A method and system for high-dimensional feature matching of multi-view images

Abstract

This invention discloses a high-dimensional feature matching method and system for multi-view images, belonging to the fields of computer vision and remote sensing mapping technology. The method includes the following steps: S1: Extracting high-dimensional feature descriptors and feature point coordinates from the multi-view images to construct a feature dataset; S2: Encoding the high-dimensional feature descriptors into binary form based on the Multi-View Feature Adaptive Hash (MFAH) method, and then generating initial multi-view feature trajectories through fast bitwise operations in the hash space; S3: Performing gross error removal on the initial multi-view feature trajectories, preferably by introducing a Multi-View Local Geometric Consistency (MLGC) strategy, to obtain an optimized set of multi-view feature trajectories. The technical solution of this invention can achieve efficient simultaneous matching and lightweight gross error removal of high-dimensional features from multiple views, significantly improving matching efficiency while maintaining considerable matching quality. It is suitable for multi-view image processing scenarios such as large-scale 3D reconstruction and remote sensing image change monitoring.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and remote sensing mapping technology, specifically relating to a method and system for achieving fast and simultaneous matching of high-dimensional features and lightweight gross error removal for large-scale multi-view images by utilizing feature adaptive hashing and local geometric consistency constraints. Background Technology

[0002] Multi-view image matching is a core technology for large-scale 3D reconstruction, visual localization, and remote sensing image analysis. By establishing feature correspondences between images from different viewpoints, it provides a foundation for camera pose estimation, point cloud triangulation, and scene understanding. Existing multi-view image matching mainly employs a framework of pairwise matching and dual-view verification. This framework first decomposes the image set into multiple sets of common-view image pairs, then performs feature matching and gross error removal independently for each pair, and finally merges the matching results into a multi-view feature trajectory. However, this type of method has two major drawbacks: first, it is inefficient, as the number of image pairs increases quadratically with the total number of images, leading to a sharp increase in the computational cost of matching and gross error removal, making it difficult to adapt to large-scale multi-view image processing; second, it has poor reliability, as gross error removal relies on dual-view verification methods, and unremoved gross errors are amplified during trajectory merging, leading to continuous trajectory errors and ultimately affecting the quality of subsequent tasks.

[0003] Compared to the inherent limitations of pairwise matching, simultaneous multi-view matching strategies hold greater potential. These methods aim to unify the modeling of feature correspondences across all images within a unified framework, directly yielding globally consistent multi-view feature trajectories, theoretically offering higher matching efficiency and trajectory reliability. However, existing simultaneous multi-view matching methods still face two core challenges: First, high-dimensional feature aggregation is difficult. When processing a large number of images simultaneously, the aggregated feature data forms massive high-dimensional matrices, leading to a severe computational burden. Second, there is a lack of lightweight simultaneous gross error removal schemes. Existing methods often simply extend dual-view gross error removal strategies to multi-view scenarios, sacrificing the global consistency advantage of simultaneous matching and further increasing computational costs, making it difficult to achieve efficient processing while maintaining matching accuracy.

[0004] To address the aforementioned issues, a multi-view image matching method is needed that can efficiently aggregate high-dimensional features and simultaneously remove gross errors. This method should improve matching efficiency while ensuring the reliability of matching results, thus meeting the needs of large-scale multi-view image processing. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing multi-view image matching methods, such as low efficiency, difficulty in high-dimensional feature aggregation, poor gross error removal, and low adaptability. It provides a high-dimensional feature matching method and system for multi-view images. This invention achieves efficient aggregation and simultaneous matching of high-dimensional features through a feature adaptive hashing method. Furthermore, it preferably combines a multi-view local geometric consistency strategy to achieve lightweight simultaneous gross error removal, significantly improving matching efficiency while maintaining matching quality comparable to pairwise matching.

[0006] Therefore, the present invention provides the following technical solution: On one hand, the present invention provides a high-dimensional feature matching method for multi-view images, comprising the following steps: S1: Feature construction, extracting high-dimensional feature descriptors and feature point coordinates from multi-view images to construct a feature dataset; S2: Trajectory generation. The high-dimensional feature descriptor is encoded based on the Multi-View Feature Adaptive Hash (MFAH) method. Then, the initial multi-view feature trajectory is generated through fast bit operations in the hash space. Each multi-view feature trajectory represents the correspondence between feature points of the same physical point in space in various images. S3: Trajectory correction, which removes gross errors from the initial multi-view feature trajectory to obtain the optimized multi-view feature trajectory.

