Distributed image registration method, system, medium, computer equipment and terminal

Abstract

The present invention belongs to the technical field of image registration, and discloses a distributed image registration method, system, medium, computer equipment and terminal. The original image to be registered is cut to obtain sub-images, and training samples are generated using the sub-images. The features of the generated samples are extracted using a pre-trained convolutional neural network (CNN). The distributed training Gaussian radial basis neural network (RBFN) learns the optimal mapping relationship between image features and registration parameters. The trained RBFN model is used to predict the registration parameters between images. The present invention extracts image features based on the CNN model to effectively obtain key image information, and at the same time proposes a distributed algorithm DD‑RBFN‑ZGS to effectively predict image registration parameters. The present invention verifies the convergence and accuracy of the algorithm through simulation experiments, and obtains better accuracy in image registration compared with end-to-end image registration algorithms and traditional image registration algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of image registration, and in particular relates to a distributed image registration method, system, medium, computer equipment and terminal. Background Art

[0002] At present, image registration technology is widely used in remote sensing data analysis, UAV automatic cruising, computer vision and other fields. There are many types of image registration problems, such as: image registration at different shooting angles, image registration under multiple shooting devices, and registration between images taken at different times.

[0003] Many classic image registration algorithms can be roughly divided into two steps. First, extract the key points of the image. Find the key points in the image that are insensitive to rotation and deformation, and describe the key points in a reasonable way, such as feature point descriptors such as Scale-invariant feature transform (SIFT) or Oriented FAST and Rotated BRIEF (ORB). Second, design a matching algorithm. Find the most likely matching point pair set from the key point sets of the two images to be registered, and use it to calculate the homography matrix parameters.

[0004] In recent years, neural networks have shown strong versatility in the field of computer vision, and using neural networks for image registration has become a preferred method.

[0005] There are many ways to use neural network technology in image registration problems. Some studies follow the traditional two-stage registration idea and use feedforward neural networks (FNNs) to learn the mapping relationship between feature point information and registration parameters. In 2000, Elhanany et al. proposed a registration algorithm based on FNN. They generated a batch of training samples by randomly rotating the template image. The rotation parameters were used as labels and the discrete cosine transform (DCT) features of the image were used to train FNN for image registration. In 2016, Yan Min et al. improved the algorithm of Elhanany et al. by using more powerful SIFT features and increasing the number of homography transformation parameters to eight. This algorithm requires training the model at each registration, resulting in efficiency being affected. To solve this problem, the algorithm only generates a small number of training samples each time and uses the extreme learning (ELM) algorithm to calculate the FNN parameters to improve efficiency.

[0006] Some studies use pre-trained CNN models to extract image features, and then use these features to calculate registration parameters. In 2018, Famao Ye et al. used some convolutional layers of a trained CNN classification model to build key point feature descriptors of images. In 2021, Fei Zhu et al. first proposed a deep CNN based on Laplacian eigenmap to extract intrinsic features of images for image registration.

[0007] There are also some studies that build efficient image registration models in an end-to-end manner, and the registration parameters can be directly predicted by the deep model. In 2016, DeTone et al. proposed a deep CNN similar to VGG to complete image registration in an end-to-end manner. In 2018, T. Nguyen et al. then proposed an improved form of the above algorithm to train CNN in an unsupervised manner. Compared with the registration algorithm based on key point matching, these two algorithms have higher registration accuracy.

[0008] As we know from the previous article, the algorithm proposed by Yan Min et al. is retrained on each individual registration task. Although it can learn the feature information of the image to be registered, it is heavily dependent on the stability of the SIFT algorithm. If the feature point extraction fails, the algorithm cannot correctly register the image. In addition, for large images, direct SIFT feature point calculation requires huge memory resources and long computing time. The algorithm is executed on a single computer with limited performance. The registration process only generates 200 training samples, and the sample richness is insufficient. The end-to-end deep network of DeTone et al. can extract features stably, but requires a huge training set and repeated parameter tuning to ensure the generalization ability of the model. Therefore, if the advantages of these two algorithms are integrated, it is expected to break through the limitations of each algorithm.

[0009] Through the above analysis, the problems and defects of the prior art are as follows:

[0010] (1) The existing FNN-based image registration algorithm trains the model each time it is registered. Considering the efficiency of the algorithm, only a small number of training samples are generated each time.

[0011] (2) Existing technologies rely heavily on the stability of the SIFT algorithm; if feature point extraction fails, the algorithm cannot correctly register the images.

[0012] (3) For large images, directly calculating SIFT feature points requires huge memory resources and a long calculation time. Too large an image will cause memory overflow (OOM), making the algorithm unable to calculate.

[0013] (4) Existing end-to-end deep networks can extract features stably, but they require large training sets and repeated parameter tuning to ensure the generalization ability of the model.

[0014] The difficulty of solving the above problems and defects is:

[0015] (1) For problems (1) and (3), if we continue to use a single-machine algorithm, we must transform the algorithm. For example, we can transform the SIFT algorithm into multiple feature extractions in different regions, which will take longer to calculate; we can reduce the number of SIFT pyramid layers, which will reduce the accuracy of feature points; it is difficult to improve the efficiency of the algorithm without losing accuracy.

[0016] (2) For problem (2), it is necessary to design a more robust artificial feature point extraction algorithm, which requires research on graphics technology, integrated signal processing, computer graphics and other disciplines, and is difficult to develop.

[0017] (3) For problem (4), it is necessary to provide a larger training data set for network training, or provide a quasi-data set for fine-tuning in actual use, or modify the network structure and use a more advanced attention model. However, the cost of such models is extremely high and there is a certain degree of difficulty in cost control.

[0018] The significance of solving the above problems and defects is:

[0019] (1) Propose a distributed algorithm that uses the power of multiple computers to solve problem (1)(3), which can take into account both algorithm efficiency and algorithm effect.

[0020] (2) For problem (2), the step of manually constructing features is omitted, which reduces the difficulty of data analysis and makes the algorithm more applicable and stable.

[0021] (3) For problem (4), instead of using an end-to-end training method, the basic model only needs to be trained once, saving training costs while ensuring the generalization ability of the model. Summary of the invention

[0022] In view of the problems existing in the prior art, the present invention provides a distributed image registration method, system, medium, computer device and terminal, and more particularly relates to a distributed image registration method, system, medium, computer device and terminal based on image cutting.

