extraction is performed on the two input images to obtain the features; Based on the aforementioned features, bidirectional registration is performed to obtain the registration result.

4. The image registration method based on deep learning according to claim 3, characterized in that, Feature extraction is performed on the two input images to obtain the features, including: Feature extraction is performed on the two input images to obtain the first feature and the second feature.

5. The image registration method based on deep learning according to claim 4, characterized in that, Based on the aforementioned features, bidirectional registration is performed to obtain the registration result, including: The first mapping relationship from the first feature to the second feature is determined based on a pre-defined neural network. The second mapping relationship from the second feature to the first feature is determined based on a pre-defined neural network. Based on a preset parameter regression network, the first transformation parameter corresponding to the first mapping relationship is determined, and the second transformation parameter corresponding to the second mapping relationship is determined. The final parameters are obtained by weighted summation of the first and second transformation parameters. Based on the final parameters, the registration result is determined.

6. The image registration method based on deep learning according to claim 3, characterized in that, Feature extraction is performed on the two input images to obtain the features, including: Based on the channel attention mechanism, features are extracted from the two images, namely the third and fourth features.

7. The image registration method based on deep learning according to claim 6, characterized in that, Based on the aforementioned features, bidirectional registration is performed to obtain the registration result, including: The third mapping relationship between the third feature and the fourth feature is determined based on a pre-defined neural network. The fourth mapping relationship between the fourth feature and the third feature is determined based on a pre-defined neural network. Based on a preset parameter regression network, the third transformation parameter corresponding to the third mapping relationship is determined, and the fourth transformation parameter corresponding to the fourth mapping relationship is determined. The final parameters are obtained by weighted summation of the third and fourth transformation parameters. Based on the final parameters, the registration result is determined.

8. An image registration device based on deep learning, characterized in that, include: The acquisition module is used to acquire the original image and the target image; The registration module is used to register the original image and the target image using a preset deep learning model to obtain a target registered image; The deep learning model is used to register the original image and the target image using an algorithm based on bidirectional cross-correlation.

9. An electronic device, characterized in that, include: A processor, and a memory for storing a processor-executable program; The processor is configured to implement the deep learning-based image registration method as described in any one of claims 1 to 7 by running a program in the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the deep learning-based image registration method as described in any one of claims 1 to 7.

Images