Method and system for constructing homography estimation model based on cross-mamba dynamic feature fusion optimization

Abstract

The application discloses a method and system for constructing a homography estimation model based on cross-mamba dynamic feature fusion optimization, comprising the following steps: building the overall framework of the network, the network is entirely composed of neural network layers, including an input module feature pyramid, a coarse matching module, and a fine matching module; setting the structure of the cross-mamba model; setting the number, connection mode, and type of the cross-mamba model in the coarse matching module; training the network to obtain a homography estimation network; the mamba network is used for multi-modal image fusion, better exploring intra-modal and inter-modal features, and dynamically enhancing the detail texture information of the source image and the specific information of each modal; for the feature extraction and reconstruction part of the network, a mamba model suitable for the fusion task is designed by combining a visual state space model with dynamic convolution and channel attention, which not only maintains the performance and global modeling capability of the mamba, but also reduces channel redundancy and enhances the local feature extraction capability.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence image registration, and particularly relates to a CrossMamba dynamic feature fusion optimization homography estimation model construction method and system. BACKGROUND

[0002] With the development of computer vision and the sharp increase in demand for robot positioning slam, homography estimation on image pairs has also become a hot topic, especially for homography estimation in complex scenes, which has broad application prospects. The traditional method of homography estimation usually solves the problem through the following series of processes: feature extraction, corresponding feature matching for homography estimation. At present, the more popular methods mainly fall into two categories: matching-based methods and unsupervised deep homography methods. However, the former lacks discriminative features when dealing with textureless or blurred images, and they can be very sensitive to errors or ambiguities in the data, which makes them difficult to handle image pairs with large baselines; the latter has high learning cost, and the attention tends to spread over a large area rather than focusing on the corresponding area, so it may introduce noise in the feature fusion update process, leading to incorrect matching.

[0003] Moreover, these methods cannot fully capture the cross-modal correlation between the two modalities, and have high computational complexity. SUMMARY

[0004] In order to solve the problems existing in the prior art, the application provides a CrossMamba dynamic feature fusion optimization homography estimation model construction method, which utilizes the feature pyramid (FPN) extraction image feature capability, the Mamba global information modeling capability and the CrossMamba dynamic feature fusion capability to realize end-to-end homography estimation.

[0005] In order to achieve the above purpose, in a first aspect, the application provides a CrossMamba dynamic feature fusion optimization homography estimation model construction method, comprising the following steps:

[0006] The network model comprises an input module, a coarse matching module and a fine matching module; the input module is used to extract input features of different scales of images, the coarse matching module is used to generate coarse matching, and the fine matching module is used to generate the final fine matching array;

[0007] The feature size of the input module is set, the structure of the coarse matching module is set, and the CrossMamba model is used inside the coarse matching module; the algorithm of the fine matching module is set, and the random sample consensus geometric method is used;

[0008] The internal weight connection parameters of the network model are optimized to obtain a homography estimation network model.

[0009] Further, in the overall framework of the network model:

[0010] The input module is a feature pyramid network. For a given pair of input images, the coarse-level feature map Cl with an eighth of the original image dimension and the fine-level feature map Fl with half of the original image dimension are extracted. The coarse-level feature map Cl and the fine-level feature map Fl are input into the coarse matching module and the fine matching module, respectively.

[0011] The coarse matching module inputs two coarse features C0 and C1 into the CrossMamba module group for intra-modal and inter-modal learning, generating a pixel-to-pixel confidence matrix Pc. By thresholding Pc and performing mutual nearest neighbor search, the coarse matching array Mc is determined.

[0012] The fine matching module uses the classic random sample consensus geometric method. Given the coarse matching array Mc and the previous fine features F0 and F1, the coarse matching positions are located on the fine feature map, the matching probability is calculated by cropping and converting, and finally the sub-pixel matching array Mf is generated.

[0013] Further, the coarse matching module adopts a dynamic feature fusion model based on the Mamba model, CrossMamba, as follows:

[0014] The same two features x1 are input into the self CrossMamba model for intra-modal learning, and the same two features x2 are input into the self CrossMamba model for intra-modal learning. After obtaining new features x1 and x2, they are combined as (x1, x2) and (x2, x1), and then input into the cross CrossMamba for inter-modal learning.