[0023] The present invention is implemented as follows: a distributed image registration method, the distributed image registration method comprising:

[0024] The original image is cut and the pre-trained convolutional neural network (CNN) is used to extract the features of the cut image. The distributed training Gaussian radial basis neural network (RBFN) learns the optimal mapping relationship between image features and registration parameters. The trained RBFN model is used to predict the registration parameters between images.

[0025] Furthermore, the distributed image registration method further includes:

[0026] The input of the model is the image to be registered. After the image to be registered is cut into sub-images of different sizes, the main image, template image and the cut sub-images are sent to the distributed network; in the distributed network, each agent automatically generates training samples according to the received sub-images, and extracts image features through the trained CNN model; using the image features, each agent node trains RBFN, and all RBFNs share the same central parameter c; the parameters of the RBFN model trained by multiple agents are globally optimized through the ZGS algorithm to obtain the final trained RBFN; the features of the image to be registered are extracted using the same method as the agent node, and the features are input into the RBFN, thereby obtaining the final prediction result of the registration parameters.

[0027] Furthermore, the distributed image registration method comprises the following steps:

[0028] Step 1: Perform image cutting to obtain sub-images;

[0029] Step 2: Build a distributed network;

[0030] Step 3: Distribute the subgraphs. Each agent in the distributed network receives a subgraph.

[0031] Step 4: Each distributed agent generates samples and extracts features;

[0032] Step 5: construct a distributed RBFN optimization algorithm;

[0033] Step six: predict image registration parameters.

[0034] Image segmentation can distribute images to agents in a distributed network for processing, which is the premise of the distributed algorithm. At the same time, each agent processes a sub-image, which can fully learn the local information of each part of the original image, which is conducive to the subsequent registration parameter prediction. The distributed optimization algorithm can make full use of the computing power of multiple computers and solve the performance problem of a single computer when processing large images. Feature extraction (CNN feature extraction) of the image enables the algorithm to accurately perform registration even when there is a lot of noise in the image.

[0035] Furthermore, the image cutting in step 1 includes:

[0036] The image is cut into sub-images of different sizes to learn the local information of each area of ​​the original image. The image to be registered is cut into 4 equal parts and 9 equal parts.

[0037] The sample generation and feature extraction in step 3 include:

[0038] In the distributed network, each agent obtains a sub-image and uses it to generate training samples. In the center of the sub-graph rectangle, and in Randomly move the four vertices of the rectangle within the range to obtain a new set of vertices. Let the coordinates of the four vertices of the rectangle be (x i ,y i ),i=1,2,3,4, the coordinates of the moved vertices are (x i ′,y i ′), i=1,2,3,4, we get H 4p =((x1-x1′,y1-y1′),...,(x4-x′4,y4-y′4)) T , the corresponding homography matrix is ​​H = DLT (H 4p ). Generate a batch of homography matrices using the described method.

[0039] Features are extracted using a CNN model that accepts two superimposed images as input to predict alignment parameters; each agent extracts image features of the training data using CNN.

[0040] According to the following formula, the generated homography matrix H is used to perform homography transformation on the sub-image received by the agent, and the transformed image is stacked on the sub-image. The stacked image is input into the CNN model for feature extraction, and the output of the CNN expansion layer is the required feature vector.

[0041]

[0042] Furthermore, the construction of the distributed network in step 4 includes:

[0043] The distributed network has 28 agents, the information transmission of the distributed network is directed, and the network topology is directed; the distributed network is connected and satisfies the weight balance, Each agent in the distributed network has a FNN with the same structure, and the selected FNN is RBFN.

[0044] The construction of the distributed RBFN optimization algorithm in step 4 includes:

[0045] (1) Gaussian Radial Basis Network

[0046] The expression of RBFN with L hidden layer nodes is as follows:

[0047]

[0048] Where x is the input of RBFN, W = [w1,w2,...,wL ] T is the output layer weight, c is called the Gaussian kernel center, ρ i (||xc i ||) = exp(-β||xc i || 2 ), h(x)=[ρ1(x),ρ2(x),...,ρ L (x)], β is the hyperparameter of the Gaussian radial basis. The loss function is defined as:

[0049]

[0050] Among them, y is the label of the sample and λ is the regularization parameter.

[0051] The parameter optimization method of Gaussian radial basis is extreme learning (ELM). Since the ZGS optimization condition requires that all distributed agents need to have the same central parameter c, the center c of RBFN is initialized randomly so that the RBFN center parameters of all agents are the same and remain unchanged during the entire training process.

[0052] After fixing the central parameter c, the expression of the RBFN of the distributed agent is:

[0053]

[0054] Among them, H rbfi =h(X i ) T represents the hidden layer output calculated by the i-th agent using local data, and the global loss function of the distributed network is:

[0055]

[0056] The global loss function formula of the distributed network satisfies the ZGS optimizability condition.

[0057] (2) Distributed Optimization Formula

[0058] According to the global loss function formula of the distributed network, we get:

[0059]

[0060] The ZGS distributed optimization formula is:

[0061]

[0062] Combined with the ZGS distributed optimization formula, the expression of the DD-RBFN-ZGS algorithm is:

[0063]

[0064] Among them, W i (k) represents the value of the RBFN output layer parameters of agent i at the kth iteration, and the parameter update is performed in discrete time; γ is a hyperparameter of the ZGS algorithm; a ij is an element in the adjacency matrix A; Ni represents the set of neighbor agent subscripts of agent i; the distributed network is connected and the weight is balanced.

[0065] (3) Convergence analysis

[0066] It is known that the objective function is quadratically differentiable and strongly convex, based on the Lyapunov function:

[0067]

[0068] The convergence of the DD-RBFN-ZGS algorithm satisfies the following formula:

[0069]

[0070] Among them, κ∈(0,1) is a constant related to the distributed network structure, and k represents the number of iterations.

[0071] Further, the image registration parameter prediction in step 5 includes:

[0072] The template image and the subject image are feature extracted in the same way as in the distributed agent, and the registration parameters are predicted using the RBFN trained by the distributed network.

