Efficient panoramic image splicing method and system based on multi-view fusion

Abstract

The invention relates to an image processing technology, and discloses an efficient panoramic image splicing method and system based on multi-view fusion, and the method comprises the steps: collecting a plurality of images from different views; performing multi-scale feature extraction on each image, generating a feature descriptor for each extracted feature point, and matching a corresponding feature point pair; performing multi-view geometric constraint screening on the feature point pairs; according to the screened feature point pairs, estimating a homography matrix between adjacent images, and carrying out global optimization on the homography matrix; aligning all the images into the same coordinate system; determining an overlapping region between adjacent images; and according to the pixel information in the overlapping areas, fusing the overlapping areas by adopting a self-adaptive weighted fusion algorithm so as to splice the plurality of images into a panoramic image. The invention further discloses a control device and a computer readable storage medium. The invention aims to improve the efficiency and accuracy of generating the multi-view fused panoramic image.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to an efficient panoramic image stitching method based on multi-perspective fusion, a control device, an equipment system, and a computer-readable storage medium. Background Art

[0002] Panoramic images offer a wide field of view and are widely used in fields such as VR, security, and surveying and mapping. While the widespread availability of image acquisition devices has made it easier to acquire multi-viewpoint images, efficiently and accurately stitching them together into a panoramic view remains a challenge.

[0003] Traditional methods perform poorly in the feature extraction and matching stages, are slow when processing large numbers of images, and cannot meet real-time requirements. Furthermore, the lack of effective geometric constraints results in a large number of erroneous matching point pairs in the matching results, increasing the difficulty and error of the subsequent homography matrix estimation, requiring multiple iterations for correction, further reducing efficiency. Furthermore, the complex geometric relationships between images from different viewpoints easily lead to a large number of erroneous matching points, which affect homography matrix estimation, resulting in misaligned stitching and ghosting, and in turn causing local distortion and deformation of the panoramic image.

[0004] In the image fusion stage, some methods adopt a fixed fusion strategy without considering the differences in pixel information in overlapping areas, resulting in poor fusion effect. It may be necessary to repeatedly adjust parameters for optimization, wasting a lot of time and computing resources.

[0005] The above content is only used to assist in understanding the technical solution of this application and does not constitute an admission that the above content is prior art. Summary of the Invention

[0006] The main purpose of this application is to provide an efficient panoramic image stitching method based on multi-perspective fusion, a control device, an equipment system and a computer-readable storage medium, aiming to improve the efficiency and accuracy of generating multi-perspective fusion panoramic images.

[0007] To achieve the above objectives, the present application provides an efficient panoramic image stitching method based on multi-view fusion, comprising the following steps: Control multiple cameras to capture multiple images from different perspectives and perform image preprocessing; Perform multi-scale feature extraction on each image and generate a feature descriptor for each extracted feature point; Based on the fast matching algorithm, the nearest neighbor search is performed on the feature descriptors between different images to match the corresponding feature point pairs; Based on the constraint model, multi-view geometric constraint screening is performed on the feature point pairs; wherein the constraint model is pre-constructed using the relative position relationship between adjacent images and the projection geometry principle; Estimate the homography matrix between adjacent images based on the filtered feature point pairs; The homography matrix is globally optimized using the global consistency optimization algorithm; Perform perspective transformation on the image based on the optimized homography matrix to align all images to the same coordinate system; Determining an overlapping area between adjacent images and analyzing pixel information within the overlapping area, wherein the pixel information includes at least pixel positions; According to the pixel information in the overlapping area, an adaptive weighted fusion algorithm is used to fuse the overlapping area to stitch multiple images into a panoramic image.

[0008] To achieve the above objectives, the present application further provides a control device, comprising: The acquisition module is used to control multiple cameras to acquire multiple images from different perspectives and perform image preprocessing; The extraction module is used to extract multi-scale features from each image and generate a feature descriptor for each extracted feature point; The matching module is used to perform nearest neighbor search on feature descriptors between different images based on a fast matching algorithm to match corresponding feature point pairs; A constraint module, configured to perform multi-view geometric constraint screening on feature point pairs based on a constraint model, wherein the constraint model is pre-constructed using the relative positional relationship between adjacent images and the principle of projection geometry; An estimation module is used to estimate the homography matrix between adjacent images based on the filtered feature point pairs; An optimization module, used to perform global optimization on the homography matrix using a global consistency optimization algorithm; The alignment module is used to perform perspective transformation on the image based on the optimized homography matrix and align all images to the same coordinate system; An analysis module, configured to determine an overlapping region between adjacent images and analyze pixel information within the overlapping region, wherein the pixel information includes at least pixel positions; The stitching module is used to fuse the overlapping areas according to the pixel information in the overlapping areas using an adaptive weighted fusion algorithm to stitch multiple images into a panoramic image.

[0009] To achieve the above-mentioned objectives, the present application also provides a device system, which includes: a memory, a processor, and a computer program stored on the memory and capable of running on the processor. When the computer program is executed by the processor, the steps of the above-mentioned efficient panoramic image stitching method based on multi-perspective fusion are implemented.

[0010] To achieve the above objectives, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned efficient panoramic image stitching method based on multi-perspective fusion are implemented.

[0011] The present application provides an efficient panoramic image stitching method, control device, equipment system and computer-readable storage medium based on multi-perspective fusion. The efficient panoramic image stitching method based on multi-perspective fusion can not only comprehensively capture image features, but also greatly improve matching efficiency, while reducing mismatching, reducing the difficulty and error of homography matrix estimation, and using a global consistency optimization algorithm to globally optimize the homography matrix, avoiding multiple iterative corrections, reducing local error accumulation, and improving the overall accuracy of stitching. At the same time, it can improve the overall efficiency of panoramic image stitching; and an adaptive weighted fusion algorithm is used in the image fusion stage, which can dynamically adjust the fusion weights without repeatedly adjusting parameters, and can obtain better fusion effects, further improving the accuracy of panoramic image stitching. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a schematic diagram of the steps of an efficient panoramic image stitching method based on multi-view fusion in one embodiment of the present application; Figure 2 This is a schematic diagram of a control device in an embodiment of the present application; Figure 3 Schematic diagram of the internal architecture of a device system according to an embodiment of the present application.