[0015] The above process is repeated 4 times, i.e., 4 [self, cross] modules, to obtain the detailed texture information of the dynamic enhanced source image and the information of each modality.

[0016] Further, CrossMamba uses a cross-scan scanning mechanism and adopts a four-way scanning strategy to integrate information from all other positions in different directions for each element in the feature, forming a global receptive field and incorporating a local attention component, HomoAttention, to establish connections between input features from different modalities.

[0017] Further, the homography attention (HomoAttention) fuses feature information in a two-stage method, learns image features through convolution and activation function, and finally generates feature weight values. Specifically, in the first stage, the image features (x1, x2) of different modalities are respectively learned by the visual state space module to learn the internal information of each image. The output (y1, y2) obtained is calculated by the homography attention (HomoAttention) module to obtain the local attention weight U. According to the importance of the features, the initial (y1, y2) features are dynamically weighted and combined.

[0018] In the second stage, the mixed features are further input into the visual state space module to enhance the learned fusion feature information, and the original feature information (x1, x2) is dynamically weighted and combined.

[0019] Further, traversing the image along the four scanning paths includes the following steps:

[0020] Flatten a 2D feature into a one-dimensional vector in four different directions;

[0021] S6 block independently sends the four one-dimensional vectors obtained in the previous step to the S6 model for processing;

[0022] Fuse the one-dimensional vectors obtained by the S6 model processing into a 2D feature output.

[0023] In a second aspect, based on the above method, the application provides a CrossMamba dynamic feature fusion optimization homography estimation model construction system, which comprises a network structure construction module, a network parameter preset module and a training module.

[0024] The network structure construction module is used to construct the overall framework of the network model, wherein the network model is composed of a neural network layer, including an input module, a coarse matching module and a fine matching module. The input module is used to extract input features of different scales of images, the coarse matching module is used to generate coarse matching, and the fine matching module is used to generate the final fine matching array.

[0025] The network parameter preset module is used to set the extraction feature size of the input module feature pyramid (FPN); set the structure of the coarse matching module, and the CrossMamba model with special connection is used in the coarse matching module; set the algorithm of the fine matching module, and the random sample consensus geometric method is adopted;

[0026] The training module is used to optimize the internal weight connection parameters of the network model to obtain a homography estimation network model.

[0027] In a third aspect, based on the model obtained by the method, the application provides a homography estimation method based on CrossMamba dynamic feature fusion optimization, comprising the following steps: obtaining a pair of (gray) input images, passing through an end-to-end homography estimation network model, and outputting the homography relationship between the pair of images, that is, a homography matrix; the end-to-end homography estimation network model adopts the CrossMamba dynamic feature fusion optimization homography estimation model construction method.

[0028] In a fourth aspect, the application further provides a computer device comprising a processor and a memory, the memory being used to store a computer executable program, the processor reading part or all of the computer executable program from the memory and executing, and the processor implementing the CrossMamba dynamic feature fusion optimization homography estimation model construction method or the CrossMamba dynamic feature fusion optimization homography estimation method when executing part or all of the computer executable program.

[0029] Meanwhile, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program can implement the CrossMamba dynamic feature fusion optimization homography estimation model construction method or the CrossMamba dynamic feature fusion optimization homography estimation method when executed by a processor.

[0030] Compared with the prior art, the application has at least the following beneficial effects: the method focuses on feature fusion to enhance the model. The CrossMamba model is introduced, which is a homography estimation network using a mamba as a core module; the CrossMamba utilizes the inherent cross-modal relationship between the two to realize fusion, filter redundant modal features and obtain fine fusion results, extract global information from two views and perform efficient feature fusion, and due to the efficient feature extraction and fusion capability, the model has surpassed the current advanced method and achieved excellent fusion results. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a schematic diagram of the overall architecture of the network in the application.

[0032] Figure 2 It is a structural schematic diagram of the coarse matching module in the application.

[0033] Figure 3 It is a structural schematic diagram of the CrossMamba module in the application.

[0034] Figure 4A schematic diagram of a visual manba module scanning method used in the present application. DETAILED DESCRIPTION

[0035] The problem to be solved by the present application is to estimate homography of images with different perspectives. A pair of input images is obtained, and an end-to-end homography estimation network model is used to output the homography relationship between the two images, that is, the homography matrix.

