A robust model fitting method based on double purification dynamic graph network

Abstract

A robust model fitting method based on double purification dynamic graph network relates to the technical field of computer vision. For the data set containing noise and outliers, an initial correspondence set is generated by scale-invariant feature transform, and a dynamic graph network is introduced for processing. A double dynamic purification module is proposed, including a regional channel preference block and an overall space consistency block, which realizes step-by-step sampling and outlier elimination from local and global perspectives respectively, and obtains the purified candidate correspondence set and the inlier probability set. The regional channel preference block strengthens local features through a preference attention mechanism, and the overall space consistency block captures long-range dependencies through a consistent attention mechanism. The model parameters are fitted using the purified candidate correspondence set, and the model estimation loss and point classification loss are calculated in combination with the inlier probability set. The dynamic graph network characteristics and the preference consistency strategy are fully utilized, which is significantly better than the prior art in terms of noise resistance and fitting accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and specifically relates to a robust model fitting method based on a double purification dynamic graph network. BACKGROUND

[0002] Robust model fitting is the process of accurately estimating model parameters from data containing noise and outliers, which plays an important role in many applications of computer vision, such as epipolar estimation, relative pose estimation, motion segmentation, 3D reconstruction, stereo disparity estimation and image registration, etc. Robust model fitting usually includes two key steps: the first step is to generate a set of model hypotheses by sampling and purifying multiple data subsets from image pairs or video sequences, where each subset is composed of the minimum number of data points required to construct a model hypothesis; the second step is to select the optimal hypothesis from the generated model hypotheses as the estimated model and verify it. There is a high degree of dependency between the two steps, and the results of data sampling and purification in the first step directly determine the accuracy of model estimation and verification in the second step. Therefore, model fitting mainly faces two challenges in practice. First, the subsets generated by sampling often mix in outliers and noise, especially in scenes involving wide-baseline stereo, complex lighting conditions, occlusion and blur, etc., the selection of inliers becomes extremely complex, significantly increasing the computational burden; second, the data points of model fitting are usually unordered and unevenly distributed, which can lead to ambiguous model hypotheses, thereby reducing the accuracy of fitting.

[0003] Traditional methods rely on hypothesis generation and verification mechanisms to filter data points that meet specific models through consistent strategies or preference analysis. However, these methods are difficult to effectively capture the context information of the data when dealing with a large number of outliers, which adversely affects the quality of model fitting. To address these issues, the corresponding learning methods proposed in recent years use deep neural networks to extract data context information and achieve data purification, without relying on model hypothesis generation, thus providing a promising alternative. These methods enhance the generalization ability of the network and reduce the complexity of the model by avoiding explicit model hypothesis generation. However, this type of method also has significant limitations. On the one hand, convolutional neural networks are mainly used to process ordered data, while model fitting essentially involves unordered and unevenly distributed data, which increases the complexity of the convolution process and limits the effective capture of global context information; on the other hand, these methods usually ignore the joint use of preference information and consistency information, which are crucial for robust model fitting and outlier removal. Therefore, how to efficiently integrate preference information and consistency information without generating explicit model hypotheses, and extract key context features from unordered data, has become a key research problem in the field of robust model fitting. SUMMARY

[0004] The purpose of the present application aims to overcome the deficiencies in the prior art, provide a robust model fitting method based on double purification dynamic graph network, which combines preference analysis and consistent strategy, effectively alleviates the interference of noise and outliers, improves the estimation accuracy of model parameters without generating model assumptions, and generates high-quality models. The method uses dynamic graph network and attention mechanism to dynamically purify data from the dual perspectives of data points and hypothesis, and realizes robust model fitting.