[0013] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0014] The embodiments of the present application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be understood as limiting the present application. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application without making any creative efforts shall fall within the scope of protection of the present application.

[0015] In addition, any descriptions of "first," "second," etc., in this application are for descriptive purposes only (e.g., to distinguish identical or similar features) and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include at least one such feature. Furthermore, the technical solutions of various embodiments may be combined with each other, but this must be based on the ability of a person of ordinary skill in the art to implement them. If the combination of technical solutions contradicts or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0016] Reference Figure 1 In one embodiment, an efficient panoramic image stitching method based on multi-view fusion includes: Step S10: controlling multiple cameras to capture multiple images from different viewing angles and performing image preprocessing; Step S20: performing multi-scale feature extraction on each image and generating a feature descriptor for each extracted feature point; Step S30: performing nearest neighbor search on the feature descriptors between different images based on a fast matching algorithm to match corresponding feature point pairs; Step S40: Perform multi-view geometric constraint screening on the feature point pairs based on the constraint model; wherein the constraint model is pre-constructed using the relative position relationship between adjacent images and the projection geometry principle; Step S50: estimating the homography matrix between adjacent images based on the filtered feature point pairs; Step S60: using a global consistency optimization algorithm to globally optimize the homography matrix; Step S70: performing perspective transformation on the images based on the optimized homography matrix to align all images into the same coordinate system; Step S80: determining an overlapping area between adjacent images, and analyzing pixel information within the overlapping area, wherein the pixel information includes at least pixel positions; Step S90: Adopting an adaptive weighted fusion algorithm to fuse the overlapping areas according to the pixel information in the overlapping areas, so as to stitch the multiple images into a panoramic image.

[0017] In this embodiment, the execution terminal of the embodiment may be a device system, or may be other equipment or devices (such as a control device) that controls the device system.

[0018] As described in step S10, in order to capture comprehensive images with overlapping areas from different perspectives, it is necessary to rationally arrange multiple cameras based on the actual scene and stitching requirements. For example, multiple cameras can be evenly distributed on a circle so that their shooting range covers the entire indoor space, and the shooting areas of adjacent cameras overlap to a certain extent. An overlap rate of 20% to 50% is generally appropriate to ensure accurate image stitching.

[0019] Use a hardware control device or software program to synchronize the shooting operations of multiple cameras. The hardware control device can be a specially designed camera controller that connects to each camera via a cable and sends synchronized shooting commands. The software program can use the communication interface between the computer and the camera (such as USB or network interface) to achieve remote control of the camera.

[0020] Choose the appropriate image resolution based on the application requirements and subsequent processing capabilities of the panoramic image. If high-definition panoramic images are required for large-screen display or high-precision analysis, choose a higher-resolution camera, such as 4K or even 8K resolution.

[0021] Optionally, to improve image quality and reduce noise, you can capture multiple frames for each viewpoint. These frames can then be merged into a single frame using an image fusion algorithm (such as average fusion or median fusion) to improve the signal-to-noise ratio and image clarity. For example, in scenes with unstable lighting conditions, capturing 3-5 frames and then fusing them can effectively reduce random noise in the image.

[0022] Optionally, due to the characteristics of the camera sensor and interference from the external environment, the captured image may contain noise, such as Gaussian noise and salt and pepper noise. Use a denoising algorithm to process the image to improve image quality. Common denoising algorithms include Gaussian filtering and median filtering. Gaussian filtering is suitable for removing Gaussian noise by smoothing the image by taking a weighted average of each pixel in the image and its neighborhood. Median filtering is more effective at removing salt and pepper noise by sorting the pixel values in the neighborhood and taking the median value as the current pixel value.

[0023] Optionally, image enhancement algorithms can be used to enhance image contrast and brightness, making details clearer. For example, histogram equalization is a commonly used image enhancement method. It adjusts the image's grayscale histogram to make the grayscale distribution more uniform, thereby improving image contrast. Alternatively, the color-adaptive histogram equalization (CLAHE) algorithm can be used, which enhances image contrast while avoiding image distortion caused by over-enhancement.

[0024] Alternatively, by using a calibration plate of known size and shape (such as a checkerboard calibration plate), multiple images can be captured at different positions and angles. These images can be used to calculate the camera's intrinsic parameters (such as focal length, principal point coordinates, etc.) and distortion coefficients. Distortion correction can then be performed on the captured images based on these parameters.

[0025] As described in step S20, the preprocessed image is first convolved with a series of Gaussian kernels with different standard deviations (σ). The standard deviation σ determines the degree of smoothing. A smaller σ corresponds to a smaller scale, resulting in less smoothed images and greater preservation of detail. A larger σ corresponds to a larger scale, resulting in a smoother image and reduced detail, but allowing for the detection of features over a wider range. The images convolved with the different Gaussian kernels are arranged in ascending order of scale to form a Gaussian pyramid. Each level of the pyramid contains multiple images of different scales.

[0026] Based on the Gaussian pyramid, the Gaussian smoothed images of two adjacent layers are subtracted to obtain a Gaussian difference image. The DoG image highlights the areas in the image that change significantly at different scales. These areas often correspond to feature points in the image, such as corners and edges.

[0027] Find local extreme points in the DoG scale space. These extreme points are possible feature points. Specifically, for each pixel, it is compared with the 8 adjacent pixels at the same scale and the 18 pixels at the upper and lower scales. If the value of the pixel is greater or less than the values of these adjacent pixels, the pixel is considered a local extreme point.

[0028] The detected local extreme points may only be approximate feature point locations and require further precise positioning. A three-dimensional quadratic function fit is performed on the local DoG function to determine the exact location and scale of the feature points. At the same time, some low-contrast points and points located at the edges are removed to improve the quality of the feature points. Low-contrast points may be caused by noise or minor image changes and do not contribute much to feature matching. Points located at the edges are less stable across different viewing angles and may cause matching errors.