[0073] Assume that the sizes of the template image and the main image are (h A ,w A ),(h B ,w B ), uniformly scaled to 128×128 size images, and use CNN to extract features. The RBFN obtained in step 4 makes a registration parameter H on this set of features. 4p =((x1-x1',y1-y1'),...,(x4-x4',y4-y4')).

[0074] The registration parameters of the original image are as follows:

[0075]

[0076] By calculating H = DLT (H' 4p ), the registration homography parameters between the original images can be obtained.

[0077] Another object of the present invention is to provide a distributed image registration system for implementing the distributed image registration method, the distributed image registration system comprising:

[0078] Image cutting module, used to cut the original image;

[0079] The sample generation and feature extraction module is used in the distributed network for each agent to automatically generate training samples based on the received sub-images and extract image features through the trained CNN model;

[0080] Distributed network construction module, used to build distributed networks;

[0081] Optimization algorithm building module, used to build distributed RBFN optimization algorithm;

[0082] The image registration parameter prediction module is used to predict the registration parameters between images using the RBFN model.

[0083] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0084] The input of the model is the image to be registered. After the image to be registered is cut into sub-images of different sizes, the main image, template image and the cut sub-images are sent to the distributed network; in the distributed network, each agent automatically generates training samples according to the received sub-images, and extracts image features through the trained CNN model; using the image features, each agent node trains RBFN, and all RBFNs share the same central parameter c; the parameters of the RBFN model trained by multiple agents are globally optimized through the ZGS algorithm to obtain the final trained RBFN; the features of the image to be registered are extracted using the same method as the agent node, and the features are input into the RBFN, thereby obtaining the final prediction result of the registration parameters.

[0085] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps:

[0086] The input of the model is the image to be registered. After the image to be registered is cut into sub-images of different sizes, the main image, template image and the cut sub-images are sent to the distributed network; in the distributed network, each agent automatically generates training samples according to the received sub-images, and extracts image features through the trained CNN model; using the image features, each agent node trains RBFN, and all RBFNs share the same central parameter c; the parameters of the RBFN model trained by multiple agents are globally optimized through the ZGS algorithm to obtain the final trained RBFN; the features of the image to be registered are extracted using the same method as the agent node, and the features are input into the RBFN, thereby obtaining the final prediction result of the registration parameters.

[0087] Another object of the present invention is to provide an information data processing terminal, which is used to implement the distributed image registration system.

[0088] Combined with all the above technical solutions, the advantages and positive effects of the present invention are as follows: the distributed image registration method provided by the present invention specifically relates to a distributed image registration algorithm based on image cutting. The algorithm is proposed to solve the registration problem of noisy images. First, the present invention cuts the original image and distributes the cut sub-images to the distributed network. Each agent in the distributed network generates a batch of training samples using the received sub-images, and uses a pre-trained convolutional neural network (CNN) to extract the image features of the training samples; then, a Gaussian radial basis neural network (RBFN) is distributedly trained to learn the optimal mapping relationship between image features and registration parameters; finally, the trained RBFN model is used to predict the registration parameters between the images to be registered. In view of the above-mentioned distributed problem, the present invention improves the zero gradient sum (ZGS) distributed optimization algorithm and proposes a distributed discrete-time model based on RBFN, called DD-RBFN-ZGS (Distributed discrete-time learning modelbased on RBFN with an improved ZGS algorithm). The convergence of DD-RBFN-ZGS is proved by Lyapunov stability theory. In this distributed network, coefficients are exchanged between communication nodes instead of original data, which means that DD-RBFN-ZGS is a privacy protection algorithm, and the original image cannot be restored by intercepting the communication data between distributed network nodes. Finally, the present invention verifies the convergence and accuracy of the algorithm through simulation experiments. Compared with the end-to-end deep network image registration algorithm and the traditional image registration algorithm based on feature point extraction, the registration algorithm based on image cutting, sample generation, and DD-RBFN-ZGS distributed training obtains better accuracy in image registration.

[0089] The present invention proposes a distributed image registration algorithm based on image cutting. The algorithm is divided into four steps. First, because the images to be registered usually have different overlapping areas in different situations, in order to make the training samples cover as many situations as possible, the present invention cuts the original image into multiple sub-images and constructs a distributed optimization network, and each agent in the network processes a sub-image respectively; each agent uses the sub-image to generate a batch of training samples, uses a trained CNN to extract image features, and distributes the training of RBFN to learn the mapping relationship between image features and registration parameters; the RBFN parameters of each agent in the distributed network are optimized by the ZGS algorithm, and the globally optimal RBFN parameters are obtained through information exchange between agents. Finally, the trained RBFN is used to predict the homography parameters between the images to be registered.

[0090] The contributions of the present invention are summarized as follows:

[0091] (1) The present invention proposes a distributed image registration algorithm based on image cutting. In order to solve the distributed problem caused by registration, the present invention proposes the DD-RBFN-ZGS algorithm and proves its convergence through Lyapunov stability analysis. The mapping relationship between the noise image features and the registration parameters is effectively learned in a distributed manner. Experimental results show that the algorithm is effective.

[0092] (2) The present invention trains a specific RBFN in each registration process and predicts the registration parameters based on the current image information. This avoids the problem of insufficient generalization ability of the end-to-end model under special circumstances. The difference between the algorithm of the present invention and the work of Yan Min et al. is that the algorithm of the present invention cuts the image at different scales and generates more training samples to ensure that the RBFN can learn the local information of the image at different scales and positions, and uses CNN instead of SIFT to extract features, which can extract feature information more robustly.

[0093] (3) The present invention transforms the end-to-end CNN model proposed by DeTone et al. to extract image features, which can effectively obtain key image information. Experiments have shown that the feature extraction method proposed in the present invention is also effective for images with a lot of noise.

[0094] (4) For the proposed distributed algorithm, the original image information is not directly transmitted between proxy nodes through the communication network. This means that the algorithm is a privacy-preserving algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0096] Figure 1 It is a flow chart of a distributed image registration method provided by an embodiment of the present invention.

[0097] Figure 2 It is a flow chart of a distributed image registration algorithm provided by an embodiment of the present invention.

[0098] Figure 3 is a structural block diagram of a distributed image registration system provided by an embodiment of the present invention;

[0099] In the figure: 1. Image cutting module; 2. Sample generation and feature extraction module; 3. Distributed network construction module; 4. Optimization algorithm construction module; 5. Image registration parameter prediction module.