[0007] Optionally, the specific process of trajectory generation in step S2 includes: S21: Initialize hash parameters and generate a one-dimensional Gaussian distribution. random projected hash matrix ; S22: By randomly projecting the hash matrix Each high-dimensional feature descriptor of each image is binary encoded to obtain a hash code, thereby mapping the high-dimensional features to a low-dimensional hash space; S23: Calculate the distance between feature points of different images based on hash codes, aggregate feature points with a distance less than a preset threshold into the same hash bucket, and ensure that the number of feature points of a single image in each hash bucket is no greater than 1. Optionally, step S21 may also define a multi-level hash bit-width sequence. , For the first Layer hash bit width, S The number of bit-width sequence layers; wherein, based on the multi-layer hash bit-width sequence A progressive trajectory generation strategy is adopted, specifically as follows: For the first-layer hash bit-width sequence, multi-view feature trajectories are generated through steps S22-S24; For the feature points that were not aggregated in step S23, the next layer of trajectory generation is entered, that is, using the second layer hash bit width sequence, multi-view feature trajectories are generated for the feature points that were not aggregated through steps S22-S24. By iterating sequentially, the final multi-view feature trajectory is obtained; The hash code based on binary encoding is: ; In the formula, For random projection hash matrix random vectors in For the first The first image A high-dimensional feature descriptor, For a single feature descriptor The corresponding binary hash code is defined as follows: ; L The hash bit width of a certain layer, These are the 1st, 2nd, and 3rd bits of the hash code width in the first layer, respectively. L Each binary hash code has a value of 0 or 1.

[0008] Optionally, the distance mentioned in step S23 is the Hamming distance, and the corresponding formula is:

[0009] in, For the first The first image The hash code of each feature, For the first The first image The hash code of each feature, For the first The value of the bit hash code, For the first The first image The first feature and the second The first image Hamming distance for each feature.

[0010] Optionally, the process of introducing the multi-view local geometric consistency strategy MLGC to remove gross errors from the initial multi-view feature trajectories in step S3 includes at least the following: S31: Initialize MLGC parameters, set the number of nearest neighbors. Topological similarity threshold Regularization parameters And determine the neighborhood of each feature point in the corresponding image; S32: Geometric consistency loss calculation, based on normalized topological similarity of the displacement vectors from feature points to their neighborhoods, and according to a topological similarity threshold. Quantize the geometric consistency loss of the displacement vector; S33: Construct a cost function based on geometric consistency loss, determine whether feature points are retained by minimizing the cost function, obtain an initial set of interior point feature trajectories, and realize trajectory update.

[0011] Optionally, the formula for calculating normalized topological similarity in step S32 is:

[0012] in, To normalize topological similarity, For the first The first image The feature point is located in its neighborhood. The displacement vectors of the nearest neighbor points It is the first The first feature point and the second The nearest neighbor points correspond to the displacement vectors in the u-th image; The total number of images, Let be the magnitude of the displacement vector. The cosine similarity of the displacement vectors. .

[0013] Optionally, the cost function in step S33 is: ; in, Let cost function be This represents the total number of current multi-view feature trajectories. The total number of images, For the first The first of the trajectories Binary determination values ​​of feature points in an image. For the first The trajectory in the first Geometric constraint cost of feature points in an image; For the current set of multi-view feature trajectories, Represents feature points.

[0014] Optionally, geometric constraint cost Satisfying: ; in, K For the neighborhood The number of nearest neighbors in the data. For the first Located on the trajectory of the first Feature points of an image for The Nearest neighbor points, They are respectively and Corresponding to the Matching points of the images; This represents the Euclidean distance between two points. For feature points In the Neighborhood under the image, For feature points In the The neighborhood under the image; The rule for solving the binary determination value of feature points is as follows: ; When it is 1, the first one is reserved. Corresponding feature points on the trajectory; Remove the first when the value is 0. The corresponding feature points on the trajectory.

