A Method for Registration of Infrared and Visible Images under Complex Backgrounds

Abstract

The present invention discloses a method for registering infrared and visible light images under complex backgrounds. Aiming at the problems of large non-linear intensity differences, low similarity, and great registration difficulty between infrared and visible light images under complex backgrounds, firstly, edge feature images of the source images are respectively extracted through a gray distribution window (GDW), which excludes the interference of background region information and reduces the types of gray value pairing in the overlapping region; then, a similarity metric function (GDW-NMI) based on the combination of GDW and normalized mutual information (NMI) is constructed, which transforms the image registration problem into the problem of solving the optimal solution of the similarity metric, with small local extreme value interference and prominent global optimal solution; finally, an improved parallel search wolf pack algorithm (PSWPA) is used, which can obtain the global optimal solution of GDW-NMI through step size adjustment and multi-dimensional parallel search strategy as the geometric transformation parameters for image registration, realizing the accurate registration of infrared images and visible light images.

Description

Technical Field

[0001] The present invention belongs to the technical fields of equipment status monitoring, security monitoring, environmental monitoring, military reconnaissance, etc. using thermal infrared and visible light images, and particularly relates to a method for registering infrared and visible light images in complex backgrounds, which is based on mutual information similarity measurement and intelligent optimization algorithms. Background Art

[0002] Thermal infrared and visible light sensors have good complementarity in imaging characteristics and application scenarios. The target detection and recognition technologies based on infrared and visible light modality images have important application values. Since the images captured by these two types of sensors, infrared and visible light, are often loosely related, show different characteristics in imaging, and there are also differences in the spatial position, angle, and size of the target objects in the imaging scene, it is necessary to use image registration technology to eliminate the deviation existing in the geometric space between infrared and visible light images, so that each spatial point in the fused image scene has the same pixel position in the infrared image and the visible light image, that is, to achieve geometric alignment. By using the registration and fusion technologies of infrared and visible light, an image containing important features and richer information in the observed scene can be generated. There are two main categories of existing image registration methods: one is the feature-based image registration method, and the other is the region-based image registration method. The feature-based method uses a descriptor operator to extract the significant features of two images and uses a feature matching algorithm to establish the spatial transformation relationship between the images; while the region-based method uses a similarity measurement to judge the overall similarity degree of the images and obtains the best result of image registration by continuously searching for the maximum similarity measurement value.

[0003] Due to the difference in the imaging principles of infrared and visible light, there is a large non-linear intensity difference between the two modality images. It is suitable to use the region-based image registration method, which has advantages such as low requirements for image quality and content, and strong anti-interference ability. However, in practical applications, outdoor target objects are often affected by complex backgrounds in the scene, such as trees, foliage, clouds, buildings, surrounding equipment, etc., which greatly reduces the similarity degree of the backgrounds, textures, and contours, etc. at the gray level of the two modality images collected from the same scene. This situation of complex backgrounds causes difficulties in image registration and low accuracy.

[0004] Since complex background regions usually have the greatest differences in details between infrared images and visible light images, conventional algorithms such as Sobel and Canny are used to extract image edge features in existing region-based registration techniques, and a simple NMI function is used for similarity measurement. To address the problem of low image similarity in existing technologies, the general approach is to enrich the detail information of infrared images by enhancing them to match the rich textures of visible light images. However, infrared images reflect the temperature distribution of the scene, and the texture information present in visible light images may not exist in infrared images. Image enhancement can only highlight unclear contours, but textures that do not exist inherently cannot be presented through image enhancement. If one blindly pursues extracting more image information, a large amount of the extracted information will exist in the image difference regions, which will instead reduce similarity and increase the difficulty of image registration. Additionally, in existing technologies, conventional optimization algorithms such as the particle swarm algorithm and the wolf pack algorithm are often used to solve NMI. However, due to the non-linear intensity difference between infrared and visible light images, and the increased local differences between images caused by complex backgrounds, the convergence of these optimization algorithms is poor. At the same time, because a simple NMI function is affected by a large number of local extrema, it is difficult for these optimization algorithms to quickly and accurately search for the global optimal solution. Therefore, the practicality of existing region-based infrared and visible light image registration techniques under complex backgrounds is limited. Summary of the Invention