[0029] A neighborhood of a specific size is defined around the detected feature point. The size of this neighborhood is typically related to the scale of the feature point to ensure that appropriate local information is captured at different scales. The gradient magnitude and direction are calculated for each pixel within the neighborhood. The gradient magnitude indicates the degree of change in the pixel within the local area, while the gradient direction reflects the direction of change.

[0030] A dominant direction is assigned to a feature point to achieve rotation invariance. Specifically, a histogram of the gradient directions of pixels within a neighborhood is calculated. The direction corresponding to the peak of the histogram is the dominant direction of the feature point. If there are multiple peaks close to each other, multiple dominant directions can be assigned to the feature point.

[0031] Divide the neighborhood into multiple subregions, such as 4×4 subregions. Within each subregion, calculate the gradient direction histogram of the pixels. Each histogram typically contains eight direction bins. Combining the histograms of all subregions forms a 128-dimensional vector, which is the SIFT feature descriptor. Normalizing this vector improves the descriptor's illumination invariance.

[0032] As described in step S30, the feature descriptors of different images extracted in step S20 are stored as feature descriptor sets. For example, if there are images A and B, their feature descriptor sets D are obtained respectively. A and D B , where each feature descriptor is a vector representing the feature information of the corresponding feature point.

[0033] Alternatively, the fast matching algorithm uses a KD-Tree (K-Dimensional Tree). A KD-Tree is a binary tree in which each node represents a hyperrectangular region in k-dimensional space. The space is continuously divided into two subregions, so that data points in each subregion are assigned to the corresponding subtree. The division is based on selecting a coordinate axis and dividing the data points into two parts according to the median value of the data points on that axis.

[0034] Optionally, select a coordinate axis and loop through the selection process, selecting the first dimension, the second dimension, and so on up to the kth dimension. Calculate the median of all data points along that axis, split the data points into two parts, and use the point corresponding to the median as the current node. Recursively construct subtrees for the left and right parts of the data points until only one data point remains in the subtree or a pre-defined termination condition is met.

[0035] Optionally, the goal of the nearest neighbor search is to find, in one set of feature descriptors, a feature descriptor that is closest to a feature descriptor in another set, where the distance metric is Euclidean distance.

[0036] Starting from the root node of the KD-Tree, the search point is selected to enter the left or right subtree, depending on the value of the current search point on the partitioning axis, until a leaf node is reached. The feature descriptor corresponding to the leaf node is used as the current nearest neighbor candidate. Backtracking to the parent node, the distance between the parent node's partitioning plane and the current search point is checked. If the distance is less than the distance to the current nearest neighbor, the search is continued in another subtree for a potentially closer point. This backtracking process is repeated until all subtrees that may contain closer points have been traversed, ultimately resulting in the nearest neighbor.

[0037] Optionally, a one-to-one matching strategy can be used. For example, for each feature descriptor in image A, the nearest neighbor feature descriptor in the set of feature descriptors in image B is found as the matching point. To ensure matching accuracy, a reverse matching strategy can also be performed: for each feature descriptor in image B, the nearest neighbor feature descriptor in image A is found. Only when both bidirectional matches are successful is the pair considered a valid matching point.

[0038] In addition to finding the nearest neighbor, the next nearest neighbor is also found. The ratio of the distance between the nearest neighbor and the next nearest neighbor is calculated. If this ratio is less than a preset threshold (such as 0.8), the match is considered reliable. This is because if the distance difference between the nearest neighbor and the next nearest neighbor is large, it means that the feature point has a clear corresponding point in the other image; conversely, if the distance difference is small, there may be a match error.

[0039] As described in step S40 , the constraint model is constructed in advance using the relative positional relationship between adjacent images and projection geometry principles.

[0040] In real-world scenarios, multiple cameras capture images from different perspectives, and images captured by adjacent cameras are spatially correlated. For example, in a ring-shaped camera array, the perspectives of adjacent cameras change continuously, which means that the content of adjacent images also has a continuous transition relationship. This relative positional relationship provides spatial prior information for the constraint model.

[0041] Projective geometry describes the mathematical relationship between objects projected from three-dimensional space onto a two-dimensional image plane. In image stitching, theorems and formulas in projective geometry, such as collinearity equations and perspective transformations, can be used to establish correspondences between feature points in adjacent images. For example, based on the pinhole camera model, points in three-dimensional space are mapped onto a two-dimensional image plane through a projective transformation. Using known camera parameters and projective relationships, geometric constraints between corresponding feature points in adjacent images can be derived.

[0042] In step S30, a fast matching algorithm is used to perform a nearest neighbor search on the feature descriptors between different images, and the corresponding feature point pairs are matched. However, there may be mismatches among these feature point pairs. Mismatches may be caused by factors such as image noise, illumination changes, and limitations of feature descriptors. The purpose of multi-view geometric constraint screening is to remove these mismatched feature point pairs and retain only the correct matching point pairs that meet the multi-view geometric constraints, thereby improving the accuracy of feature point matching and providing more reliable data for subsequent homography matrix estimation.

[0043] Optionally, based on the constraint model, calculate whether each feature point pair satisfies the multi-view geometric constraints. Optional geometric constraints include: (1) Epipolar constraint: In a binocular vision system, the epipolar constraint describes the relationship between the projection points of the same 3D point on two image planes. For a pair of feature points in a pair of adjacent images, if they are correctly matched, then they should satisfy the epipolar constraint. By calculating the epipolar line of the feature point pair, if the geometric relationship between the position of the feature point pair and the epipolar line does not meet the epipolar constraint, the feature point pair is considered to be mismatched.

[0044] (2) Scale consistency constraint: Feature points of the same object in different images should have a certain consistency in scale. For example, if the scale of the object corresponding to a feature point in one image is known, then the scale of the object corresponding to the matching feature point in the adjacent image should be within a reasonable range. By comparing the scale information of the feature points, feature point pairs that do not meet the scale consistency constraint can be filtered out.