[0015] Secondly, the present invention also provides a matching system based on the above-mentioned high-dimensional feature matching method for multi-view images, comprising: Feature extraction module: used to extract high-dimensional feature descriptors and feature point coordinates from multi-view images and construct a feature dataset; Hash matching module: used to encode high-dimensional feature descriptors based on the Multi-View Feature Adaptive Hash (MFAH) method, and then generate initial multi-view feature trajectories through fast bit operations in the hash space. Each multi-view feature trajectory represents the correspondence between feature points of the same physical point in space in various images. The geometry optimization module is used to remove gross errors from the initial multi-view feature trajectory to obtain the optimized multi-view feature trajectory. The hash matching module is signal-connected to the geometry optimization module, and the output of the feature extraction module is signal-connected to the input of the hash matching module.

[0016] Thirdly, the present invention also provides an electronic device, including one or more processors and a memory storing one or more computer programs, wherein when the processor invokes the computer programs, it executes the steps of the high-dimensional feature matching method for the multi-view image.

[0017] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when invoked by a processor, executes the steps of the high-dimensional feature matching method for the multi-view image.

[0018] Compared with the prior art, the present invention achieves the following progress and effects: 1. This invention innovatively transforms the traditional random hyperplane hash algorithm to adapt to multi-view scenarios, proposing a multi-view feature adaptive hash matching method (MFAH). Through adaptive hash encoding and cross-view hash code retrieval, it achieves simultaneous aggregation of similar hash codes for all view features within a unified hash space. This method transforms the matching problem into a synchronous aggregation problem of hash codes within the same space, eliminating the need for pairwise processing in traditional pairwise matching. This design ensures the reliability of hash matching in multi-view scenarios while reducing the complexity of multi-view matching from... Reduced to linear .

[0019] Compared to the existing technology CN106886785A, which only supports dual-view aerial image matching and requires feature vector mapping and encoding via kernel methods, this invention innovatively optimizes the traditional random hyperplane hash algorithm for multi-view scene adaptability. It eliminates the kernel space mapping step and proposes a multi-layer adaptive hash bit-width sequence, cross-view single-feature-point hash bucket aggregation rules, and a multi-probe neighborhood retrieval mechanism in the preferred scheme. This allows for the direct association of unique feature points from each view within a bucket into a set based on image perspective, generating feature trajectories. The encoding process is simpler and more efficient, eliminating the need for complex kernel mapping and similarity matrix construction.

[0020] Therefore, the multi-view matching framework provided by this invention significantly improves efficiency while ensuring the reliability of matching and the quality of subsequent 3D reconstruction. The reprojection error of the reconstructed sparse point cloud is comparable to that of the traditional pairwise matching method, and the camera pose deviation is extremely small. It achieves a balance between efficiency and accuracy and is suitable for multi-view image processing in large-scale and complex scenarios.

[0021] 2. Addressing the technological gap in existing methods for synchronous mismatch removal across multiple views, this invention further optimizes the scheme by introducing a novel Multi-View Local Geometric Consistency (MLGC) strategy. This method overcomes the limitation of traditional MLGC being applicable only to two views, proposing a method for calculating the topological similarity of the neighborhood in multiple views and a linear complexity cost function. This method achieves lightweight global mismatch removal across multiple views without the need for complex solvers, accounting for only 5%~10% of the matching process runtime. The two-stage threshold strategy, from coarse to fine, effectively reduces the possibility of gross errors and ensures the reliability of trajectory quality. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the multi-view image high-dimensional feature simultaneous matching method according to an embodiment of the present invention; Figure 2 This is a comparison chart of the average correct matching rate (CMR) before and after MLGC strategy optimization in this embodiment of the invention; Figure 3 is a comparison of the running time of the embodiments of the present invention and the traditional pairwise matching method. Figure 3(a) Temple; Figure 3(b) Palace; Figure 3(c) Lighthouse; Figure 3(d) Playground. Figure 4 is a comparison of the number of feature trajectories between the embodiment of the present invention and the traditional pairwise matching method. Figure 4(a) Temple; Figure 4(b) Palace; Figure 4(c) Lighthouse; Figure 4(d) Playground; Figure 5 This is a comparison image of sparse point clouds reconstructed by embodiments of the present invention and traditional pairwise matching methods. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The technical features involved in the various embodiments of the invention described below can be combined with each other as long as they do not conflict with each other.