[0036] For example, during the movement of the camera, two pictures taken at different positions are input into the network model after encoding, and the homography relationship between the two images is output.

[0037] The input end of the network model can be various image collection devices such as cameras and cameras, and the output end can be specific monitoring indication devices, task execution devices or more complex calculation models for further analysis and processing.

[0038] The network model is entirely composed of neural network layers, and in the order of information transmission, there are three modules: an input module feature pyramid (FPN), a coarse matching module, and a fine matching module. The original pair of images with different perspectives is input into the network model, and the homography relationship between the two images is obtained at the output end. The overall architecture of the network model is shown in Figure 1 .

[0039] The network model and each module are as follows:

[0040] The input module needs to be able to effectively extract features of different sizes of images, and the application adopts a common feature pyramid (FPN) feature pyramid network. For a given pair of input images, coarse-level feature maps Cl of one-eighth of the original image dimension and fine-level feature maps Fl of one-half of the original image dimension are extracted and input into the coarse matching module and the fine matching module respectively. The feature pyramid (FPN) includes three parts, a bottom-up line, a top-down line and a horizontal connection. The horizontal connection 1*1 convolution kernel is mainly used to adjust the output channel of different feature maps to be 256, because it needs to be fused, and only by making them have the same shape can the fusion be realized, so that the other layer features after upsampling can be added without changing the size of the feature map. Bottom-up: it is actually the forward process of the network. In the forward process, the selection is performed by 2 integer times. The size of the feature map will change after passing through certain layers, and will not change when passing through other layers. Therefore, the extracted features are the output of the last layer of each stage, so as to form a feature pyramid. The network backbone of the feature selection network is scaled by 2 integer times. Top-down: the process adopts upsampling, and 2 times upsampling is realized by a simple adjacent interpolation algorithm. The horizontal connection is to fuse the results of upsampling and the feature maps of the same size generated from bottom to top. After fusion, a 3*3 convolution kernel is used to convolve each fusion result, the purpose of which is to eliminate the aliasing effect of upsampling. And it is assumed that the generated feature map results are P2, P3, P4, P5, and the original convolution results C2, C3, C4, C5 correspond one by one.

[0041] The coarse matching module (coarse matching) is shown in Figure 2 , which is used to generate coarse matching and is composed of a plurality of CrossMamba modules. The two coarse features output by the feature pyramid network feature pyramid (FPN) are input into the innovative CrossMamba module group to learn within and between modalities, so as to generate a pixel-to-pixel confidence matrix Pc. By thresholding the confidence matrix Pc and then performing mutual nearest neighbor search, a coarse matching array Mc is determined.

[0042] Since the subsequent modules use the random sampling consistency geometric method to greatly reduce the image area involved in attention calculation, they are more dependent on the output of the upstream operation coarse matching module, and the accuracy of the output seriously affects the results of the subsequent random sampling consistency algorithm. Therefore, the focus of the improvement of the present application is on coarse matching, first the same two x1 input CrossMamba model, intra-modal learning, x2 same as above, get new x1, x2, after combination (x1, x2), (x2, x1), input CrossMamba again, inter-modal learning. The above process is repeated 4 times, i.e. 4 [self, cross] modules.

[0043] The CrossMamba module is improved based on the visual Mamba model as shown in Figure 3 The key to the success of V-Mamba is the adoption of the selective scanning spatial state sequence model (S6 model), which is designed to solve natural language processing (NLP) tasks at the beginning. Unlike the attention mechanism in the ViT model, the S6 model interacts each element in the one-dimensional vector (such as a text sequence) with the information scanned before, effectively reducing the quadratic complexity to linear.

[0044] However, since visual signals (such as images) do not have the natural orderliness that text sequences have, it is not possible to simply apply the data scanning method in the S6 model directly to visual signals. For this purpose, a cross scanning mechanism is designed, and the cross scanning module (CSM) adopts a four-way scanning strategy, i.e. scanning from the four corners of the feature map at the same time, as shown in Figure 4 This strategy ensures that each element in the feature integrates information from all other positions in different directions, forming a global receptive field without increasing the linear computational complexity. On the basis of CSM, the visual Mamba network proposes a 2D selective scanning module SS2D module, which constructs a linear time complexity context information acquisition method by traversing the image along four scanning paths. The processing process of this module contains three steps:

[0045] 1) Flatten a 2D feature into a one-dimensional vector along 4 different directions (top left, bottom right, bottom left, top right).