[0045] (3) Screening feature point pairs: For each feature point pair, substitute it into the constraint model for verification. If the geometric constraints are met, the feature point pair is retained; if not, it is removed. For example, an error threshold for the epipolar constraint is set. When the epipolar constraint error of the feature point pair is less than the threshold, the feature point pair is considered to meet the constraint conditions. Otherwise, it is removed as a mismatched point pair.

[0046] After multi-view geometric constraint screening, the resulting feature point pairs are more accurate and reliable. These filtered feature point pairs are used to estimate the homography matrix between adjacent images in step S50. Accurate feature point pairs can improve the accuracy of homography matrix estimation, making subsequent image alignment and stitching more natural and accurate, and reducing problems such as misalignment and ghosting in the stitched image.

[0047] As described in step S50 , in the field of computer vision, a homography matrix is a 3×3 matrix, typically denoted by H. It describes the perspective transformation relationship between two planes. In image stitching, adjacent images are captured from different perspectives of the same scene, and a projective transformation relationship exists between them, which can be represented by a homography matrix. Using the homography matrix, points in one image can be mapped to corresponding locations in another image, thereby achieving image alignment.

[0048] Using the selected feature point pairs, the homography matrix between adjacent images is estimated using methods such as the least squares method. The least squares method solves the parameters of the homography matrix by minimizing the projection error of the feature point pairs. For example, given multiple matching feature point pairs in adjacent images, the least squares method can be used to find an optimal homography matrix that minimizes the error between these feature point pairs after the projection transformation.

[0049] As described in step S60, the goal of global consistency optimization is to minimize the stitching error between all images during the stitching process of the entire panoramic image by adjusting the homography matrices of all adjacent images, thereby ensuring the overall consistency and accuracy of the panoramic image. In other words, the goal is to ensure that all images blend seamlessly after stitching, without noticeable misalignment or distortion, presenting a natural, coherent panoramic image.

[0050] Optionally, the global consistency optimization algorithm uses a bundle adjustment algorithm to globally optimize the homography matrix.

[0051] The core idea of the bundle adjustment algorithm is to formulate the panoramic image stitching problem as a nonlinear minimization problem. It simultaneously considers the constraints between all images, including the homography matrix between adjacent images and the correspondence between feature point pairs. Through an iterative approach, the parameters of the homography matrix are continuously adjusted to minimize the global error of all stitched images.

[0052] Specifically, the bundle adjustment algorithm defines an error function that measures the projection error of all feature point pairs after the homography transformation. During the optimization process, the algorithm continuously tries different values of the homography matrix parameters to gradually reduce the value of the error function until a certain convergence condition is met.

[0053] After global consistency optimization, the homography matrices of all adjacent images are adjusted, making them more globally consistent. This prevents global inconsistencies caused by the accumulation of local errors during subsequent image stitching, resulting in a high-quality, coherent panoramic image. For example, in the stitched panorama, the overlapping areas between the images blend seamlessly, without noticeable misalignment or distortion, resulting in a more natural and aesthetically pleasing image.

[0054] As described in step S70, a perspective transformation is performed on each image based on the optimized homography matrix, transforming it from its original coordinate system to a unified coordinate system. Perspective transformation can perform operations such as image rotation, translation, and scaling, allowing images from different perspectives to be aligned within the same coordinate system. For example, for two adjacent images, a perspective transformation of one image can be performed using the homography matrix to spatially align it with the other image.

[0055] After the perspective transformation, all images are aligned to the same coordinate system, and the overlapping areas of adjacent images can be accurately aligned. At this point, images from different perspectives can be stitched together in the correct positional relationship, laying the foundation for subsequent fusion operations.

[0056] As described in step S80, the overlapping region between the adjacent images is determined by comparing their positions and ranges. After the images are aligned, the boundaries of the overlapping region can be calculated based on the image coordinate information and the homography matrix. For example, the overlapping region can be found by comparing the coordinates of the four vertices of the two adjacent images.

[0057] Analyze the pixel information within the overlapping area, focusing on its location. This information can be used in subsequent fusion algorithms to determine how to weight the pixels within the overlapping area. Additionally, analysis of pixel color, brightness, and other information can provide additional insights for fusion.

[0058] As described in step S90, for each pixel in the overlapping area, the distance from the pixel to each image boundary is calculated. Generally speaking, the closer a pixel is to an image boundary, the greater the contribution weight of the image to the pixel.

[0059] For example, if pixel p is in the overlapping region, its distance to the boundary of image I1 is d1, and its distance to the boundary of image I2 is d2. Then the distance weight of image I1 for pixel p is w1=d2 / (d1+d2), and the distance weight of image I2 for pixel p is w2=d1 / (d1+d2). In this case, w1+w2=1.

[0060] After calculating the distance weight of each pixel, for each pixel p in the overlapping area, the pixel values of different images at that pixel are weighted averaged according to the weight. Assuming that the pixel value of image I1 at pixel p is P1, and the pixel value of image I2 at pixel p is P2, then the pixel value P after fusion is f For: P f =w1×P1+w2×P2.

[0061] For color images, the above weighted averaging operation needs to be performed on the red, green, and blue channels respectively.

[0062] Alternatively, in addition to distance weighting, the consistency of brightness and color can also be considered. That is, the brightness and color differences at the pixel of the adjacent images in the overlapping area can be calculated. The smaller the difference, the higher the credibility of the image at that pixel, and the greater the weight.

[0063] For example, the brightness difference ΔL and color difference ΔC of the same pixel between two adjacent images can be calculated and then converted into weights according to the Gaussian attenuation formula. Assume that the combined brightness and color consistency weight is w LC , then the weight of the final image I1 at pixel p is W1=w1×w LC1 , the weight of image I2 at pixel p is W2=w2×w LC2 , at this time W1+W2=1, and P f =W1×P1+W2×P2.