[0024] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application. This invention provides a high-dimensional feature matching method for multi-view images. It achieves efficient aggregation of high-dimensional features through a multi-view feature adaptive hashing method, and combines a multi-view local geometric consistency strategy to achieve lightweighting and gross error removal. This solves the technical problems of low efficiency, difficulty in high-dimensional feature aggregation, and poor gross error removal in existing multi-view matching methods. The invention will be further described below with reference to embodiments.

[0025] like Figure 1 As shown, the high-dimensional feature matching method for multi-view images provided in this embodiment includes three core steps: feature extraction, simultaneous hash matching, and geometric consistency gross error removal. The specific implementation process is as follows: S1: Multi-view image feature extraction. This embodiment selects four scenes from the Tanks and Temples benchmark dataset as the test dataset. The image resolution for each scene is 1920×1080, and the number of images is 302, 509, 309, and 307 respectively. The SuperPoint algorithm is used to extract the high-dimensional feature descriptor and feature point coordinates of each image, thereby constructing the feature matrix and feature point matrix for each image. The first... Feature matrix of an image and feature point matrix ,as follows: , , in, The dimension of the high-dimensional feature descriptor, This represents the number of feature points in a single image. They are the 1st, 2nd and 3rd respectively. N A high-dimensional feature descriptor for each feature point, meaning that each feature point corresponds to a high-dimensional feature descriptor. They are the 1st, 2nd, and 3rd respectively. N The coordinates of each feature point. It should be understood that the techniques for extracting high-dimensional feature descriptors and feature point coordinates are not limited to the SuperPoint algorithm; other image feature extraction techniques are also feasible.

[0026] S2: Based on the Multi-View Feature Adaptive Hash (MFAH) method, high-dimensional features are binary encoded and the initial multi-view feature trajectory is generated through fast bitwise operations in the hash space, as detailed below: S2.1: Initialize hash matching phase parameters and generate a one-dimensional Gaussian distribution. random projected hash matrix In this embodiment, a multi-layer hash bit-width sequence is preferably set. In other feasible embodiments, it can be set to a single layer. Hamming distance threshold. Hamming radius ; They are the 1st, 2nd, and 3rd respectively. k 1 random projection vector; k The vector dimension is generally the same as the feature dimension. For the first Layer hash bit width, S This represents the number of bit-width sequence layers.

[0027] S2.2: Binary encoding for each feature descriptor Through random projection hash matrix By performing binary encoding, each feature descriptor obtains a set of hash codes with multiple bit widths, thus mapping high-dimensional features to a low-dimensional hash space.

[0028] Among them, based on feature descriptors With random projected hash matrix The corresponding binary encoding calculation formula is:

[0029] In the formula, For random projection hash matrix random vectors in For the first The first image A high-dimensional feature descriptor, For a single feature descriptor The corresponding binary hash code is defined as follows: ; The hash bit width of a certain layer, These are the 1st, 2nd, and 3rd bits of the hash code width in the first layer, respectively. L A binary code consisting of either 0 or 1.

[0030] It should be understood that the set of hash codes with multiple bit widths is as follows: .

[0031] S2.3: Hash bucket construction: Calculate the distance between feature points of different images based on hash codes, aggregate feature points with a distance less than a preset threshold into the same hash bucket, and ensure that the number of feature points of a single image in each hash bucket is no more than 1, so as to avoid multiple feature points of the same image from matching the same physical point. This embodiment calculates the Hamming distance between different image features and aggregates feature points with a Hamming distance less than a threshold into the same hash bucket. The corresponding formula for calculating the Hamming distance is as follows:

[0032] in, For the first The first image The hash code of each feature, For the first The first image The hash code of each feature, For the first The value of the bit hash code.

[0033] S2.4: Hash bucket aggregation and trajectory generation. The adjacent hash buckets of each candidate hash bucket are retrieved and aggregated with a preset radius. The feature points in the aggregated candidate hash buckets are then sequentially associated into a set according to the image view number, forming a multi-view feature trajectory with spatial physical points as the core.

[0034] In this embodiment, the Hamming radius is used. Retrieve and aggregate the adjacent hash buckets of each candidate hash bucket, and then associate the feature points within the candidate hash buckets into sets in an ordered manner according to the image view number (it should be understood that in rare cases, after aggregation, the candidate hash bucket may contain multiple feature points in the same view image, thus generating multiple trajectories), thereby generating the initial multi-view feature trajectory. ,in These are the 1st, 2nd, and Mth feature trajectories, and each trajectory This represents the correspondence between feature points of the same physical point in space across different images, where... The first The trajectory corresponds to the first, second, and third images. Feature points of an image; if a certain image If the physical point does not exist, then the corresponding .