[0100] Figure 4 It is a schematic diagram of the structure of a distributed image registration algorithm provided by an embodiment of the present invention.

[0101] Figure 5 It is a schematic diagram of cutting an image into different sizes provided by an embodiment of the present invention.

[0102] Figure 6 This is a schematic diagram of a connected and weight-balanced network topology provided by an embodiment of the present invention.

[0103] Figure 7 It is a schematic diagram of homography transformation in the image registration process provided by an embodiment of the present invention.

[0104] Figure 8 is a schematic diagram of the convergence of W under different parameters provided by an embodiment of the present invention; Figure 8 (d) shows a case where too large γ causes W to diverge.

[0105] Fig. 9 It is a schematic diagram comparing the registration algorithm of the present invention and the end-to-end CNN registration algorithm under different overlapping areas provided by an embodiment of the present invention.

[0106] Fig.10 It is a schematic diagram comparing the registration algorithm of the present invention with the end-to-end CNN registration algorithm and the SIFT registration algorithm under different overlapping areas and with added noise provided by an embodiment of the present invention.

[0107] Fig.11 It is a schematic diagram of different cutting strategies provided by an embodiment of the present invention.

[0108] Fig.12 It is a schematic diagram comparing the registration effects under random noise conditions provided by an embodiment of the present invention.

[0109] Fig.13 It is a schematic diagram comparing the registration effects under random noise and Gaussian noise provided by an embodiment of the present invention.

[0110] Fig.14 It is a schematic diagram for comparing the registration effects in a noise-free situation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0111] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0112] In view of the problems existing in the prior art, the present invention provides a distributed image registration method, system, medium, computer device and terminal. The present invention is described in detail below with reference to the accompanying drawings.

[0113] like Figure 1 As shown, the distributed image registration method provided by the embodiment of the present invention includes the following steps:

[0114] S101, performing image cutting to obtain sub-images;

[0115] S102, constructing a distributed network;

[0116] S103, distributing the subgraph, each agent in the distributed network receives a subgraph;

[0117] S104, each distributed agent generates samples and extracts features;

[0118] S105, constructing a distributed RBFN optimization algorithm;

[0119] S106, predicting image registration parameters.

[0120] like Figure 3 As shown, the distributed image registration system provided by the embodiment of the present invention includes:

[0121] Image cutting module 1, used for cutting the original image;

[0122] Sample generation and feature extraction module 2, used for each agent in the distributed network to automatically generate training samples according to the received sub-images and extract image features through the trained CNN model;

[0123] Distributed network construction module 3, used for constructing a distributed network;

[0124] The optimization algorithm construction module 4 is used to construct a distributed RBFN optimization algorithm;

[0125] The image registration parameter prediction module 5 is used to predict the registration parameters between images using the RBFN model.

[0126] The technical solution of the present invention is further described below in conjunction with specific embodiments.

[0127] 1. The present invention proposes a distributed image registration algorithm based on image cutting. The algorithm is proposed to solve the registration problem of noisy images. First, the present invention cuts the original image and distributes the cut sub-images to the distributed network. Each agent in the distributed network generates a batch of training samples using the received sub-images, and uses a pre-trained convolutional neural network (CNN) to extract the image features of the training samples; then, a Gaussian radial basis neural network (RBFN) is distributedly trained to learn the optimal mapping relationship between image features and registration parameters; finally, the trained RBFN model is used to predict the registration parameters between the images to be registered. In response to the above-mentioned distributed problem, the present invention improves the zero gradient sum (ZGS) distributed optimization algorithm and proposes a distributed discrete-time model based on RBFN, called DD-RBFN-ZGS (Distributed discrete-time learning model based on RBFNwith an improved ZGS algorithm). The convergence of DD-RBFN-ZGS is proved using Lyapunov stability theory. In this distributed network, coefficients are exchanged between communication nodes instead of raw data, which means that DD-RBFN-ZGS is a privacy protection algorithm. Finally, the convergence and accuracy of the algorithm are verified by simulation experiments. Compared with the end-to-end image registration algorithm and the image registration algorithm based on feature point extraction, the present invention obtains better accuracy in image registration.

[0128] The present invention proposes a distributed image registration algorithm based on image cutting. The algorithm is divided into four steps. First, the images to be registered have different overlapping areas under different circumstances. In order to make the training samples cover as many situations as possible, the present invention cuts the original image into multiple sub-images and constructs a distributed optimization network, and each agent in the network processes a sub-image respectively. Next, each agent uses the sub-image to generate a batch of training samples, uses a trained CNN to extract image features, and distributes the training of RBFN to learn the mapping relationship between image features and registration parameters. Then, the RBFN parameters of each agent in the distributed network are optimized by the ZGS algorithm, and the globally optimal RBFN parameters are obtained through information exchange between agents. Finally, the trained RBFN is used to predict the homography parameters between the images to be registered.

[0129] The contributions of the present invention are summarized as follows:

[0130] 1) The present invention proposes a distributed image registration algorithm based on image cutting. In order to solve the distributed problem generated by registration, the present invention proposes the DD-RBFN-ZGS algorithm and proves its convergence through Lyapunov stability analysis. The mapping relationship between the noise image features and the registration parameters is effectively learned in a distributed manner. Experimental results show that the algorithm is effective.

[0131] 2) The present invention trains a specific RBFN in each registration process and predicts the registration parameters based on the current image information. This avoids the problem of insufficient generalization ability of the end-to-end model under special circumstances. The difference between the algorithm of the present invention and the work of Yan Min et al. is that the algorithm of the present invention cuts the image at different scales and generates more training samples to ensure that the RBFN can learn the local information of different scales and positions of the image, and uses CNN instead of SIFT to extract features, which can extract feature information more robustly.

[0132] 3) The present invention improves the CNN model proposed by DeTone et al. to extract image features, which can effectively obtain key information of the image. Experiments have shown that the feature extraction method proposed by the present invention is also effective for images with a lot of noise.

[0133] 4) For the proposed distributed algorithm, the original image information is not directly transmitted between proxy nodes through the communication network. This means that the algorithm is a privacy-preserving algorithm.