[0046] 2) S6 block independently sends the 4 one-dimensional vectors obtained in the previous step into the S6 model operation.

[0047] 3) The four one-dimensional vectors obtained in the first step are fused into a 2D feature output.

[0048] In addition to the visual Mamba module, the present application innovatively introduces a homography attention (HomoAttention) module. The input of the attention module, the initial integrated quality, can greatly affect the final fusion weight. Since it is still a feature fusion problem, an intuitive method is to use another attention module to fuse the input features. The homography attention (HomoAttention) passes through a series of convolutional layers and activation functions, and finally assigns the obtained weights to different modalities, establishes the connection between the input features from different modalities, and adopts a two-stage method to fuse the feature information. This module can effectively mine the correlation information between modalities and promote better information interaction between modalities.

[0049] The following is the calculation process of the CrossMamba model:

[0050] The first stage first learns the internal information of each image through the visual state space module for the image features (x1, x2) of different modalities.

[0051]

[0052]

[0053] The obtained output (y1, y2) is input into the homography attention (HomoAttention) module to calculate the local attention weight U. W1 and W2 are point-by-point convolution.

[0054]

[0055] According to the importance of the features, the initial (y1, y2) features are dynamically weighted and combined, so as to improve the ability of the model to extract relevant information from the input signal.

[0056]

[0057] In the second stage, the mixed features are further input into the visual state space module to enhance the learned fusion feature information.

[0058]

[0059] The original feature information (x1, x2) is dynamically weighted and combined.

[0060]

[0061] The network building and training steps of the present application are as follows:

[0062] 1. Build the overall framework based on CrossMamba dynamic feature fusion network. Based on CrossMamba dynamic feature fusion network, the overall architecture includes an input module feature pyramid (FPN), a coarse matching module and a fine matching module, forming an end-to-end homography estimation network model. The eighth feature map extracted from the input module feature pyramid (FPN) is directly input into the coarse matching network; the CrossMamba module of the coarse matching network combines the visual state space model with dynamic convolution and channel attention for multi-modal image fusion; the half feature map extracted from the input end feature pyramid (FPN) is input into the fine matching network together with the coarse matching matrix output by the coarse matching network, and finally the fine matching network outputs the final homography matrix.

[0063] 2. Train the network. The network is implemented using PyTorch. According to our GPU resources (4 NVIDIA GeForce RTX 3090 cards), the larger size of the training image is set to 640 pixels. The optimizer uses Adam, β = (0.9, 0.999), and the initial learning rate is 0.001. Each small batch contains a pair of images. The maximum number of training cycles is 20. In the inference stage, given the fine matching Mf, cv2.findHomography and random sample consensus are used as robust estimators for homography fitting.

[0064] The training results are as follows, compared with the Transformer local feature matching model (LoFTR), the geometricized Transformer model (GeoFormer), and the geometricized Transformer model with quadtree (GeoFormer with quadtree), the CrossMamba model has much higher accuracy than other models at each scale, and we also verify the effectiveness of the HomoAtttion module:

[0065]

[0066] In embodiment 2, the present application provides a CrossMamba dynamic feature fusion optimized homography estimation method, which includes the following steps:

[0067] A pair of input images are obtained, which are input into an end-to-end homography estimation network model to output the homography relationship between the two images, that is, the homography matrix; the input image is a grayscale image;

[0068] In the end-to-end homography estimation network model: based on the feature pyramid (FPN) feature pyramid network extracts image features, for example, coarse level feature maps Cl of one-eighth size of the original image dimension and fine level feature maps Fl of one-half size of the original image dimension can be extracted;

[0069] Intra-modal and inter-modal learning is performed on a group of coarse level feature maps to generate a pixel-to-pixel confidence matrix Pc, and by thresholding the confidence matrix Pc and performing mutual nearest neighbor search, a coarse matching array Mc is determined.

[0070] Using the classic random sample consensus geometric method, the coarse matching position is located on the fine feature map using the given coarse matching array Mc and the previous fine features F0 and F1, the matching probability is calculated by cropping and converting, and finally the sub-pixel matching array Mf is generated.