[0035] It should be noted that this embodiment sets up a multi-layer hash bit-width sequence, and therefore adopts a progressive trajectory generation strategy, specifically as follows: For the first-layer hash bit-width sequence, multi-view feature trajectories are generated through steps S2.2-S2.4; For the feature points that were not aggregated in step S2.3, the next layer of trajectory generation is initiated. That is, using the second layer hash bit width sequence, multi-view feature trajectories are generated for the feature points that were not aggregated through steps S2.2-S2.4. By iterating sequentially, the final multi-view feature trajectory is obtained.

[0036] S3: Based on the Multi-View Local Geometric Consistency (MLGC) strategy, the initial multi-view feature trajectories are simultaneously removed to retain feature trajectories that satisfy local geometric consistency, thus obtaining reliable multi-view feature matching results.

[0037] S3.1: Initialize MLGC parameters. In this embodiment, the first-stage coarse matching parameter is set to the number of nearest neighbors. Topological similarity threshold Regularization parameters For each feature point Constructing local neighborhoods in the corresponding images The neighborhood is represented as the local topological relationship between the feature point and its eight neighbors. It should be understood that other feasible embodiments are not limited to the parameter values ​​described above, and the parameter values ​​can be adaptively adjusted based on application accuracy requirements.

[0038] S3.2: Geometric consistency loss calculation. The displacement vectors from feature points to their neighboring points in the neighborhood are calculated sequentially. The normalized topological similarity under multiple views is further calculated, and the geometric consistency loss is quantified based on the comparison between the topological similarity threshold and the normalized topological similarity.

[0039] In this embodiment, feature points are calculated sequentially to their neighborhoods. Displacement vectors of the inner 8 neighboring points Solving for normalized topological similarity under multiple views ,according to Quantization geometric consistency loss The loss is 0 when the similarity is greater than 0.6, and 1 otherwise; this calculation process is repeated for all 8 neighbors in the neighborhood.

[0040] The formula for calculating normalized topological similarity is:

[0041] in, For the first The first image The feature point to its i.e. The displacement vectors of the nearest neighbor points It is the first The first feature point and the second The nearest neighbor points correspond to the displacement vectors in the u-th image; The total number of images, Let be the magnitude of the displacement vector. The cosine similarity of the displacement vectors. ; The quantitative formula for geometric consistency loss is:

[0042] in, for The corresponding geometric consistency loss value, This is the topological similarity threshold.

[0043] S3.3: Initial interior point extraction, constructing a cost function based on multi-view local geometric consistency, and solving for the binary decision value of each feature point. Such as geometric constraint cost If the condition is met, retain the relevant points; otherwise, discard them to obtain the initial set of interior point feature trajectories. ; Cost function of local geometric consistency in multiple views for:

[0044] Solving the cost function yields the value of each feature point in the trajectory set. Based on the binary determination value It can determine whether a certain feature point in the trajectory set is retained. Among them, This represents the total number of current multi-view feature trajectories. The total number of images, For the first The first of the trajectories Binary determination values ​​of feature points in an image. For regularization parameters, For the first The trajectory in the first Geometric constraint cost of feature points in an image; This is the current set of multi-view feature trajectories.

[0045] The rule for solving the geometric constraint cost is as follows:

[0046] in, K For the neighborhood The number of nearest neighbors in the data. The total number of images, For the first Located on the trajectory of the first Feature points of an image for The Nearest neighbor points, They are respectively and Corresponding to the Matching points of the images; This represents the Euclidean distance between two points. Representing feature points In the Similarly, the neighborhood under the image, Representing feature points In the The neighborhood under the image.

[0047] Indicates the first Corresponding feature points on the trajectory The geometric constraint cost is used to measure the degree to which the matching result satisfies the geometric constraints. The smaller the value, the more accurate the matching result and the smaller its impact on the objective function; A larger value indicates an error in the matching result. For multi-view matching tasks, each feature trajectory can define a set of multi-view neighborhoods, thus enabling the calculation of the corresponding [value] for each matching point. value.