[0134] 2. Preliminary knowledge

[0135] 2.1 Graph Theory

[0136] The connection network of a distributed multi-agent system is usually described by a graph g = {v, ε}. Here, v is a set of points representing the set of n agents in the distributed system, and ε is used to represent the connection (edge) between agents. For example, (i, j) ∈ ε means that there is a connection between agents i and j. The neighbor set N i ={j∈v:(i,j)∈ε} represents the set of all agents connected to agent i. A = [a ij ]∈R n×n Used to represent the adjacency matrix of the graph, where a ij > 0 if there is a connection between agents i and j, otherwise a ij = 0. If A is a symmetric matrix, then the graph g is an undirected graph.

[0137] 2.2 Multi-agent distributed system

[0138] Consider a first-order dynamical system with n agents. The state of the agent can be expressed as: i (t+1)=x i (t)+u i (t),

[0139] where x i (t)∈R m Used to represent the state of the i-th agent at time t, u i (t)∈R m It is used to represent the state adjustment made by the i-th agent at time t according to the states of other agents. A multi-agent system reaches consensus if and only if x converges to t→∞.

[0140] In the present invention, a multi-agent system is used to solve the prediction problem of distributed image registration coefficients. Therefore, the multi-agents not only need to reach a consensus, but also the corresponding RBFN parameters when each agent reaches a consensus need to be able to predict the optimal solution of the image registration coefficients.

[0141] 2.3 Zero Gradient Sum Algorithm

[0142] The zero gradient sum algorithm (ZGS) is a distributed optimization algorithm used to solve the optimization problem of the "sum" form. The objective function expression of the optimization problem of the "sum" form is:

[0143]

[0144] The ZGS algorithm requires that the objective function is strongly convex and second-order differentiable. If the objective function meets the conditions, ZGS can be used for distributed optimization. The ZGS distributed optimization formula is as follows:

[0145]

[0146] where x i (k)∈R n Represents the global optimal solution x in distributed node i * The estimate, represents f i The optimal solution of (x), a ij is the element in the i-th row and j-th column of the adjacency matrix, k is the number of iterations, and γ>0 is a hyperparameter of the algorithm used to control the learning rate.

[0147] 2.4 Homography Image Transformation

[0148] Images taken from different angles of the same scene can be considered to have a homography mapping. Generally, the present invention refers to the image to be transformed as the subject image and the target image as the template image. The homography mapping between the two images is expressed using the matrix H, and the homography transformation for image registration is defined as follows:

[0149]

[0150] Among them, (u,v) represents the pixel coordinates in the main image, which will be mapped to the coordinates (u',v').

[0151] The calculation formula for (u',v') is:

[0152]

[0153]

[0154] It can be seen that multiplying the homography matrix by any non-zero constant will not change the result of the homography transformation. Therefore, the present invention can multiply the homography matrix by Normalize it to standard form:

[0155]

[0156] It can be seen from formula (6) that the homography transformation of the image has eight parameters.

[0157] Since the eight parameters of H usually have different orders of magnitude and are not suitable for direct fitting by machine learning, DeTone proposed an equivalent representation of homography parameters:

[0158]

[0159] Where Δu i =u' i -u i ,Δv i =v' i -v i, (u,v)→(u',v') represents the coordinates of the two matching points in the subject image and the template image. 4p The matrix can be transformed to H through direct linear transformation (DLT). 4p Each element of has the same value range, making the model easier to train.

[0160] 3. Distributed Image Registration Algorithm

[0161] The algorithm structure proposed by the present invention is as follows Figure 4 As shown, the process is as follows Figure 2 As shown. The input of the model is the image to be registered. These images are cut into sub-images of different sizes. Then, the main image, template image and the cut sub-images are sent to the distributed network. In the distributed network, each agent automatically generates a batch of training samples based on the received sub-images, and extracts image features through the trained CNN model. Using these features, each agent node trains an RBFN, and all RBFNs share the same central parameter c. The parameters of the RBFN model trained by multiple agents are globally optimized using the ZGS algorithm to obtain the final trained RBFN. Use the same method as the agent node to extract the features of the image to be registered and input them into the RBFN to obtain the final prediction result of the registration parameters.

[0162] This invention will introduce information about each part of the distributed image registration network.

[0163] 3.1 Image Slicing

[0164] The image registration algorithm proposed in the present invention relies on the information of the overlapping parts of the images to be registered. However, in different cases, the images to be registered usually have different overlapping areas. Therefore, the present invention cuts the image into sub-images of different sizes to learn the local information of each area of ​​the original image. Figure 5 As shown, the image to be registered is cut into 4 equal parts and 9 equal parts.

[0165] 3.2 Sample generation and feature extraction

[0166] In the distributed network, after each agent obtains a sub-image, it will use this sub-image to generate a batch of training samples. Assume that the size of the original image to be registered is N×N, with The sample generation method is introduced as an example of a sub-graph of size. rectangle, and then Randomly move the four vertices of the rectangle within the range to obtain a new set of vertices. Let the coordinates of the four vertices of the rectangle be (x i ,y i ),i=1,2,3,4, the coordinates of the moved vertices are (x i ′,y i′), i=1,2,3,4, we can get H 4p =((x1-x1′,y1-y1′),...,(x4-x′4,y4-y′4)) T , the corresponding homography matrix is ​​H = DLT (H 4p ).

[0167] The present invention uses the CNN model designed by DeTone et al. to extract features. The model accepts two superimposed images as input to predict the registration parameters between them. This model is selected because it can effectively extract the spatial information between the two images by inputting two superimposed images. Compared with feature extraction algorithms such as SIFT, the output of CNN is stable and less affected by noise. According to formulas (4) and (5), the sub-image received by the agent is homographed using H generated above, and the transformed image is stacked on the sub-image and input into the CNN model for feature extraction. The output of the CNN expansion layer is the required feature vector.

[0168] 3.3 Distributed Network Structure

[0169] Since the cut template image and the main image generate 28 sub-images including the original image, the distributed network used in the present invention has 28 agents.

[0170] The distributed network used in the present invention has directional information transmission, that is, the network topology is directional. The present invention requires that the distributed network is connected and satisfies the weight balance (i.e. ). Figure 6 A network topology that meets the conditions is given. Each agent has a FNN with the same structure, and the algorithm of the present invention uses RBFN.