[0005] The objective of the present invention is to provide a suitable image registration method for target detection and recognition technologies based on the integrated fusion of infrared and visible light images. Aiming at the problems of large non-linear intensity differences, low similarity, and high registration difficulty between infrared and visible light images under complex backgrounds, the method adopts a region-based registration idea, uses the gray information of the images to measure the overall similarity of the images, constructs a similarity measurement function (GDW-NMI) based on the combination of a Grayscale Distribution Window (GDW) and Normalized Mutual Information (NMI), transforms the image registration problem into the problem of solving the optimal solution of the similarity measurement function, and then uses a Parallel Search Wolf Pack Algorithm (PSWPA), which simulates the predation behavior of wolf packs in intelligent optimization algorithms, to solve the global optimal solution and achieve accurate registration of infrared and visible light images.

[0006] To achieve the above-mentioned invention and creation objectives, the present invention adopts the following conceptions:

[0007] First, the gray distribution characteristics of infrared images and visible light images under complex backgrounds and the correlation of mutual information are analyzed. Feature extraction is performed on the two source images to be registered respectively through a gray distribution window GDW, and a similarity metric function GDW-NMI based on GDW and normalized mutual information NMI is constructed. Its local extreme interference is small and the global optimal solution is prominent. Then, referring to the exploring wolf wandering mechanism in the wolf pack algorithm, a parallel search wolf pack algorithm PSWPA is obtained through variable step size and multi-dimensional search strategies. The global optimal solution of GDW-NMI is obtained by using it as the geometric transformation parameters for image registration, realizing accurate registration of infrared images and visible light images.

[0008] According to the above concept, the present invention adopts the following five steps to realize the registration of infrared and visible light images:

[0009] Step 1: Construct a GDW function, and extract the significant edge features of the infrared image and the visible light image to be registered respectively to obtain two feature images.

[0010] Step 2: Regard the two feature images as two random variables, and construct a GDW-NMI function for measuring the similarity degree between the two images.

[0011] Step 3: Initialize the PSWPA algorithm, and randomly distribute the artificial wolves into the optimization solution space of GDW-NMI.

[0012] Step 4: Execute the PSWPA algorithm. In the solution space of GDW-NMI, the artificial wolves find all the extreme values through parallel search, continuously adjust the step size, iteratively eliminate the local extreme values, and screen out the global optimal solution.

[0013] Step 5: Perform geometric transformation on the image parameters corresponding to the global optimal solution and apply them to the source images to be registered to obtain the registered infrared and visible light images.

[0014] Compared with the prior art, the present invention has the following prominent substantial features and remarkable progress:

[0015] 1) The present invention considers the correlation between the gray distribution characteristics of infrared images and visible light images under complex backgrounds and mutual information, uses GDW to extract the significant contours of images, divides the inconsistent detail information regions between images into flat regions and sets the corresponding gray values to zero, excluding the interference of background region information, reducing the types of gray value pairings in the overlapping regions. The surface of the constructed GDW-NMI function is smooth and the main peak is sharp. When using this function for similarity measurement, it can avoid the interference of local extreme values in the search space, which is beneficial to improving the registration accuracy and the correct registration rate.

[0016] 2) The parallel search wolf pack algorithm (PSWPA) constructed in the present invention utilizes multi-dimensional parallel search and variable step sizes, has a high degree of global search freedom, enables artificial wolves to not miss extreme points during the summoning and besieging processes, searches for all extreme values in the search space of similarity measures through parallel search by artificial wolves, and filters out the global optimal solution, which can significantly improve the accuracy of image registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of a method for registering infrared and visible light images under complex backgrounds according to the present invention.

[0018] Figure 2 It is a flowchart of extracting image saliency edge features based on GDW according to the present invention.

[0019] Figure 3 It is a flowchart of the PSWPA algorithm according to the present invention.