[0048] The rule for solving the binary determination value of feature points is as follows:

[0049] When the binary determination value of a feature point is 1, the feature point is retained in the trajectory; when the binary determination value of a feature point is 0, the feature point is removed, thus updating the trajectory.

[0050] S3.4: Adjust the second-stage fine-matching parameters as follows , , Based on the updated trajectory Reconstructing the local neighborhood It should be understood that after the trajectory is updated, the corresponding neighborhood also changes dynamically because some feature points have been removed. The parameters in this stage are more stringent than those in the first stage, and can generally be appropriately increased. And decrease The number of neighboring points can be appropriately increased, but generally the same number as in the first stage is used. The final high-precision multi-view feature matching trajectory is obtained by iteratively calculating the cost function and binary decision values. The number of iterations and the iteration termination condition can be adaptively set and adjusted according to application requirements.

[0051] This embodiment verifies the matching effect of the above method. The method of this invention is compared with the traditional pairwise matching method SuperPoint+SuperGlue (SP_SG) under the same hardware environment: Intel Xeon E5-2680 v3 processor, NVIDIA GeForce RTX 2080 Ti graphics card, 188GiB memory, and the same SfM backend parameters. Running time, correct matching rate (CMR), root mean square error (RMSE), reprojection error, and camera pose error are used as evaluation indicators. The verification results are as follows: Matching efficiency: As shown in Figure 3, the initial trajectory generated by this invention is slightly less than that of the traditional pairwise matching method. However, the running time of the traditional pairwise matching method increases exponentially, while the running time of the method of this invention increases linearly with the number of images, as shown in Figure 4. Especially in the scenario of 200 images, the running time of the method of this invention is 6-10 times faster than that of pairwise matching.

[0052] Gross error removal effect: The MLGC strategy reduces the matching RMSE of each scene from 78~279 pixels to 1~2 pixels, and improves the CMR from 40%~44% to 92%~93%, effectively removing gross errors caused by local geometric inconsistencies. The continuity and consistency of feature trajectories are significantly improved, such as... Figure 2 As shown; Table 1 Point cloud reprojection error

[0053] 3D Reconstruction Quality: The number of feature trajectories generated by the method of this invention is 60%~80% of that of traditional methods. However, for different datasets, the reprojection error of the reconstructed sparse point cloud is comparable to that of traditional methods, both being 0.7~0.8 pixels, as shown in Table 1. The translation error, rotation error, and scale ratio of the camera pose are similar, as shown in Table 2, ensuring the accuracy of subsequent 3D reconstruction. Figure 5 Both methods shown can achieve good sparse point cloud reconstruction results.

[0054] Table 2 Comparison of camera poses (200 images)

[0055] In some feasible embodiments, the present invention also provides a matching system based on a high-dimensional feature matching method for multi-view images, which includes at least: a feature extraction module, a hash matching module, and a geometric optimization module.

[0056] The feature extraction module is used to acquire a multi-view image set, extract high-dimensional feature descriptors and feature point coordinates of each image, and construct the feature matrix and feature point matrix of each image. The hash matching module is used to execute the above-mentioned multi-view feature adaptive hashing method, which performs binary encoding, hash bucket construction and fusion on high-dimensional feature descriptors, realizes simultaneous aggregation of multi-view features, and generates initial multi-view feature trajectories. The geometry optimization module is used to remove gross errors from the initial multi-view feature trajectories, resulting in optimized multi-view feature trajectories. Preferably, the aforementioned multi-view local geometric consistency strategy is introduced to simultaneously remove gross errors from the initial multi-view feature trajectories, outputting high-precision multi-view feature matching results.

[0057] In some feasible embodiments, the hash matching module further includes a hash parameter initialization unit, a binary encoding unit, a hash bucket construction unit, and a trajectory generation unit. Specifically, the hash parameter initialization unit generates a random projection hash matrix and sets parameters such as hash bit width, Hamming distance threshold, and Hamming radius; the binary encoding unit performs binary encoding on high-dimensional feature descriptors, mapping them to a low-dimensional hash space; the hash bucket construction unit calculates the Hamming distance of feature hash codes and aggregates similar feature points to form candidate hash buckets; and the trajectory generation unit retrieves and merges adjacent hash buckets, performs multi-level hash bucket trajectory fusion, and generates an initial multi-view feature trajectory.