[0171] 3.4 Distributed RBFN Optimization Algorithm

[0172] The details of the DD-RBFN-ZGS algorithm are given in this chapter. The first part gives an improved RBFN that meets the requirements of the ZGS algorithm. In the second part, the algorithm formula of DD-RBFN-ZGS on a discrete-time directed connected network is given. The third part analyzes the convergence of the algorithm.

[0173] 3.4.1 Gaussian Radial Basis Network

[0174] The expression of RBFN with L hidden layer nodes is as follows:

[0175]

[0176] Where x is the input of RBFN, W = [w1,w2,...,w L ] Tis the output layer weight, c is called the Gaussian kernel center, ρ i (||xc i ||) = exp(-β||xc i || 2 ), h(x)=[ρ1(x),ρ2(x),...,ρ L (x)], β is the hyperparameter of the Gaussian radial basis. The loss function is defined as:

[0177]

[0178] Where y is the label of the sample and λ is the regularization parameter.

[0179] The ZGS optimization condition requires that all distributed agents have the same central parameter c. Therefore, the central parameter c of the RBFN will be uniformly initialized randomly so that the central parameter of the RBFN of all agents is the same and remains unchanged during the entire training process. After fixing the central parameter c, the expression of the RBFN of the distributed agent can be written as:

[0180]

[0181] Among them, H rbfi =h(X i ) T represents the hidden layer output calculated by the i-th agent using local data, and the global loss function of the distributed network is:

[0182]

[0183] Formula (11) is a strongly convex and second-order differentiable summation function that satisfies the ZGS optimizability condition.

[0184] 3.4.2 Distributed Optimization Formula

[0185] According to formula (11), the present invention can obtain:

[0186]

[0187] Therefore, combined with formula (2), the expression of the DD-RBFN-ZGS algorithm can be written as:

[0188]

[0189] Where W i (k) represents the value of the RBFN output layer parameter of agent i at the kth iteration, and the update of this parameter is performed in discrete time, γ is the hyperparameter of the ZGS algorithm, and a ij is an element in the adjacency matrix A. The present invention requires that the distributed network is connected and satisfies weight balance.

[0190] 3.4.3 Convergence analysis

[0191] It can be proved that the discrete-time parameter iteration according to formula (13) can converge.

[0192] It is known that the objective function is quadratically differentiable and strongly convex, in a convex set D∈R n×L , the following formula holds:

[0193]

[0194]

[0195]

[0196] Consider a Lyapunov function:

[0197]

[0198] Theorem: The convergence of the DD-RBFN-ZGS algorithm satisfies the following formula:

[0199]

[0200] Where κ∈(0,1) is a constant related to the distributed network structure, and k represents the number of iterations.

[0201] Proof: According to formulas (14) and (17), there exists a positive constant ρ such that the following formula holds on the convex set C i = {W i ∈D:Loss i (W * )-Loss i (W)-▽Loss i (W) T (W * -W)≤V(W(0))} holds:

[0202]

[0203] Where Γ represents the Laplace matrix of the distributed network topology. Differentiating (17) yields:

[0204]

[0205] Since the present invention requires that the topology of the network satisfies weight balance, there exists a positive constant ε that satisfies εΓΓ T ≤Γ+Γ TIn formula (13), the parameters satisfy W(k+1)=W(k)-γD(k)ΓW(k), where D(k)=diag{d i (k)}, If W i (k+1)≠W i (k), otherwise then θ=min{θ i} is established. Combining (19) and (20), we can get:

[0206]

[0207]

[0208]

[0209] Proof

[0210] Therefore, the DD-RBFN-ZGS iteration can make the parameter W converge at an exponential rate. From formula (22), it can be seen that a too small λ will slow down the convergence, while a too large λ will cause the parameter to diverge. According to formula (19), in the distributed optimization process, agents only exchange parameters, not raw data directly, and the algorithm is privacy-preserving.

[0211] 3.5 Image Registration Parameter Prediction

[0212] In order to use the RBFN trained by the distributed network to predict the registration parameters, the present invention needs to extract features from the template image and the main image in the same way as in the distributed agent. Assume that the sizes of the template image and the main image are (h A ,w A ),(h B ,w B ), uniformly scaled to 128×128 images, and extract features using CNN. The RBFN obtained by Algorithm 1 makes a registration parameter H on this set of features. 4p =((x1-x1',y1-y1'),...,(x4-x4',y4-y4')). Figure 7 As shown, the scaling of the image is also a homography transformation, and its homography matrices are:

[0213]

[0214] Therefore, according to the registration parameters of the scaled image predicted by RBFN, the present invention can obtain the registration parameters of the original image before scaling as follows:

[0215]

[0216] Finally, by calculating H = DLT (H' 4p ), the registration homography parameters between the original images can be obtained.

[0217] 4. Experiment

[0218] 4.1 Convergence and Accuracy Experiments

[0219] The algorithm proposed in this paper is compared with the SIFT key point extraction registration algorithm and the end-to-end CNN registration algorithm proposed by DeTone et al. The MS-COCO dataset is used as the test set. The original images in the dataset are randomly homographed to generate the images to be registered. The accuracy of the registration is determined by the predicted value and the actual H 4p The mean square error (MSE) is measured by two indicators.

[0220] For the SIFT algorithm, the standard OpenCV implementation was used in the experiment. In order to unify the metric standard, the present invention converts the homography matrix H calculated by SIFT into H 4p , and use H 4p Calculate the registration error. Because the SITF algorithm may fail in some cases (for example, if the number of key points matched is less than 4, the homography matrix cannot be calculated), the present invention sets the SIFT predicted H 4p is a 0 matrix. When performing SIFT key point registration, only key points with a similarity greater than a threshold are retained. After obtaining the paired key points, the RANSAC algorithm is used to estimate the registration parameters.

[0221] The present invention tests the performance of the distributed algorithm under different model parameters. The parameters that need to be controlled are the number of samples N generated by each agent, the number of nodes L in the hidden layer of RBFN, and the hyperparameter γ of ZGS. The maximum number of iterations of ZGS is uniformly set to 5000, and the regularization parameter λ of RBFN is set to 1e -3 , the radial basis parameter β is set to 5e -5 . Figure 8 The convergence process of the RBFN coefficient W in the distributed agent is demonstrated. Table 1 shows the mean square error of the distributed algorithm registration accuracy under different parameter combinations. It can be seen that the optimal combination of parameters is {N = 500, L = 256, γ = 5e -5}, this set of parameters is used in subsequent experiments.