[0071] Intra-modal learning is performed on the same two features x1 and x2, and after obtaining new features x1 and x2, cross combination is performed (x1, x2), (x2, x1); inter-modal learning is performed; the above process is repeated 4 times to obtain the detailed texture information of the dynamic enhanced source image and the information of each modality; using a cross-scan scanning mechanism and a four-way scanning strategy, each element in the feature integrates information from all other positions in different directions to form a global receptive field, and based on a homography attention (HomoAttention) local attention component, feature weight values are learned by convolution and activation function to establish connections between input features from different modalities.

[0072] The application can also provide a computer device comprising a processor and a memory, the memory being used to store computer executable programs, the processor reading part or all of the computer executable programs from the memory and executing, and the processor executing part or all of the computer executable programs can realize the CrossMamba dynamic feature fusion optimized homography estimation model construction method.

[0073] On the other hand, the application provides a computer readable storage medium, the computer readable storage medium storing a computer program, the computer program being executed by a processor, and the CrossMamba dynamic feature fusion optimized homography estimation model construction method and the CrossMamba dynamic feature fusion optimized homography estimation method can be realized.

[0074] The computer device can be a notebook computer, a desktop computer, a vehicle-mounted computer or a workstation.

[0075] For the memory described in the present application, it can be an internal storage unit of a notebook computer, a desktop computer, a vehicle-mounted computer or a workstation, such as a memory or a hard disk; or an external storage unit, such as a mobile hard disk or a flash card.

[0076] The computer-readable storage medium can include computer storage medium and communication medium. The computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. The computer-readable storage medium can include read-only memory (ROM), random access memory (RAM), solid state disk (SSD) or optical disk, etc. Among them, the random access memory can include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

Claims

1. A method for constructing a homography estimation model based on cross-mamba dynamic feature fusion optimization, characterized in that: The following steps are involved: The network model is constructed by input module, coarse matching module and fine matching module; the input module is used to extract input features of different scales of the image, the coarse matching module is used to generate coarse-level matching, and the fine matching module is used to generate the final fine matching array; Set the extracted feature size of the input module, set the coarse matching module structure, and use the cross-mamba model inside the coarse matching module; set the fine matching module algorithm and use the random sampling consistency geometry method; Optimize the internal weight connection parameters of the network model to obtain the homography estimation network model; In the overall framework of the network model: The input module is a feature pyramid network. For a given pair of input images, it extracts a coarse feature map Cl with a size of one-eighth the original image dimension and a fine feature map Fl with a size of one-half the original image dimension. The coarse feature map Cl and the fine feature map Fl are input into the coarse matching module and the fine matching module respectively. The coarse matching module inputs the two coarse features C0 and C1 into the Cross Mamba module group for intra-modal and inter-modal learning, generates a pixel-to-pixel confidence matrix Pc, and determines the coarse matching array Mc by thresholding Pc and performing mutual nearest neighbor search. The fine matching module uses the classic random sampling consistency geometry method to locate the coarse matching position on the fine feature map given the coarse matching array Mc and the previous fine features F0 and F1, perform cropping transformation to calculate the matching probability, and finally generate the sub-pixel matching array Mf; The same two features x1 are fed into the self-crossed Mamba model for intra-modal learning. The same two features x2 are fed into the self-crossed Mamba model for intra-modal learning. After obtaining new features x1 and x2, they are cross-combined (x1, x2) and (x2, x1). These are then fed into the cross-crossed Mamba model for inter-modal learning. The above process is repeated multiple times, i.e., multiple [self, cross] modules, to obtain the detailed texture information of the dynamically enhanced source image and the information of each modality. Cross Mamba uses a cross-scanning mechanism and a four-way scanning strategy to enable each element in the feature to integrate information from all other positions in different directions to form a global receptive field, incorporate homography attention components, and establish connections between input features from different modalities; Homography attention uses a two-stage approach to fuse feature information, learning image features through convolution and activation functions and finally generating feature weight values. Specifically, in the first stage, the image features (x1, x2) of different modalities are first learned through the visual state space module to learn the internal information of each image. The output (y1, y2) is passed through the homography attention module to calculate the local attention weight U, and the initial (y1, y2) features are dynamically weighted and combined according to the importance of the features. In the second stage, the mixed features are input into the visual state space module to enhance the learned fusion feature information and dynamically weight and combine the original feature information (x1, x2).