[0058] In some feasible embodiments, the geometry optimization module further includes a geometry parameter initialization unit, a neighborhood construction unit, a loss calculation unit, and a trajectory optimization unit. Specifically, the geometry parameter initialization unit sets MLGC parameters such as the number of nearest neighbors, topological similarity threshold, and regularization parameters; the neighborhood construction unit constructs a K-nearest neighbor local neighborhood for each feature point under multiple views; the loss calculation unit calculates the normalized topological similarity of the feature point's neighborhood displacement vector, quantifying the geometry consistency loss; and the trajectory optimization unit constructs a cost function based on geometry consistency, solves for the binary judgment value of the feature point through a two-stage threshold strategy, and achieves gross error removal and fine optimization of the initial trajectory.

[0059] It should be understood that the simultaneous matching system provided by this invention corresponds to the simultaneous matching method described above. Therefore, for specific details, please refer to the description of the method above; this invention will not repeat them here. Furthermore, the above division of functional modules is merely a logical functional division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Simultaneously, the integrated units described above can be implemented in hardware or as software functional units.

[0060] Furthermore, the present invention also provides an electronic device, including one or more processors and a memory storing one or more computer programs, wherein the memory stores the computer programs, and the processor calls the computer programs to execute: S1: extracting high-dimensional feature descriptors and feature point coordinates of multi-view images to construct a feature dataset; S2: encoding high-dimensional features into binary based on the Multi-View Feature Adaptive Hash (MFAH) method, and generating initial multi-view feature trajectories through fast bit operations in the hash space; S3: performing gross error removal on the initial multi-view feature trajectories to obtain optimized multi-view feature trajectories, wherein preferably, a Multi-View Local Geometric Consistency (MLGC) strategy is introduced to perform gross error removal on the initial trajectories to obtain an optimized set of multi-view feature trajectories.

[0061] For details on the implementation process, please refer to the description of the method embodiment.

[0062] The present invention also provides a readable storage medium storing a computer program, which is called by a processor to execute: S1: Extracting high-dimensional feature descriptors and feature point coordinates from multi-view images to construct a feature dataset; S2: Encoding high-dimensional features in binary form based on the Multi-View Feature Adaptive Hash (MFAH) method, and generating initial multi-view feature trajectories through fast bit operations in the hash space; S3: Performing gross error removal on the initial multi-view feature trajectories to obtain optimized multi-view feature trajectories, wherein preferably, a Multi-View Local Geometric Consistency (MLGC) strategy is introduced to perform gross error removal on the initial trajectories to obtain an optimized set of multi-view feature trajectories.

[0063] The specific implementation process can be referred to the description of the foregoing method. It should be understood that in this embodiment of the invention, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0064] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the controller described in any of the foregoing embodiments, such as the controller's hard drive or memory. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both the controller's internal storage unit and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0065] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1. A high-dimensional feature matching method for multi-view images, characterized in that, Includes the following steps: S1: Feature construction, extracting high-dimensional feature descriptors and feature point coordinates from multi-view images to construct a feature dataset; S2: Trajectory generation. The high-dimensional feature descriptor is encoded based on the Multi-View Feature Adaptive Hash (MFAH) method. Then, the initial multi-view feature trajectory is generated through fast bit operations in the hash space. Each multi-view feature trajectory represents the correspondence between feature points of the same physical point in space in various images. S3: Trajectory correction, which removes gross errors from the initial multi-view feature trajectory to obtain the optimized multi-view feature trajectory; The specific process of trajectory generation in step S2 includes: S21: Initialize hash parameters and generate a one-dimensional Gaussian distribution. random projected hash matrix ; S22: By randomly projecting the hash matrix Each high-dimensional feature descriptor of each image is binary encoded to obtain a hash code, thereby mapping the high-dimensional features to a low-dimensional hash space; S23: Calculate the distance between feature points of different images based on hash codes, aggregate feature points with a distance less than a preset threshold into the same hash bucket, and ensure that the number of feature points of a single image in each hash bucket is no greater than 1. S24: Generate initial multi-view feature trajectories based on feature points in the hash bucket.