[0005] The present application proposes a double dynamic purification module, which uses dynamic graph structure for step-by-step sampling and purification at the regional and overall levels. Unlike traditional methods that rely on consistent strategy or preference analysis, the proposed method introduces regional channel preference blocks and overall space consistent blocks. The regional channel preference block captures the data point region features corresponding to the hypothesis through the preference attention mechanism, enhances the local context information and improves the preference feature extraction capability; the overall space consistent block extracts global consistent information through the consistent attention mechanism, and generates consistent weights to distinguish outliers and inner points. By dynamically constructing regional and overall graphs, the double dynamic purification module effectively eliminates low-weight outliers and retains high-weight candidate values, thereby significantly improving the precision and robustness of model fitting.

[0006] The present application comprises the following steps:

[0007] A. Input the data set containing noise and outliers, and use the scale invariant feature transformation method to generate an initial correspondence set;

[0008] B. The initial correspondence set obtained in step A is step-by-step sampled and outlier-removed by the double dynamic purification module to generate a purified candidate correspondence set and an inner point probability set;

[0009] C. Construct a regional channel preference block, in the double dynamic purification module of step B, use the regional channel preference block to partition the initial correspondence set and extract multi-channel preference features; capture preference context information through the regional preference attention mechanism, construct a dynamic regional graph to generate preference features of query, key and value; obtain preference weight to strengthen inner point detection, and generate purified preference features through preference weight and multi-region feature aggregation, to guide the estimation and optimization of inner point probability from a local perspective;

[0010] D. Construct an overall space consistent block, input the preference features obtained in step C into the overall space consistent block, use the space consistent attention mechanism to extract global consistency information, and generate initial consistent weights; calculate the consistent affinity by constructing a dynamic overall graph, and integrate consistent context information by combining the weighted adjacency matrix and graph convolution operation to generate overall consistent features to improve the reliability of inner point selection and the precision of model fitting; obtain the purified candidate correspondence set and the inner point probability set through the point cloud convolution block and the multi-layer perception layer;

[0011] E. The cleaned candidate correspondence set and the inlier probability set obtained in step D are input into the weighted eight-point method solver to estimate the model and its model parameters;

[0012] F. The model parameters and the inlier probability set obtained in step E are input into the predictor to calculate the model estimation loss and the point classification loss, which are used to optimize the model fitting process.

[0013] In step A, the original data set containing noise and outliers is input, and the scale-invariant feature transformation method is used to construct the initial correspondence set, denoted as , where each represents a hypothetical correspondence, represents the number of channels, represents the number of hypothetical correspondences, and represent the normalized coordinates of the th data point in the two corresponding images; the normalization operation is based on the camera intrinsic parameters of the image pair , which maps the spatial coordinates to the standard range ;

[0014] In step B, the initial correspondence set obtained in step A is processed by the double dynamic cleaning module, which gradually samples data points and removes outliers; this module combines preference analysis and consistency strategy to dynamically explore the interaction between data points and hypothetical correspondences without generating explicit model assumptions, thereby significantly reducing the negative impact of outliers on model fitting; the double dynamic cleaning module consists of a regional channel preference block and a global spatial consistency block, generating a cleaned candidate correspondence set and an inlier probability set , denoted as , where represents the proposed dynamic graph network, represents the network parameters; is the inlier probability set, where is the inlier probability of the th data point, and when , is considered as an inlier, otherwise as an outlier;

[0015] In step C, the double dynamic cleaning module in step B constructs a regional channel preference block to perform multi-region division on the initial correspondence set and extract multi-channel preference features, generating preference weights to enhance inlier detection and optimize model inlier probability evaluation; the regional channel preference block designs a partition residual structure to split the initial correspondence set into , where The number of regions in the channel dimension is represented, the preference context information from different regions is captured, and the diversity and robustness of feature extraction are enhanced. First, a region preference attention mechanism is proposed to dynamically capture the preference features in each region, and the preference weight of the channel is measured by using residual information, so as to strengthen the detection ability of the inner point and highlight the representativeness of the important region. The extraction of the preference feature is based on the query , keyword and value , which are represented as follows:

[0016] (1)

[0017] Wherein, represents a cascade operation, represents an average pooling operation and a maximum pooling operation on the first region; second, a dynamic region graph is constructed for each region, which is represented as follows:

[0018] (2)

[0019] Wherein, , the vertex of which represents the first nearest neighbor of the feature , and the edge represents the affinity of and its neighborhood; the dynamic region graph is dynamically updated according to the feature order at each layer of the network, which is different from the fixed graph structure in the traditional convolutional neural network; then, the region channel preference block further aggregates the preference features by using the ring convolution layer, the point cloud convolution block and the multi-layer perception layer to obtain the preference weight , and enhances the feature diversity by inheriting the preference weight of the previous region, which is represented as follows:

[0020] (3)

[0021] (4)

[0022] Wherein, represents a normalized exponential activation function, represents an element product, represents an element sum; by integrating the preference features of all regions, the purified preference features are generated, and they are used to guide the evaluation and corresponding optimization of the model inner point probability; the region channel preference block focuses on the support tendency of a single data point and the corresponding hypothesis, rather than the consistency of the whole data set; it significantly improves the feature extraction ability in complex data and the robustness of model fitting, and provides high-quality input features for the subsequent region channel preference block;

[0023] In step D, the preference features obtained in step C are introduced into the whole space consistent block, the consistency information of the features is extracted and strengthened through the space consistent attention mechanism, and the initial consistent weight is generated , and a dynamic whole graph is constructed based on the consistent weight; combined with the weighted adjacency matrix and the graph convolution operation, the whole consistent feature is generated to improve the reliability of the inner point extraction and the accuracy of the model fitting; specifically, the whole space consistent block identifies and refines the irregular feature space between the inner points and the outliers through the space consistent attention mechanism, emphasizes the consistent features, suppresses irrelevant features, and improves the integrity and accuracy of the feature representation; which is expressed as follows:

[0024] (5)

[0025] wherein, represents a sigmoid activation function, represents batch normalization and multilayer perceptron operation; the consistent weight describes the consistency of the data points in the key feature space, and provides a basis for the construction of the dynamic whole graph; the dynamic whole graph is constructed to establish the long-range dependency relationship between the data points; the vertices of the dynamic whole graph represent the hypothesis correspondence, and the edges represent the consistent affinity between the vertices, which is expressed as follows:

[0026] (6)

[0027] wherein, represents the dynamic whole graph, and the edges can be represented as ; only when two correspondences have a high consistent weight, a strong association is established; otherwise, the association is weak or non-existent; the whole consistent feature is derived through the establishment of the weighted adjacency matrix and the graph convolution operation, which is expressed as follows:

[0028] (7)

[0029] wherein, is the diagonal matrix of the weighted adjacency matrix, is a learnable spectral matrix used to propagate consistent properties; the dynamic whole graph is dynamically updated between different network layers, and the output whole consistent feature The consistency of the features at the global and spatial levels is reflected, especially between highly correlated inliers; then, the relevance of the disordered data is further purified by the point cloud convolution block, the invariance of the disordered data is processed, and the consistent features are encoded by combining the multi-layer perceptron to estimate the inlier probability set; this process strengthens the ability to distinguish inliers from outliers and ensures the preservation of key data points; to avoid missing effective inliers, the consistent weight is gradually passed to the DDP module of the next stage for further optimization, thereby improving the reliability of inlier extraction and the accuracy of model fitting; through global optimization of the global spatial consistency block, reliable support is provided for subsequent model fitting;

[0030] In step E, the purified candidate correspondence set and the inlier probability set obtained in step D are input into the weighted eight-point method solver to estimate the model and its parameters, which can be represented as wherein, represents the model estimated by the proposed network, represents the model parameters estimated by the proposed network, represents the weighted eight-point method solver, and are the hyperbolic tangent and the rectified linear unit activation functions, respectively; the estimated model on the candidate correspondence set is tested and verified through the validation operation, and the optimized model is obtained, which can be represented as wherein, represents the full-size validation function used to evaluate the fitting accuracy and reliability of the model;

[0031] In step F, the model parameters and the inlier probability set obtained in step E are input into the predictor to calculate the model estimation loss and the point classification loss , which are used to optimize the network performance, and are represented as follows:

[0032] (8)

[0033] (9)

[0034] wherein, is the binary cross-entropy loss function, represents the classification score of the inliers and outliers in the th double dynamic purification module, represents the true label of the inliers and outliers in the th double dynamic purification module, is the number of double dynamic purification modules; is the true model parameter, and for the corresponding coordinate set of the image pair, and respectively represent the first and the second element in the vector.