2. The method for constructing a homography estimation model based on cross-Mamba dynamic feature fusion optimization according to claim 1 is characterized in that: Traversing the image along the four scan paths involves the following steps: Flatten a 2D feature into a one-dimensional vector along 4 different directions; The S6 block independently feeds the four one-dimensional vectors obtained in the previous step into the S6 model for processing; The one-dimensional vectors obtained by S6 model processing are fused into a 2D feature output.

3. A system for constructing a homography estimation model based on Cross Mamba dynamic feature fusion optimization, characterized in that: Including network structure construction module, network parameter preset module and training module; The network structure building module is used to build the overall framework of the network model, where the network model is composed of neural network layers, including input module, coarse matching module and fine matching module; the input module is used to extract input features of different scales of the image, the coarse matching module is used to generate coarse-level matching, and the fine matching module is used to generate the final fine matching array; The network parameter preset module is used to set the feature size of the input module feature pyramid FPN; set the coarse matching module structure, which uses a specially connected cross-mamba model; and set the fine matching module algorithm, which uses a random sampling consistency geometry method. The training module is used to optimize the internal weight connection parameters of the network model to obtain the homography estimation network model; In the overall framework of the network model: The input module is a feature pyramid network. For a given pair of input images, it extracts a coarse feature map Cl with a size of one-eighth the original image dimension and a fine feature map Fl with a size of one-half the original image dimension. The coarse feature map Cl and the fine feature map Fl are input into the coarse matching module and the fine matching module respectively. The coarse matching module inputs the two coarse features C0 and C1 into the Cross Mamba module group for intra-modal and inter-modal learning, generates a pixel-to-pixel confidence matrix Pc, and determines the coarse matching array Mc by thresholding Pc and performing mutual nearest neighbor search. The fine matching module uses the classic random sampling consistency geometry method to locate the coarse matching position on the fine feature map given the coarse matching array Mc and the previous fine features F0 and F1, perform cropping transformation to calculate the matching probability, and finally generate the sub-pixel matching array Mf; The same two features x1 are fed into the self-crossed Mamba model for intra-modal learning. The same two features x2 are fed into the self-crossed Mamba model for intra-modal learning. After obtaining new features x1 and x2, they are cross-combined (x1, x2) and (x2, x1). These are then fed into the cross-crossed Mamba model for inter-modal learning. The above process is repeated multiple times, i.e., multiple [self, cross] modules, to obtain the detailed texture information of the dynamically enhanced source image and the information of each modality. Cross Mamba uses a cross-scanning mechanism and a four-way scanning strategy to enable each element in the feature to integrate information from all other positions in different directions to form a global receptive field, incorporate homography attention components, and establish connections between input features from different modalities; Homography attention uses a two-stage approach to fuse feature information, learning image features through convolution and activation functions and finally generating feature weight values. Specifically, in the first stage, the image features (x1, x2) of different modalities are first learned through the visual state space module to learn the internal information of each image. The output (y1, y2) is passed through the homography attention module to calculate the local attention weight U, and the initial (y1, y2) features are dynamically weighted and combined according to the importance of the features. In the second stage, the mixed features are input into the visual state space module to enhance the learned fusion feature information and dynamically weight and combine the original feature information (x1, x2).

4. A homography estimation method based on cross-mamba dynamic feature fusion optimization, characterized in that: The method comprises the following steps: obtaining a pair of grayscale input images, passing them through an end-to-end homography estimation network model, and outputting the homography relationship between the pair of images, that is, the homography matrix; the end-to-end homography estimation network model is obtained by adopting the cross-mamba dynamic feature fusion optimization homography estimation model construction method according to claim 1 or 2.

5. A computer device, characterized in that: The invention comprises a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, it can realize the method for constructing a homography estimation model based on cross-mamba dynamic feature fusion optimization as described in claim 1 or 2 or the method for constructing a homography estimation model based on cross-mamba dynamic feature fusion optimization as described in claim 4.

6. A computer-readable storage medium, characterized in that A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the method for constructing a homography estimation model based on cross-mamba dynamic feature fusion optimization as described in claim 1 or 2 or the method for constructing a homography estimation model based on cross-mamba dynamic feature fusion optimization as described in claim 4.

Images