[0020] Figure 4 It is a schematic diagram of the method for registering infrared and visible light images under complex backgrounds according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following further describes specific embodiments of the present invention with reference to the accompanying drawings.

[0022] As Figure 1 and Figure 4 shown, a method for registering infrared and visible light images under complex backgrounds is implemented as follows:

[0023] Step 1: Construct a GDW function, and respectively extract the saliency edge features of the infrared image and the visible light image to be registered to obtain two feature images. The specific implementation process is as follows:

[0024] 1) Respectively use the infrared and visible light images to be registered as the floating image and the reference image to be registered (or vice versa). Draw circles with the central pixel points of the two images as the centers and a radius of (N - 1) / 2, create a window GDW with a size of (N×N), and set the coordinates of its central pixel point as (x c , y c ) and the pixel value as g(x c , y c ), where N is generally taken as 3 or 5.

[0025] 2) As Figure 2 shown, sort the pixel values of all pixel points in the current window to obtain the maximum pixel value g max and the minimum pixel value g min, if the difference between the two is less than the set threshold, the pixel points within the current window are regarded as being in a flat area and the pixel value of its center point is set to 0; otherwise, this area is regarded as an edge area. In this way, all pixel points within the current window can be divided into three areas: a low-pixel flat area (assuming the pixel value interval size is α), a high-pixel flat area (assuming the pixel value interval size is β), and an edge area. Then, the pixel value of the center point is updated according to the area where the center point is located. If it is in the edge area, a new gray value is set by removing the maximum and minimum values and then taking the average; otherwise, it is set to 0. The detailed execution process is shown in the following flowchart:

[0026] Output: The new pixel value g′(x c ,y c ) of the central pixel point (x c ,y c )

[0027] If g min +α≥g max -β then

[0028] g′(x c ,y c )=0

[0029] Otherwise

[0030] If g(x c ,y c )∈(g min +α,g max -β) then

[0031]

[0032] Otherwise

[0033] g′(x c ,y c )=0;

[0034] 3) As Figure 2 shown, traverse all pixel points of the entire image with the GDW window and perform the pixel value update operation described in 2) to obtain the feature image.

[0035] Step 2: Regard the two feature images extracted in Step 1 as two random variables. Denote the feature map of the reference image as A and the feature map of the floating image as B, and form a similarity metric function GDW-NMI to measure the similarity degree between the two images. Its normalized mutual information NMI(A,B) is expressed as:

[0036]

[0037] Among them, H(A) and H(B) respectively represent the information entropy of A and B, and H(A,B) represents the joint entropy of A and B, which can be calculated through probabilities, as follows:

[0038]

[0039] Among them, P A (a) represents the occurrence probability of the gray value a in A; P A (a)log2P A (a) represents the amount of information of the gray value a in A; P AB (a,b) represents the probability that the gray value of the pixel corresponding to the pixel with gray value a in A is b in B, as follows:

[0040] P AB (a,b)≤P A (a) (3)

[0041] Assume that the types of gray values of all pixels corresponding to the pixel with gray value a in A in B can be taken as variables x, y, and z, then the following relationship is satisfied:

[0042] P A (a)=P AB (a,x)+P AB (a,y)+P AB (a,z) (4)

[0043] Then when x∈(0,1), y∈(0,1) and x + y∈(0,1), the following holds:

[0044] xlog2x<xlog2(x + y) (5)

[0045] ylog2y<ylog2(x + y) (6)

[0046] xlog2x + ylog2y<(x + y)log2(x + y) (7)

[0047] If multiple variables are all within the interval (0,1), and their sum is also within the interval (0,1), that is, x∈(0,1), y∈(0,1), z∈(0,1),… and x + y + z +…∈(0,1), then the following relationship is satisfied:

[0048] xlog2y + ylog2y + zlog2z +…<(x + y + z +…)log2(x + y + z +…) (8)

[0049] Therefore, when keeping other gray values unchanged, the joint entropy of all pixels with gray value a in A corresponding to all pixel gray values x, y, and z in B is greater than the joint entropy when the gray value in B is only b, that is:

[0050] -P AB (a,b)log2P AB (a,b)<-P AB (a,x)log2P AB (a,x)-P AB (a,y)log2P AB (a,y)-P AB (a,z)log2P AB (a,z)-… (9)

[0051] So far, the problem of registering two images is transformed into the problem of solving the optimal solution of the similarity measure GDW-NMI function. The wolf pack algorithm PSWPA described in the following steps 3 and 4 is used for solving. When using the PSWPA algorithm to search for the optimal solution through multiple iterations and using the corrected parameters for image geometric transformation in each iteration, finally, the mutual information of the two images reaches the maximum value, which means the registration of the two images is completed, expressed as:

[0052] T best =argmax(NMI(A,T(B))) (10)

[0053] where T(B) represents the geometric transformation performed on image B, as described in formula (15); T best represents the geometric transformation parameters when the two images are completely registered, that is, the global optimal solution of GDW-NMI.

[0054] Step 3: Randomly distribute artificial wolves into the optimization solution space of the GDW-NMI function to initialize the PSWPA algorithm. This algorithm has three types of artificial wolves with different functions (scout wolves, lead wolves, and fierce wolves), three optimization behaviors (scout wolf wandering, lead wolf summoning, and fierce wolf besieging), and two intelligent rules ("the winner takes all" lead wolf competition rule and "the fittest survive" wolf pack update rule) to complete iterative optimization. The initialization parameters include the search space S, the total number of artificial wolves M in the wolf pack, the maximum dimension number D of geometric transformation, the maximum number of loop times T scout for the wandering behavior, the maximum number of loop times T calling for the summoning and besieging behaviors, the maximum number of iterations T max , the iteration number influence factor c1, and the GDW-NMI change influence factor c2.

[0055] As Figure 1 shown, first, randomly initialize the positions of all artificial wolves, calculate the objective function value of formula (10) at the position of each artificial wolf, set the artificial wolf with the highest objective function value as the lead wolf (denoted as lead), divide the other artificial wolves into scout wolves and fierce wolves, and then set the initial search range of the scout wolves.

[0056] Step 4: As shown in Figure 3 i , execute the PSWPA algorithm. Let the initial position of artificial wolf i be X i , and the GDW-NMI objective function value at this position is called the prey odor concentration Y i in the PSWPA algorithm. Set the artificial wolf at the position with the highest prey odor concentration as the lead wolf. The number of search space dimensions of the artificial wolf is D, which is the different directions in the GDW-NMI geometric space. As can be seen from equation (15), they represent the horizontal displacement, vertical displacement, scaling ratio, and rotation angle in the image geometric transformation respectively. For each search dimension d ∈ [1, D], set the closest distance between the fierce wolf and the lead wolf in this dimension as d near , and the execution process of the three behaviors of the artificial wolf in this dimension is as follows:

[0057] 1) The exploring wolf wanders. The search step size cur of the exploring wolf i at the current iteration number T is:

[0058]

[0059] where S u and S l represent the search upper and lower limits of this dimension respectively. During the wandering process of the exploring wolf, it continuously moves towards the position with a higher (better) prey odor concentration. If it encounters a better position, update the current position, and the update formula is:

[0060]

[0061] where k is the update number, X i (k) is the original position of the exploring wolf i, and X i (k + 1) is its updated position; λ0 is a random number between [-1, 1], used to adjust the search step size of the exploring wolf. During the wandering process of the exploring wolf, the ratio of the increment of the prey odor concentration (the prey odor concentration at the new position minus the prey odor concentration at the previous position) to the prey odor concentration at the previous position is called the prey odor concentration change rate. The exploring wolf dynamically adjusts its wandering step size according to the prey odor concentration change rate, which is expressed as:

[0062]

[0063] where C represents the number of times the prey odor concentration change rate of the exploring wolf i ranks among the top in the entire wolf pack during the current wandering process. A high ranking of the prey odor concentration change rate represents a sudden increase in the GDW-NMI value. At this time, the wandering step size of this exploring wolf can be quickly reduced through equation (13), and only a small area is searched in detail, while other exploring wolves continue to maintain the step size of the previous wandering. This allows all exploring wolves to search in parallel without interfering with each other.