[0035] Compared with the prior art, the present application has the following outstanding technical effects and advantages:

[0036] 1) A robust model fitting network combining preference analysis and consistent strategy is proposed, which effectively purifies outliers and noise from the dual perspectives of data points and hypothesis correspondence based on dynamic graph network, significantly improving the accuracy and robustness of model fitting.

[0037] 2) An effective dual dynamic purification module is designed, which extends from local area to global space through dynamic sampling and optimization, effectively reducing the interference of disordered data and random abnormal distribution. This module can achieve efficient data point screening and outlier elimination without generating explicit model assumptions.

[0038] 3) A regional channel preference block and an overall space consistent block are proposed, which use dynamic graph structure and attention mechanism to accurately estimate preference scores and consistent scores respectively. The regional channel preference block focuses on feature extraction and optimization in local areas, and the overall space consistent block analyzes the long-range dependence relationship between data points from a global perspective. Through the synergistic effect of the two, a large number of outliers are effectively purified. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a schematic diagram of a robust model fitting method based on a dual purification dynamic graph network according to the present application.

[0040] Figure 2 is a schematic diagram of a regional channel preference block according to the present application.

[0041] Figure 3 is a schematic diagram of an overall space consistent block according to the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical scheme and advantages of the present application clearer, the following examples will further illustrate the present application with reference to the accompanying drawings. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0043] As shown in Figure 1 , the implementation of the embodiment of the present application includes the following steps:

[0044] A. Given a data set containing noise and outliers, an initial correspondence set is constructed by a scale-invariant feature transformation method as follows: where represents a hypothesis correspondence,​ represents the number of channels, represents the number of hypothesis correspondences, and represents the spatial position of the th data point pair within the range; these positions are normalized by the intrinsic functions of the cameras of the image and pairs. The normalization operation is based on the intrinsic parameters of the image pairs, mapping the spatial coordinates to the standard range .

[0045] B. Introduce the initial correspondence set obtained in step A into the dual dynamic purification module, which integrates preference analysis and consistency strategy together, gradually explores the interaction between data points and hypothesis correspondences, and dynamically purifies the adverse effects of outliers without generating model hypotheses; the dual dynamic purification module consists of a regional channel preference block and an overall spatial consistency block; the two blocks are trained collaboratively to improve the performance of the network; the regional channel preference block estimates a set of preference weights from the initial correspondence set , and then calculates a set of consistency weights through the overall spatial consistency block; this process produces a set of purified candidate correspondences and a set of inlier probabilities of these candidate correspondences , which can be represented as , where represents a learnable network, represents network parameters; the set of inlier probabilities is represented as , where is assigned to ; if , then is considered as an inlier, otherwise is considered as an outlier.

[0046] C. In the dual dynamic purification module of step B, a regional channel preference block is constructed, as shown in Figure 2 , for accurately extracting preference information required for model fitting from the initial correspondence set; the initial correspondence set is divided into multiple regions by designing a regional residual structure, represented as , where represents the number of regions in the channel dimension, capturing preference context information from multiple channels; the channel properties of each region extract preference features through regional division, and combine with regional preference attention mechanism to enhance the detection ability of inliers, emphasizing the importance of representative channels, while improving robustness and accuracy in complex data; the regional preference attention mechanism utilizes the residual information of each region to generate query , key , and value and aggregate the preference features within the region, denoted as follows:

[0047] (1)

[0048] wherein, denotes the concatenation operation, denotes the average pooling and max pooling operation along the dimension.