[0222] Table 1. Registration mean square error of distributed registration algorithm under different parameters

[0223]

[0224] The present invention divides the data into three categories according to the degree of image overlap: the overlap is >60% (large offset), >70% (medium offset), >80% (small offset). At the same time, Gaussian noise and fog noise are added to the main image. Fig. 9 The comparison of the registration accuracy without adding noise is shown in Figure 2. The test results obtained from these three types of data are as follows: Fig.10 As shown in the figure, end-to-end CNN refers to the algorithm proposed by DeTone et al.

[0225] In experiments, the SIFT algorithm always fails when there is a lot of noise in the image. This is because it is difficult to extract enough valid key points (more than four) from the image in the presence of noise.

[0226] The distributed registration algorithm can achieve higher registration accuracy. Although the distributed algorithm and the end-to-end CNN have the same convolutional layer to extract features, for each registration, the training samples are dynamically generated by the current target image and template image to be registered, so the distributed network better learns the mapping relationship between the current image features and the registration parameters. The cutting strategy ensures that the generated samples cover the image overlap intervals in different situations as much as possible, and at least one distributed agent can learn the key features of the overlapping area. After sufficient communication between agents, the global optimal parameters are reached. Therefore, the RBFN that integrates sub-image information has generalization ability.

[0227] 4.2 Experiments on different cutting schemes

[0228] In order to prove the effectiveness of the image cutting strategy proposed in the present invention, the registration accuracy of the algorithm under different cutting times was tested. The network structure and sample generation method are the same as before. The present invention adjusts the number of sub-images, and the number of agents in the network needs to be the same as the number of sub-images. Taking strategy a as an example, the two images to be registered will generate 10 sub-images. Therefore, the distributed network requires 10 agents.

[0229] like Fig.11 As shown, the present invention adds 7 cutting strategies, including (a) no cutting + 4 equal divisions; (b) no cutting + 9 equal divisions; (c) 4 equal divisions + 9 equal divisions; (d) no cutting + 4 equal divisions + 9 equal divisions + 16 equal divisions; (e) 4 equal divisions; (f) 9 equal divisions; (g) no cutting; (h) no cutting + 4 equal divisions + 9 equal divisions. For these schemes, the number of distributed agents needs to be adjusted to the same as the number of sub-images. The model parameters are adjusted during the test process to make the distributed network converge under various strategies. The test results are shown in Table 2. It can be seen that the cutting strategy used in the present invention has certain advantages. Note that the distributed algorithm will degenerate into a centralized algorithm of strategy (g). Therefore, the corresponding parameters are not listed in the experimental data.

[0230] Table 2 Effects of distributed registration algorithm under different cutting strategies and the number of iterations required for convergence

[0231]

[0232] 4.3 Comparative Experiment

[0233] In order to demonstrate the effect of image registration, the present invention draws rectangles representing the registration positions on the template image and the subject image respectively. If the two images are accurately registered, the two rectangles will completely overlap. Note that these rectangles are not drawn when calculating the registration parameters.

[0234] 4.3.1 Comparison of registration in the presence of noise

[0235] The present invention adds noise to the subject image to simulate the registration of the noise image. Fig.12 shows the registration in the case of random noise, Fig.13 The registration for random noise and Gaussian noise cases is shown. It can be seen that the SIFT-based registration algorithm fails due to the inability to correctly extract feature points, and the algorithm of the present invention and the end-to-end CNN can give better registration. Compared with the end-to-end CNN, the algorithm of the present invention has better registration accuracy.

[0236] 4.3.2 Comparison of registration without noise

[0237] This experiment did not add noise to the image, and the registration results are as follows Fig.14 When the image to be registered is noise-free, the extraction of SIFT feature points is not disturbed, and good registration is achieved. Fig.14 The third row of shows the occasional inaccurate registration of the end-to-end CNN due to the lack of generalization ability. The effectiveness of the algorithm of the present invention is further demonstrated.

[0238] In the present invention, the present invention studies the image registration problem of neural network. Drawing on the work of Yan Min, DeTone and others, a distributed image registration algorithm based on image cutting is proposed. The algorithm cuts the image to be registered into sub-images, distributes the learning of local feature information of the original image, and completes the image registration. In view of the distributed problem, a distributed training algorithm under a discrete time directed network is proposed, called DD-RBFN-ZGS. The algorithm does not transmit the original image information during the training process and is a privacy protection algorithm. Using Lyapunov stability theory, it is proved that the DD-RBFN-ZGS algorithm is convergent.

[0239] The present invention experiments on images with different noise levels and compares the effects of different cutting strategies on registration accuracy. The results show that compared with end-to-end deep networks and registration algorithms based on feature point extraction, the algorithm of the present invention has good registration accuracy and robustness. Unlike the end-to-end algorithm, the algorithm proposed in the text uses the image to be registered for distributed training at each registration, overcoming the disadvantage of insufficient generalization ability of the end-to-end algorithm. Compared with the registration algorithm based on traditional feature point extraction, the algorithm overcomes the problem that feature points cannot be effectively extracted when there is a large amount of noise in the image. Moreover, when registering large images, feature extraction algorithms such as SIFT are difficult to execute on a single machine. The algorithm proposed in the present invention can process images in a distributed manner and use the computing power of multiple computers to deal with this situation.

[0240] This work was supported by the National Natural Science Foundation of China (Grant Nos. 62106186, 61772391, and 61966030) and the Fundamental Research Funds for the Central Universities (Grant No. JB210701).