[0064] Of course, for color images, it is still necessary to perform the above weighted averaging operation on the red, green, and blue channels respectively.

[0065] Among them, the brightness weight is calculated using the Gaussian attenuation formula: , where σ L is the standard deviation of the Gaussian function of brightness, which controls the rate at which the weight changes with the brightness difference. L The larger the value, the smoother the weight changes with ΔL; L The smaller it is, the more sensitive the weight is to changes in ΔL.

[0066] And, the color weight is calculated using the Gaussian attenuation formula: ,σ C is the standard deviation of the Gaussian function of color, which acts similarly to σ L Similarly, it controls the rate at which the weights change with color differences.

[0067] Comprehensive consistency weight w LC The same can be done by L and w C Perform weighted combination to obtain , μ is the relative importance coefficient of brightness and color, ranging from 0 to 1, and is positively correlated with the importance of brightness. Its specific value can be set according to actual needs, and is generally set to 0.5.

[0068] The core idea of the adaptive weighted fusion algorithm is to assign an appropriate weight to each pixel based on the specific information of the pixels in the overlapping area. The weighted average of the pixel values at that pixel from the different images is then calculated, achieving a smooth transition and avoiding the appearance of stitching seams. This adaptive approach dynamically adjusts the weights based on the actual image conditions (such as brightness and color), resulting in a more natural fusion effect.

[0069] In addition, for non-overlapping areas, the pixel values of the corresponding images are directly copied to the panoramic image. Since these areas are only covered by one image, there is no fusion problem.

[0070] Optionally, to further improve the fusion effect, a multi-resolution fusion method can be used. First, decompose the image into pyramids of different resolutions (such as a Gaussian pyramid), perform a weighted fusion operation on each pyramid layer, and then reconstruct the fused pyramid images to obtain the final fused image. This method can effectively reduce high-frequency noise in the image fusion process, resulting in smoother fusion.

[0071] Through the adaptive weighted fusion algorithm described above, the overlapping areas of all adjacent images are fused, while the non-overlapping areas are carefully processed, ultimately stitching all images together into a complete panoramic image. In this panoramic image, the transitions between adjacent images are natural, with no noticeable seams, making the overall image appear more harmonious and coherent. For example, when stitching a landscape panorama, the fused panorama, captured from different perspectives, seamlessly blends together. Elements such as the sky, mountains, and rivers transition naturally at the joints, presenting a vast and beautiful panoramic view.

[0072] In one embodiment, an efficient panoramic image stitching method based on multi-perspective fusion can not only fully capture image features, but also greatly improve matching efficiency, while reducing false matching, reducing the difficulty and error of homography matrix estimation, and using a global consistency optimization algorithm to globally optimize the homography matrix, avoiding multiple iterative corrections, reducing local error accumulation, and improving the overall accuracy of stitching. At the same time, it can improve the overall efficiency of panoramic image stitching; and an adaptive weighted fusion algorithm is used in the image fusion stage, which can dynamically adjust the fusion weights without repeatedly adjusting parameters, and can obtain better fusion effects, further improving the accuracy of panoramic image stitching.

[0073] In one embodiment, based on the above embodiment, before the step of performing nearest neighbor search on feature descriptors between different images based on a fast matching algorithm to match corresponding feature point pairs, the method further includes: The feature enhancement module based on deep learning abstracts and enhances feature descriptors.

[0074] In this embodiment, before performing nearest neighbor search on feature descriptors between different images based on a fast matching algorithm to match corresponding feature point pairs, a feature enhancement module based on deep learning is introduced to abstract and enhance the feature descriptors. This step can significantly improve the accuracy and stability of subsequent feature matching.

[0075] Alternatively, a lightweight CNN architecture, such as MobileNet or ShuffleNet, can be used as the deep learning architecture. These can improve processing speed while ensuring certain performance when computing resources are limited.

[0076] A feature enhancement module is designed based on a lightweight CNN architecture. This module consists of multiple convolutional layers, activation layers, and normalization layers. Convolutional layers extract features, activation layers (such as ReLU) introduce nonlinearity to the network, and normalization layers (such as Batch Normalization) accelerate network convergence and improve stability.

[0077] The input of the module is the feature descriptor extracted in the previous step, and the output is the abstracted and enhanced feature descriptor.

[0078] The feature descriptors extracted from different images are organized into a unified format and dimension. Typically, feature descriptors are represented as vectors, ensuring that the length of each feature descriptor vector is consistent so that they can be input into the deep learning model.

[0079] The organized feature descriptor dataset is divided into a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to adjust the model's hyperparameters during training, and the test set is used to evaluate the model's final performance.

[0080] Choose an appropriate loss function to measure the difference between the model's output and the expected output. Common loss functions for feature enhancement tasks include mean squared error (MSE) and cross-entropy loss. MSE is suitable for regression problems, measuring the distance between the enhanced feature descriptor and the ideal feature descriptor. Cross-entropy loss is suitable for classification problems and can be used if feature enhancement involves classification.

[0081] Choose an appropriate optimizer to update the model's parameters to minimize the loss function. Common optimizers include stochastic gradient descent (SGD), Adam, and Adagrad. The Adam optimizer combines the advantages of momentum and adaptive learning rate, achieving good training results in most cases.

[0082] The training set data is fed into the feature augmentation module, and the model output is calculated through forward propagation. The loss function is then used to calculate the loss value. Next, the optimizer is used to update the model parameters based on the loss value. This process is repeated until the model performance reaches a satisfactory level. During training, the validation set is regularly used to evaluate the model's performance, and the model's hyperparameters, such as the learning rate and batch size, are adjusted based on the validation results.

[0083] After model training is complete, the feature descriptors to be processed are fed into the trained feature enhancement module for forward propagation. The module performs a series of convolution, activation, and normalization operations on the input feature descriptors to extract more abstract and representative features.