[0049] Subsequently, a dynamic region graph is constructed based on the preference features, the vertices of the dynamic region graph representing the nearest neighbor data point sets in the feature space, and the edges representing the affinity between the data points within the neighborhood, denoted as follows:

[0050] (2)

[0051] wherein, , denotes the dynamic region graph of the th region, denotes the set of field vertices, denotes the corresponding set of directed edges; the dynamic region graph is dynamically updated according to the feature order at each layer of the network, unlike the fixed graph structure in traditional convolutional neural networks; the region channel preference block further aggregates the preference features to derive the preference weights using a ring convolution layer, a point cloud convolution block, and a multi-layer perception layer , and enhances the feature diversity by inheriting the preference weights of the previous region, denoted as follows:

[0052] (3)

[0053] (4)

[0054] wherein, is a normalized exponential activation function, denotes the element product, denotes the element sum; by integrating the preference features of all regions, the purified preference features are generated; the reliability of the preference weights is used to guide the initial inlier probability estimation and optimization in the dual dynamic purification module; the region channel preference block enhances the accuracy of local analysis by improving the corresponding ability of the single data point support hypothesis, and further optimizes the performance and robustness of the model fitting with the assistance of the global spatial consistency block.

[0055] D. The preference features obtained in step C are introduced into the global spatial consistency block, as Figure 3As shown, the model fitting is effectively performed by capturing consistent information; first, the spatial consistent attention mechanism is used to identify and refine the irregular feature space of inliers and outliers, and the integrity and accuracy of feature representation are improved by emphasizing consistent features and suppressing irrelevant features; based on the initial consistent weight The spatial consistent attention mechanism is calculated as follows:

[0056] (5)

[0057] where, is a sigmoid activation function, represents batch normalization and multi-layer perceptron operations; the consistent weight is used to describe the consistency information of data points in the feature .

[0058] Then, a dynamic global graph is constructed to establish long-range dependency relationships between data points; the vertices of the dynamic global graph represent the corresponding hypotheses, and the edges represent the consistent affinity between vertices, which is represented as follows:

[0059] (6)

[0060] where, represents the dynamic global graph, and the elements of the edge can be represented as , where, represents the consistent affinity between vertices; only when two correspondences have a high consistent weight, a strong association is established; otherwise, the association is weak or non-existent; the global consistent feature is derived by constructing a weighted adjacency matrix and graph convolution operation, which is represented as follows:

[0061] (7)

[0062] where, is the diagonal matrix of the weighted adjacency matrix, is a learnable spectral matrix used to propagate consistent properties; the dynamic global graph is dynamically updated between different network layers, and the output global consistent feature reflects the feature consistency at the global and spatial levels, especially between highly correlated inliers.

[0063] Finally, the point cloud convolution block is used to further purify the relevance of unordered data and handle the invariance of unordered data; the consistent feature is further encoded by a multi-layer perceptron, which is used to estimate the inlier probability set; to avoid misclassification of valid inliers, the consistent weight is gradually passed to the next dual dynamic purification module for further optimization, thereby improving the reliability of inlier extraction and the accuracy of model fitting.​

[0064] E. The purified candidate correspondence set and the inlier probability set obtained in step D are input into the weighted eight-point solver to estimate the model and the model parameters; this process can be represented as wherein, represents the model estimated by the proposed network, represents the model parameters estimated by the proposed network, represents the weighted eight-point solver, and is an activation function; the validation operation is used to test and verify the estimated model on the candidate correspondence set , and obtain the final model , which is represented as wherein, represents the full-size validation function.

[0065] F. The model parameters and the inlier probability set obtained in step E are input into the predictor to calculate the model estimation loss and the point classification loss , which are used to optimize the network performance, and are represented as follows:

[0066] (8)

[0067] (9)

[0068] wherein, is a binary cross-entropy loss function, represents the classification score of the inliers and outliers of the th dual dynamic purification module, represents the true label of the inliers and outliers of the th dual dynamic purification module, is the number of dual dynamic purification modules; is the true model parameter, and are the correspondence coordinate sets of the image pairs, and represent the th and the th elements in the vectors, respectively.