2. The high-dimensional feature matching method for multi-view images according to claim 1, characterized in that, Step S21 also sets a multi-level hash bit-width sequence. , For the first Layer hash bit width, S The number of bit-width sequence layers; Among them, based on the multi-layer hash bit width sequence A progressive trajectory generation strategy is adopted, specifically as follows: For the first-layer hash bit-width sequence, multi-view feature trajectories are generated through steps S22-S24; For the feature points that were not aggregated in step S23, the next layer of trajectory generation is entered, that is, using the second layer hash bit width sequence, multi-view feature trajectories are generated for the feature points that were not aggregated through steps S22-S24. By iterating sequentially, the final multi-view feature trajectory is obtained; The hash code based on binary encoding is: ； In the formula, For random projection hash matrix random vectors in For the first The first image A high-dimensional feature descriptor, For a single feature descriptor The corresponding binary hash code is defined as follows: ; L The hash bit width of a certain layer, These are the 1st, 2nd, and 3rd bits of the hash code width in the first layer, respectively. L Each binary hash code has a value of 0 or 1.

3. The high-dimensional feature matching method for multi-view images according to claim 1, characterized in that, Step S3 introduces the Multi-View Local Geometric Consistency (MLGC) strategy to remove gross errors from the initial multi-view feature trajectories. The implementation process includes at least the following: S31: Initialize MLGC parameters, set the number of nearest neighbors. Topological similarity threshold Regularization parameters And determine the neighborhood of each feature point in the corresponding image; S32: Geometric consistency loss calculation, based on normalized topological similarity of the displacement vectors from feature points to their neighborhoods, and according to a topological similarity threshold. Quantize the geometric consistency loss of the displacement vector; S33: Construct a cost function based on geometric consistency loss, determine whether feature points are retained by minimizing the cost function, obtain the initial set of interior point feature trajectories, and realize trajectory update.

4. The high-dimensional feature matching method for multi-view images according to claim 3, characterized in that, The formula for calculating normalized topological similarity in step S32 is: ； in, To normalize topological similarity, For the first The first image The feature point is located in its neighborhood. The displacement vectors of the nearest neighbor points It is the first The first feature point and the second The nearest neighbor points correspond to the displacement vectors in the u-th image; The total number of images, Let be the magnitude of the displacement vector. The cosine similarity of the displacement vectors. .

5. The high-dimensional feature matching method for multi-view images according to claim 3, characterized in that, The cost function mentioned in step S33 is: ； in, Let cost function be This represents the total number of current multi-view feature trajectories. The total number of images, For the first The first of the trajectories Binary determination values ​​of feature points in an image. For the first The trajectory in the first Geometric constraint cost of feature points in an image; For the current set of multi-view feature trajectories, Represents feature points.

6. The high-dimensional feature matching method for multi-view images according to claim 5, characterized in that, Geometric constraint cost satisfy: ； in, K For the neighborhood The number of nearest neighbors in the data. For the first Located on the trajectory of the first Feature points of an image for The Nearest neighbor points, They are respectively and Corresponding to the Matching points of the images; This represents the Euclidean distance between two points. For feature points In the Neighborhood under the image, For feature points In the The neighborhood under the image; The rule for solving the binary determination value of feature points is as follows: ； When it is 1, the first one is reserved. Corresponding feature points on the trajectory; Remove the first when the value is 0. The corresponding feature points on the trajectory.

7. A matching system based on the high-dimensional feature matching method for multi-view images according to any one of claims 1-6, characterized in that: include: Feature extraction module: used to extract high-dimensional feature descriptors and feature point coordinates from multi-view images and construct a feature dataset; Hash matching module: used to encode high-dimensional feature descriptors based on the Multi-View Feature Adaptive Hash (MFAH) method, and then generate initial multi-view feature trajectories through fast bit operations in the hash space. Each multi-view feature trajectory represents the correspondence between feature points of the same physical point in space in various images. The geometry optimization module is used to remove gross errors from the initial multi-view feature trajectory to obtain the optimized multi-view feature trajectory. The hash matching module is signal-connected to the geometry optimization module, and the output of the feature extraction module is signal-connected to the input of the hash matching module.

8. An electronic device, characterized in that: It includes one or more processors and a memory storing one or more computer programs, wherein when the processor invokes the computer programs, it performs the steps of the high-dimensional feature matching method for multi-view images according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when invoked by a processor, executes the steps of the high-dimensional feature matching method for multi-view images according to any one of claims 1-6.

Images