[0084] Through the deep learning model's multi-layer nonlinear transformations, feature descriptors are further abstracted, removing some noise and redundant information while enhancing the distinguishability and stability of features. For example, after enhancement, the distance between previously similar but difficult-to-distinguish feature descriptors in feature space becomes more distinct, making them easier to accurately identify during subsequent matching.

[0085] In one embodiment, based on the above embodiment, the constraint model uses epipolar geometry constraints, determines whether feature point pairs satisfy epipolar constraint conditions by calculating a basic matrix, and eliminates feature point pairs that do not satisfy epipolar constraints.

[0086] In this embodiment, epipolar geometry is used to describe the geometric relationship between two images from different perspectives. The fundamental matrix is an algebraic representation of the epipolar geometry, mapping points in one image to epipolar lines in the other. Using the fundamental matrix, we can determine whether feature point pairs satisfy the epipolar geometry relationship, thereby selecting reliable feature point pairs.

[0087] Previously, a fast matching algorithm was used to obtain feature point pairs between different images. To calculate the fundamental matrix, at least eight non-collinear feature point pairs were selected from these pairs. This is because the fundamental matrix has nine elements, but due to scale uncertainty, it actually only has eight degrees of freedom. Therefore, at least eight feature point pairs are required to solve the linear system.

[0088] For each pair of feature points (p1, p2), where p1=(x1, y1) is the feature point in the first image and p2=(x2, y2) is the matching feature point in the second image. According to the epipolar geometry constraint, we have , after expansion, we get a linear equation about the elements of the fundamental matrix F. Substituting eight or more pairs of feature points into this formula, we can construct a linear equation system A×f=0, where A is a matrix consisting of the coordinates of the feature points and f is a vector consisting of the elements of the fundamental matrix F.

[0089] Use the singular value decomposition (SVD) method to solve the linear equation system A×f=0. Perform SVD decomposition on the matrix A and get A=UΣV T , where Σ is a diagonal matrix. Take the last column of V as the solution of f, rearrange f into a 3×3 matrix form, and get the initial basic matrix F.

[0090] Optionally, the rank of the basic matrix is 2, and the initial basic matrix obtained by the above method may not meet the rank 2 condition. Therefore, it can be rank-constrained, and F can be decomposed again by SVD, the last singular value of the Σ diagonal matrix is set to 0, and then F is recalculated to obtain a basic matrix that meets the rank constraint.

[0091] For each pair of feature points (p1, p2), substitute it into the epipolar constraint equation In the calculation, a value is obtained. This value can be regarded as the error measure between the feature point pair and the epipolar constraint.

[0092] A pre-set error threshold ϵ is used to determine whether a feature point pair satisfies the epipolar constraint. The choice of this threshold needs to be adjusted based on the specific application scenario and image data. Usually, an appropriate value can be determined through experiments.

[0093] If the calculated error value is less than the threshold ϵ, the feature point pair is considered to satisfy the epipolar constraint and is a reliable matching point pair; if the error value is greater than or equal to the threshold ϵ, the feature point pair is considered to not satisfy the epipolar constraint and may be a mismatched point pair.

[0094] Traverse all feature point pairs, and according to the above judgment results, remove the feature point pairs that do not meet the epipolar constraint conditions from the feature point pair set.

[0095] After screening, the remaining feature point pairs that meet the epipolar constraints are used in subsequent image stitching steps, such as estimating the homography matrix between adjacent images. These reliable feature point pairs can improve the accuracy of homography matrix estimation, thereby improving the quality of panoramic image stitching.

[0096] In one embodiment, based on the above embodiment, before the step of estimating the homography matrix between adjacent images based on the filtered feature point pairs, the method further includes: The filtered feature point pairs are optimized based on the random sampling consistency algorithm.

[0097] In this embodiment, RANSAC is an iterative algorithm. Its core idea is to find the set of inliers that best fit a certain mathematical model (in this scenario, the homography transformation model) from the dataset by means of random sampling. In the optimization process of feature point pairs, it can help us exclude a small number of possible incorrect matching points, so as to obtain a purer and more accurate set of feature point pairs.

[0098] Set a maximum number of iterations N, which determines the number of running rounds of the algorithm. The more iterations, the greater the possibility of finding the optimal solution, but the computing time will also increase accordingly. N can be determined according to experience or through the formula where q is the expected probability of obtaining the correct model (such as 0.99), z is the proportion of inliers among all data points (which can be roughly estimated initially), and K is the minimum number of data points required for each sampling (for homography matrix estimation, K = 4).

[0099] Set a distance threshold D to determine whether a certain feature point pair is an inlier. This threshold represents the projection error of the feature point pair after homography transformation. When the error is less than this threshold, the feature point pair is considered an inlier; otherwise, it is an outlier.

[0100] Randomly select K pairs (for homography matrix estimation, K = 4) of feature point pairs from the filtered set of feature point pairs. This is because the homography matrix is a 3×3 matrix but has scale invariance and actually has 8 degrees of freedom, so at least 4 pairs of non - collinear feature point pairs are required to solve it.

[0101] Use the extracted K pairs of feature point pairs to calculate the homography matrix H by methods such as the least - squares method. The homography matrix describes the projection transformation relationship between two images, which can map points in one image to the corresponding positions in the other image.

[0102] For each pair of feature points (p1, p2) in the filtered set of feature point pairs, transform p1 through the calculated homography matrix H to obtain the projected point P2 = H×p1. Calculate the distance (such as the Euclidean distance) e between the projected point P2 and the actual feature point p2. If e < D, then the feature point pair is considered an inlier, and count the number of inliers n.

[0103] If the number of inliers n obtained in the current iteration is greater than the previously recorded maximum number of inliers n max , then update n max =n, and take the currently calculated homography matrix H as the optimal homography matrix H best .

[0104] The algorithm stops when the number of iterations reaches the preset maximum number of iterations N. If in a certain iteration, the ratio of the number of inliers n to the total number of feature point pairs reaches a preset threshold (such as 0.8), the algorithm can also be terminated early.