[0069] To comprehensively evaluate the performance of the method of the present application, the present application is tested on two data sets, outdoor YFCC100M and indoor SUN3D, and compared with the existing state-of-the-art methods; in the experiment, SIFT is used to generate initial correspondences, and the dimension of each correspondence is ; the region division dimension is ​is set to 4; in order to realize efficient processing, two continuous double dynamic purification modules are constructed; in terms of data set division, the known scenes in the YFCC100M and SUN3D data sets are divided into a training set, a verification set and a test set according to a ratio of 6:2:2; during the training process, the learning rate is set to , so as to ensure efficient learning and stable convergence of the model.

[0070] Table 1 is the average precision comparison experiment result of the basic matrix estimation on the YFCC100M and SUN3D data sets.

[0071] Table 1

[0072]

[0073] The experimental results show that the performance of the present application on the two data sets is better than that of the existing most advanced method. In the unknown outdoor scene, the present application improves the best hypothesis generation method (PARSAC) by 7.54% and 6.01% respectively in the mAP5º and mAP20º indicators; in the known outdoor scene, the present application improves the best corresponding learning method (SSLNet) by 7.01% in the mAP20º indicator. In addition, the present application also shows significant performance improvement in indoor scenes. These results fully prove that the present application effectively reduces the adverse effects of outliers on model fitting by combining consistent strategy and preference analysis, and at the same time shows its robustness and generalization ability in various scenes.

[0074] Table 2 is the average precision comparison experiment result of the essential matrix estimation on the YFCC100M and SUN3D data sets.

[0075] Table 2

[0076]

[0077] The experimental results show that the present application is significantly better than the existing method on known and unknown indoor and outdoor data sets. This excellent performance verifies the significant advantage of the learning-based method in inlier identification compared with the traditional hypothesis generation method. On the YFCC100M data set of unknown scenes, the present application improves MAGSAC++ by 19.30%. Even in the case of large scene changes, the dynamic graph learning and attention mechanism embedded in the present application can accurately identify inliers, showing excellent adaptability. In addition, compared with the consistent corresponding learning method or the preference analysis method, the present application shows stronger robustness to different data characteristics, can effectively refine the corresponding relationship, and at the same time maintains the flexibility to noise interference and complex spatial relationship, further improves the reliability and accuracy of model fitting.

[0078] Table 3 is the precision (P), recall (R) and F-score (F) of the outlier rejection on YFCC100M and SUN3D datasets.

[0079] Table 3

[0080]

[0081] The quantitative results of the experiment show that on the two datasets, the present application is superior to other methods in terms of P and F. However, it is slightly lower than some of the comparative methods in terms of R. This is because in the process of gradually rejecting a large number of outliers, the present application inevitably misdeletes some ambiguous inliers. Although there is a certain gap in the recall rate, the present application performs best in terms of comprehensive performance (F). In addition, in the unknown scene of the SUN3D dataset, the present application improves the P index by 5.72% compared with the second best method (SSLNet). These results fully prove the significant advantage of the present application in dealing with outliers, and demonstrate its excellent flexibility and robustness in learning and capturing intrinsic contextual information.

[0082] A large number of experiments show that the proposed robust model fitting method based on double purification dynamic graph network has significant advantages in improving the model fitting performance.

[0083] The above embodiments are only the preferred embodiments of the present application and cannot be considered to limit the scope of the present application. Any equivalent changes and improvements made in accordance with the scope of the present application should still belong to the scope of the present application.