[0241] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0242] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A distributed image registration method, characterized in that: The distributed image registration method comprises: cutting the original image, using the cut sub-images to generate training samples in a distributed manner, and using a pre-trained convolutional neural network (CNN) to extract features of the generated samples; distributed training of a Gaussian radial basis neural network (RBFN) to learn the optimal mapping relationship between image features and registration parameters; and using the trained RBFN model to predict the registration parameters between the images to be registered; The distributed image registration method comprises the following steps: Step 1: Perform image cutting to obtain sub-images; Step 2: Build a distributed network; Step 3: Distribute the subgraphs. Each agent in the distributed network receives a subgraph. Step 4: Each distributed agent generates samples and extracts features; Step 5: construct a distributed RBFN optimization algorithm; Step six, predicting image registration parameters; The construction of the distributed network in step 2 includes: the distributed network has 28 agents, the distributed network information transmission is directed, and the network topology is directed; the distributed network is connected and satisfies weight balance, Where aij is the i-th row and j-th column element in the adjacency matrix of the distributed network topology; each agent in the distributed network has an FNN with the same structure, and the selected FNN is RBFN; The construction of the distributed RBFN optimization algorithm in step 5 includes: (1) Gaussian Radial Basis Network The expression of RBFN with L hidden layer nodes is as follows: Where x is the input of RBFN, W = [w1,w2,...,w L ] T is the output layer weight, c is called the Gaussian kernel center, ρ i (||xc i ||) = exp(-β||xc i || 2 ), h(x)=[ρ1(x),ρ2(x),...,ρ L (x)], β is the hyperparameter of the Gaussian radial basis; the loss function is defined as: Among them, y is the label of the sample, λ is the regularization parameter; The parameter optimization method of Gaussian radial basis is extreme learning ELM. Since the ZGS optimization condition requires that all distributed agents need to have the same central parameter c, the center c of RBFN is initialized randomly so that the RBFN center parameters of all agents are the same and remain unchanged during the entire training process. After fixing the central parameter c, the expression of the RBFN of the distributed agent is: in, represents the hidden layer output calculated by the i-th agent using local data, and the global loss function of the distributed network is: The global loss function formula of the distributed network is a strongly convex and second-order differentiable summation function, which satisfies the ZGS optimizability condition; (2) Distributed Optimization Formula According to the global loss function formula of the distributed network, we get: The ZGS distributed optimization formula is: Combined with the ZGS distributed optimization formula, the expression of the DD-RBFN-ZGS algorithm is: Among them, W i (k) represents the value of the RBFN output layer parameters of agent i at the kth iteration, and the update of the parameters is performed in discrete time; γ is a hyperparameter of the ZGS algorithm; a ij is an element in the adjacency matrix A; Ni represents the set of neighbor agent subscripts of agent i; the distributed network is connected and satisfies weight balance; (3) Convergence analysis It is known that the objective function is quadratically differentiable and strongly convex. In the convex set D∈R n×L , the following formula holds: Where θ is a positive real number based on the Lyapunov function: The convergence of the DD-RBFN-ZGS algorithm satisfies the following formula: Among them, κ∈(0,1) is a constant related to the distributed network structure, and k represents the number of iterations.

2. The distributed image registration method according to claim 1, characterized in that: The distributed image registration method also includes: the input of the model is the image to be registered, after the image to be registered is cut into sub-images of different sizes, the main image, the template image and the cut sub-images are sent to the distributed network; in the distributed network, each agent automatically generates training samples according to the received sub-images, and extracts image features through the trained CNN model; using the image features, each agent node trains RBFN, and all RBFNs share the same central parameter c; the parameters of the RBFN model trained by multiple agents are globally optimized through the ZGS algorithm to obtain the final trained RBFN; the features of the image to be registered are extracted using the same method as the agent node, and the features are input into the RBFN, thereby obtaining the final prediction result of the registration parameters.

3. The distributed image registration method according to claim 1, characterized in that: The image cutting in step 1 includes: The image is cut into sub-images of different sizes to learn the local information of each area of ​​the original image. The image to be registered is cut into 4 equal parts and 9 equal parts. The sample generation and feature extraction in step 4 include: in the distributed network, after each agent obtains a sub-image, the sub-image is used to generate training samples; when the size of the original image to be registered is N×N, the size of the four-equal sub-image is In the center of the sub-graph rectangle, and in Randomly move the four vertices of the rectangle within the range to obtain a new set of vertices; the coordinates of the four vertices of the rectangle are (x i ,y i ),i=1,2,3,4, the coordinates of the moved vertex are (x′ i ,y′ i ), i = 1, 2, 3, 4, and get H 4p =((x1-x′1,y1-y′1),...,(x4-x′4,y4-y′4)) T , the corresponding homography matrix is ​​H = DLT (H 4p ); Generate a batch of homography matrices using the method; Extract features using a CNN model that accepts two superimposed images as input to predict alignment parameters; each agent extracts image features of the training data using CNN; According to the following formula, the generated homography matrix H is used to perform homography transformation on the sub-image received by the agent, and the transformed image is stacked on the sub-image; the stacked image is input into the CNN model for feature extraction, and the output of the CNN expansion layer is the required feature vector; 4. The distributed image registration method according to claim 1, characterized in that: The image registration parameter prediction in step six includes: The same method as in the distributed agent is used to extract features from the template image and the subject image, and the RBFN trained by the distributed network is used to predict the registration parameters; The sizes of the template image and the main image are (h A ,w A ),(h B ,w B ), uniformly scaled to 128×128 size images, and use CNN to extract features; the RBFN obtained by Algorithm 1 makes a registration parameter H on this set of features 4p = prediction of ((x1-x1′,y1-y1′),...,(x4-x4′,y4-y4′)); The scaling of the image is a homography transformation, and the homography matrices are: According to the registration parameters of the scaled image predicted by RBFN, the registration parameters of the original image before scaling are obtained as follows: By calculating H = DLT (H' 4p ), the registration homography parameters between the original images can be obtained.

5. A distributed image registration system for implementing the distributed image registration method according to any one of claims 1 to 4, characterized in that: The distributed image registration system comprises: Image cutting module, used to cut the original image; The sample generation and feature extraction module is used in the distributed network for each agent to automatically generate training samples based on the received sub-images and extract image features through the trained CNN model; Distributed network construction module, used to build distributed networks; Optimization algorithm building module, used to build distributed RBFN optimization algorithm; The image registration parameter prediction module is used to predict the registration parameters between images using the RBFN model.

6. A computer device, characterized in that: The computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the distributed image registration method according to any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the distributed image registration method according to any one of claims 1 to 4.

8. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the distributed image registration system as claimed in claim 5.

Images