[0105] According to the optimal homography matrix H best , traverse the filtered feature point pair set again, calculate the projection error of each feature point pair, and take the feature point pairs with errors less than the distance threshold D as the final inliers to form the optimized feature point pair set.

[0106] Furthermore, feature point pairs with projection errors greater than or equal to the distance threshold D are regarded as outliers and removed from the feature point pair set.

[0107] The RANSAC algorithm optimizes the resulting feature point pairs, resulting in significantly improved reliability and accuracy. These optimized feature point pairs can then be used to more accurately estimate the homography matrix between adjacent images, laying a solid foundation for subsequent operations like panoramic image stitching.

[0108] In one embodiment, based on the above embodiment, after the step of fusing the overlapping areas using an adaptive weighted fusion algorithm based on pixel information in the overlapping areas to stitch the multiple images into a panoramic image, the method further includes: Perform feathering processing on the boundary of the fusion area of the panoramic image.

[0109] In this embodiment, the appropriate feathering window size is determined based on the complexity of the fusion region boundary and the overall image resolution. If the boundary is complex and detailed, a smaller feathering window may be required to more precisely process the boundary. If the boundary is relatively simple, a larger feathering window may be selected to improve processing efficiency. Specifically, the feathering window size can be determined based on experience or experimentation, and is typically a few to several dozen pixels.

[0110] Select an appropriate feathering window shape (such as rectangle, circle, etc.) based on the actual shape of the fusion area boundary.

[0111] Within the feathering window, the feathering weight is calculated based on the distance from the pixel to the boundary of the blending area. The closer the pixel is to the boundary, the smaller its feathering weight is; the farther the pixel is from the boundary, the larger its feathering weight is.

[0112] In addition to distance, local image features, such as brightness and color, can also be considered to adjust feathering weights. For example, if there is a significant difference in brightness across a boundary, the feathering can be increased in that area, effectively reducing the feathering weight in that area. By performing local image analysis and calculating statistics of brightness or color differences, distance-based weights can be modified based on these statistics.

[0113] For each pixel in the feathering window, the pixel values on both sides of the boundary of the pixel are weighted averaged according to the calculated feathering weight.

[0114] Starting from the boundary of the blending region and expanding outwards, the weighted average formula is applied to each pixel within the feathering window. During processing, it is necessary to ensure that the pixel value is within the legal range (for example, for an 8-bit image, the pixel value range is 0-255). If the calculated pixel value is out of range, it needs to be truncated.

[0115] Through the above detailed feathering processing steps, the stitching traces at the boundaries of the panoramic image fusion area can be effectively eliminated, making the panoramic image more natural and beautiful.

[0116] In one embodiment, based on the above embodiment, after the step of fusing the overlapping areas using an adaptive weighted fusion algorithm based on pixel information in the overlapping areas to stitch the multiple images into a panoramic image, the method further includes: Perform color correction on panoramic images.

[0117] In this embodiment, after multiple images are stitched together into a panoramic image using an adaptive weighted fusion algorithm, the images captured by different cameras may have different lighting conditions, exposure parameters, and other factors, resulting in color inconsistencies between different parts of the panoramic image, affecting the overall visual effect of the image. Therefore, it is necessary to perform color correction on the panoramic image.

[0118] First, select a reference area. This should be a representative portion of the panoramic image, with relatively uniform color and normal lighting conditions. For example, if the panoramic image is a landscape, you could choose the sky or a large area of vegetation as the reference area.

[0119] Convert the panoramic image from RGB to a color space more suitable for color analysis, such as HSV (hue, saturation, value) or LAB (luminance, a-channel, b-channel). For example, LAB separates brightness information (the l-channel) from color information (the a-channel and b-channel), making it easier to adjust the colors independently.

[0120] In the selected color space, calculate the color statistics of the reference area, such as the mean and standard deviation.

[0121] The panoramic image is divided into multiple small sub-regions, which can be divided into regular grids. For each sub-region, its color statistics, such as mean and standard deviation, are also calculated in the selected color space.

[0122] Compare the color statistics of each sub-region with the reference region and calculate the difference between them. The degree of difference can be measured using a metric such as Euclidean distance.

[0123] Based on the calculated color differences, the color of each subregion is linearly transformed. To avoid obvious color changes at the subregion boundaries, a smooth transition is required. Methods such as bilinear interpolation or Gaussian filtering can be used to smoothly blend the correction results of adjacent subregions to make the color transition more natural.

[0124] The color-corrected panoramic image is converted from the LAB color space back to the RGB color space for subsequent display and storage.

[0125] Through the above color correction processing steps, the color differences between different parts of the panoramic image can be effectively eliminated, making the color of the panoramic image more natural and consistent, and improving the overall quality and visual effect of the image.

[0126] In addition, refer to Figure 2 In an embodiment of the present application, a control device Z10 is further provided, comprising: The acquisition module Z11 is used to control multiple cameras to acquire multiple images from different perspectives and perform image preprocessing; Extraction module Z12, used to perform multi-scale feature extraction on each image and generate a feature descriptor for each extracted feature point; Matching module Z13, used to perform nearest neighbor search on feature descriptors between different images based on a fast matching algorithm to match corresponding feature point pairs; A constraint module Z14 is used to perform multi-view geometric constraint screening on feature point pairs based on a constraint model; wherein the constraint model is pre-constructed using the relative position relationship between adjacent images and the projection geometry principle; An estimation module Z15 is used to estimate the homography matrix between adjacent images based on the filtered feature point pairs; Optimization module Z16, used to perform global optimization on the homography matrix using a global consistency optimization algorithm; Alignment module Z17, used to perform perspective transformation on the image based on the optimized homography matrix and align all images into the same coordinate system; An analysis module Z18 is configured to determine an overlapping region between adjacent images and analyze pixel information within the overlapping region, wherein the pixel information includes at least pixel positions; The stitching module Z19 is used to fuse the overlapping areas according to the pixel information in the overlapping areas using an adaptive weighted fusion algorithm to stitch the multiple images into a panoramic image.