Claims

1. A robust model fitting method based on a dual-purification dynamic graph network, characterized in that... Includes the following steps: A. Input a dataset containing noise and outliers, and generate an initial correspondence set using the scale-invariant feature transformation method; The specific method for generating the initial correspondence set is to extract the hypothetical correspondence set through scale-invariant feature transformation, denoted as: ,in, Indicates the number of channels. This indicates the corresponding quantity. Indicates the first Each hypothesis corresponds to a range. Internal spatial location and These locations are determined by the image. and The corresponding camera intrinsic parameters are normalized to standardize the spatial coordinates; B. The initial correspondence set obtained in step A is gradually sampled and outliers are removed through a dual dynamic purification module to generate a purified candidate correspondence set and an interior point probability set. C. Constructing regional channel preference blocks: In the dual dynamic purification module of step B, the initial corresponding set is partitioned using regional channel preference blocks to extract preference features from multiple channels; preference context information is captured through a regional preference attention mechanism, and a dynamic regional graph is constructed to generate preference features for queries, keys, and values; preference weights are obtained to enhance inlier detection, and the purification-related preference features are generated by aggregating preference weights and multi-regional features to guide the estimation and optimization of inlier probabilities from a local perspective; D. Construct a global spatial consistency block, input the preference features obtained in step C into the global spatial consistency block, extract global consistency information using the spatial consistency attention mechanism, and generate initial consistency weights; calculate consistency affinity by constructing a dynamic global graph, and integrate consistency context information by combining a weighted adjacency matrix with graph convolution operations to generate global consistency features to improve the reliability of inlier selection and the accuracy of model fitting; obtain the purified candidate correspondence set and inlier probability set through point cloud convolutional blocks and multilayer perceptron layers; E. Input the purified candidate correspondence set and interior point probability set obtained in step D into the weighted eight-point solver for model fitting to estimate the model and its parameters, expressed as follows: ,in, This represents the proposed network estimation model. This represents the model parameters estimated by the proposed network. This indicates a weighted eight-point solver. and These are hyperbolic tangent and modified linear unit activation functions, respectively; the candidate correspondence set is validated through a verification operation. Estimation model on Conduct testing and validation to obtain an optimized model. , represented as ,in, This represents a full-size validation function used to evaluate the model's fitting accuracy and reliability. F. Input the model parameters and inlier probability set obtained in step E into the predictor, calculate the model estimation loss and point classification loss, and use them to optimize the model fitting process.

2. The robust model fitting method based on a dual-purification dynamic graph network as described in claim 1, characterized in that... In step B, the initial correspondence set obtained in step A is progressively sampled and outlier removed through a dual dynamic purification module. The dual dynamic purification module integrates preference analysis and consistency strategy to progressively explore the interaction between data points and hypothesis correspondences, and dynamically purifies the interference of outliers without generating model hypotheses. The dual dynamic purification module includes: (1) a regional channel preference block, used to select from the initial corresponding set The algorithm estimates the preference weights, extracts multi-channel preference features through partitioning operations, and captures preference context information using dynamic region graphs; (2) a global spatial consistency block is used to calculate consistency weights based on preference features and captures consistency context information through dynamic global graphs to enhance the reliability of interior point extraction; the dual dynamic purification module trains the above two blocks in tandem to improve network performance and generate a purified candidate correspondence set. and the corresponding interior point probability set , represented as ,in, Indicates the network mentioned. Represents network parameters; interior point probability set is represented as ,in Indicates correspondence The probability of an interior point; when hour, It is considered an interior point; otherwise, it is considered an outlier.

3. The robust model fitting method based on a dual-purification dynamic graph network as described in claim 1, characterized in that... In step C, a regional channel preference block is constructed to accurately extract the preference information required for model fitting from the initial correspondence set; the regional channel preference block divides the initial correspondence set into multiple regions by designing a partitioned residual structure. , It represents the number of regions in the channel dimension, capturing preference context information from multiple channels; the channel attributes of each region extract preference features through region division, and combined with the region preference attention mechanism, it enhances the detection capability of inliers, emphasizes the importance of representative channels, and improves robustness and accuracy in complex data; The region preference attention mechanism utilizes the residual information of each region to generate queries. Keywords Sum , means as follows: （1） in, Indicates a cascading operation. Indicates along the first Perform average pooling and max pooling operations on each dimension; Subsequently, a dynamic region graph is constructed based on preference information. The vertices of the dynamic region graph represent the set of nearest neighbor data points in the feature space, and the edges represent the affinity between data points in the neighborhood, as shown below: （2） in, , Indicates the first Dynamic region map of each area Represents a vertex set. The directed edge set represents affinity; the dynamic region graph is dynamically updated at each layer of the network according to the feature order, unlike the fixed graph structure in traditional convolutional neural networks; the region channel preference block further utilizes circular convolutional layers, point cloud convolutional blocks, and multi-layer perceptron layers to aggregate preference features and derive preference weights. By inheriting the preference weights and preferences from the previous region, feature diversity is enhanced, as shown below: （3） （4） in, The normalized exponential activation function is... Represents the element-wise product. Represents elements and; by integrating the preference features of all regions, a purified preference feature is generated. , represented as The reliability of preference weights is used to guide the initial interior point probability estimation and optimization in the dual dynamic cleanup module; the regional channel preference block enhances the accuracy of local analysis by improving the ability of individual data points to support the corresponding assumptions, and further optimizes the performance and robustness of model fitting with the assistance of the overall spatial consistency block.

4. The robust model fitting method based on a dual-purification dynamic graph network as described in claim 1, characterized in that... In step D, a global spatially consistent block is constructed to effectively fit the model by capturing consistent information. This global spatially consistent block uses consistent features to distinguish between inliers and outliers. First, a spatially consistent attention mechanism is used to identify and refine the irregular feature space of inliers and outliers, improving the completeness and accuracy of feature representation by emphasizing consistent features and suppressing irrelevant features. Based on... Initial consistent weights Calculated using the spatially consistent attention mechanism, it is represented as follows: （5） in, It is the sigmoid activation function. This represents batch normalization and multilayer perceptron operations; consistent weights are used to describe features. Consistency information of data points in the middle; Then, a dynamic global graph is constructed to establish long-range dependencies between corresponding vertices; the vertices of the dynamic global graph This indicates a hypothetical correspondence, with edges... The uniformity affinity between vertices is represented as follows: （6） in, Represents a dynamic overall graph, with edges The element is represented as ,in, This represents the consistent affinity between vertices; a strong association is established only when two correspondences have a high consistency weight; otherwise, the association is weak or nonexistent; this is achieved by establishing a weighted adjacency matrix. The graph convolution operation derives globally consistent features, as shown below: （7） in, It is a diagonal matrix of the weighted adjacency matrix. A learnable spectral matrix is ​​used to propagate consistent properties; the dynamic global graph is dynamically updated across different network layers, outputting overall consistent features. It reflects the consistency of characteristics at the overall and spatial levels, especially among highly related inner layers; Finally, the correlation of the disordered data is further purified by point cloud convolutional blocks to handle the invariance of the disordered data; the consistent features are further encoded by a multilayer perceptron to estimate the probability set of inliers; to avoid mis-selecting valid inliers, the consistent weights are gradually passed to the next dual dynamic purification module for further optimization, thereby improving the reliability of inlier extraction and the accuracy of model fitting.

5. The robust model fitting method based on a dual-purification dynamic graph network as described in claim 1, characterized in that... In step F, the model parameters and interior probability set obtained in step E are input into the predictor to calculate the model estimation loss. Sum of point classification loss Used to optimize network performance, as shown below: （8） （9） in, The binary cross-entropy loss function is... The logit value represents the number of... The classification scores of interior points and outliers in the dual dynamic purification module. The true classification score represents the score of the first category. The true labels of interior points and outliers in each dual dynamic purification module. This refers to the number of dual dynamic purification modules; These are the parameters of the actual model. and This is the set of coordinates corresponding to the image pairs. and They represent the first and second elements in the vector. and the Each element.

Images