[0127] Optionally, the control device Z10 may be a virtual control device (such as a virtual machine) or a physical device (such as a physical device other than a device system that can execute the corresponding method).

[0128] In addition, an embodiment of the present application also provides a device system, the internal architecture of the device system can be as follows Figure 3 As shown, it includes a processor, a memory, a communication interface, and an input interface connected via a system bus. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database is used to store data called by the computer program. The communication interface is used to communicate data with an external terminal. The input interface is used to receive signals input from an external device. When the computer program is executed by the processor, an efficient panoramic image stitching method based on multi-view fusion as described in the above embodiment is implemented.

[0129] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the present invention and does not limit the device system to which the present invention is applied. For example, in some optional embodiments, the device system may further include an output interface (not shown in the figure), and the output interface is also connected to the system bus and is used to output corresponding signals to the peripheral device.

[0130] In addition, this application also provides a computer-readable storage medium, which includes a computer program. When executed by a processor, the computer program implements the steps of the efficient panoramic image stitching method based on multi-view fusion as described in the above embodiment. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0131] In summary, the efficient panoramic image stitching method, control device, equipment system and computer-readable storage medium based on multi-perspective fusion provided in the embodiments of the present application can not only comprehensively capture image features, but also greatly improve matching efficiency, while reducing false matching, reducing the difficulty and error of homography matrix estimation, and using a global consistency optimization algorithm to globally optimize the homography matrix, avoiding multiple iterative corrections, reducing local error accumulation, and improving the overall accuracy of stitching. At the same time, it can improve the overall efficiency of panoramic image stitching; and in the image fusion stage, an adaptive weighted fusion algorithm is used, which can dynamically adjust the fusion weights without repeatedly adjusting parameters, and can obtain better fusion effects, further improving the accuracy of panoramic image stitching.

[0132] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAM bus dynamic RAM (RDRAM).

[0133] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0134] The above description is only a preferred embodiment of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An efficient panoramic image stitching method based on multi-view fusion, characterized in that: include: Control multiple cameras to capture multiple images from different perspectives and perform image preprocessing; Perform multi-scale feature extraction on each image and generate a feature descriptor for each extracted feature point; Based on the fast matching algorithm, the nearest neighbor search is performed on the feature descriptors between different images to match the corresponding feature point pairs; Based on the constraint model, multi-view geometric constraint screening is performed on the feature point pairs; wherein the constraint model is pre-constructed using the relative position relationship between adjacent images and the projection geometry principle; Estimate the homography matrix between adjacent images based on the filtered feature point pairs; The homography matrix is globally optimized using the global consistency optimization algorithm; Perform perspective transformation on the image based on the optimized homography matrix to align all images to the same coordinate system; Determining an overlapping area between adjacent images and analyzing pixel information within the overlapping area, wherein the pixel information includes at least pixel positions; According to the pixel information in the overlapping area, an adaptive weighted fusion algorithm is used to fuse the overlapping area to stitch multiple images into a panoramic image.

2. The efficient panoramic image stitching method based on multi-view fusion according to claim 1, characterized in that: The image preprocessing method includes distortion correction processing and / or image enhancement processing.

3. The efficient panoramic image stitching method based on multi-view fusion according to claim 1, characterized in that: Before the step of performing nearest neighbor search on feature descriptors between different images based on a fast matching algorithm to match corresponding feature point pairs, the method further includes: The feature enhancement module based on deep learning abstracts and enhances feature descriptors.

4. The efficient panoramic image stitching method based on multi-view fusion according to claim 1, characterized in that: The constraint model uses epipolar geometry constraints, determines whether feature point pairs satisfy epipolar constraint conditions by calculating a basic matrix, and eliminates feature point pairs that do not satisfy epipolar constraints.

5. The efficient panoramic image stitching method based on multi-view fusion according to any one of claims 1 to 4, characterized in that: Before the step of estimating the homography matrix between adjacent images based on the filtered feature point pairs, the method further includes: The filtered feature point pairs are optimized based on the random sampling consistency algorithm.

6. The efficient panoramic image stitching method based on multi-view fusion according to claim 1, characterized in that: After the step of fusing the overlapping areas using an adaptive weighted fusion algorithm based on pixel information in the overlapping areas to stitch the multiple images into a panoramic image, the method further includes: Perform feathering processing on the boundary of the fusion area of the panoramic image.

7. The efficient panoramic image stitching method based on multi-view fusion according to claim 1, characterized in that: After the step of fusing the overlapping areas using an adaptive weighted fusion algorithm based on pixel information in the overlapping areas to stitch the multiple images into a panoramic image, the method further includes: Perform color correction on panoramic images.

8. A control device, characterized in that: include: The acquisition module is used to control multiple cameras to acquire multiple images from different perspectives and perform image preprocessing; The extraction module is used to extract multi-scale features from each image and generate a feature descriptor for each extracted feature point; The matching module is used to perform nearest neighbor search on feature descriptors between different images based on a fast matching algorithm to match corresponding feature point pairs; A constraint module, configured to perform multi-view geometric constraint screening on feature point pairs based on a constraint model, wherein the constraint model is pre-constructed using the relative positional relationship between adjacent images and the principle of projection geometry; An estimation module is used to estimate the homography matrix between adjacent images based on the filtered feature point pairs; An optimization module, used to perform global optimization on the homography matrix using a global consistency optimization algorithm; The alignment module is used to perform perspective transformation on the image based on the optimized homography matrix and align all images to the same coordinate system; An analysis module, configured to determine an overlapping region between adjacent images and analyze pixel information within the overlapping region, wherein the pixel information includes at least pixel positions; The stitching module is used to fuse the overlapping areas according to the pixel information in the overlapping areas using an adaptive weighted fusion algorithm to stitch multiple images into a panoramic image.

9. A device system, characterized in that: The device system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the efficient panoramic image stitching method based on multi-perspective fusion as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the efficient panoramic image stitching method based on multi-view fusion according to any one of claims 1 to 7.

Images