[0064] 2) Head wolf summoning and fierce wolf siege. When the scout wolf wanders to reach T scout , update the head wolf and perform two independent actions: head wolf summoning and fierce wolf siege. After hearing the head wolf's summons, the fierce wolves quickly move towards the position of the head wolf. When the distance between a fierce wolf and the head wolf is less than the parameter d near , the siege action is executed. The transition of a fierce wolf from the summoning action to the siege action is also determined by the parameter d near . The position update formula for fierce wolf j is:

[0065] X j (k + 1) = X j (k) + λ1 * (X lead - X j (k)) (14)

[0066] where X j (k) is the original position of the fierce wolf, X j (k + 1) is its updated position, and X lead is the position of the current head wolf; when X j (k) ≥ d near , take λ1 as a random number between [0, 1], which can ensure that the fierce wolf moves quickly towards the head wolf; when X j (k) < d near , take λ1 as a random number between [-1, 1], which can ensure that the fierce wolf searches back and forth near the head wolf. If the head wolf is replaced during this process, the new head wolf will re - execute the summoning process; otherwise, when the number of loops reaches T calling , stop the summoning and siege actions, start the survival of the fittest mechanism, and re - initialize the artificial wolf with the worst position (lowest prey odor concentration) among the artificial wolves randomly.

[0067] According to the above processes 1) and 2), every time an artificial wolf moves one step, the parameters of all its dimensions are updated once. That is, after step 3 is executed, each dimension of the artificial wolf simultaneously executes step 4, thus realizing parallel search.

[0068] As Figure 3 shown, repeat step 4 until the maximum number of iterations T max is reached. At this time, the position where the head wolf is located is the global optimal solution T best of GDW - NMI, that is, as shown in formula (10), the geometric position within the search space where the head wolf is located is the optimal geometric transformation parameter for image registration.

[0069] Step 5: For the image to be registered, perform geometric transformation using the T best parameter to obtain the registered image. As described in equation (15), the image geometric transformation model is:

[0070]

[0071] Among them, the four parameters h, v, q, and r respectively represent the translation amount in the horizontal direction (in pixels), the translation amount in the vertical direction (in pixels), the scaling ratio, and the rotation angle of the image to be registered during the geometric transformation process.

[0072] The method for registering infrared and visible light images under complex backgrounds in the above embodiments of the present invention. Aiming at the problems of large non-linear intensity difference, low similarity, and great registration difficulty between infrared and visible light images under complex backgrounds, first, the edge feature images of the source images are respectively extracted through the Gray Distribution Window (GDW), excluding the interference of background region information and reducing the types of gray value pairing in the overlapping region; then a similarity metric function (GDW-NMI) based on the combination of GDW and Normalized Mutual Information (NMI) is constructed, transforming the image registration problem into the problem of solving the optimal solution of the similarity metric, with small local extreme interference and prominent global optimal solution; finally, an improved Parallel Search Wolf Pack Algorithm (PSWPA) is used to obtain the global optimal solution of GDW-NMI through step size adjustment and multi-dimensional parallel search strategy, as the geometric transformation parameters for image registration, realizing the accurate registration of infrared images and visible light images.

[0073] The above has described the embodiments of the present invention with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made according to the purpose of the invention of the present invention. Any changes, modifications, substitutions, combinations, or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent replacement methods. As long as they meet the invention purpose of the present invention and do not deviate from the technical principle and inventive concept of the present invention, they all belong to the protection scope of the present invention.

Claims

1. A method for infrared and visible light image registration under complex backgrounds, characterized in that: By analyzing the correlation between the gray - level distribution characteristics and mutual information of infrared images and visible - light images under complex backgrounds, a gray - level distribution window (GDW) is constructed to extract features from the two source images to be registered respectively. A similarity metric function GDW - NMI based on GDW and normalized mutual information (NMI) is obtained, which has small local - extreme interference and prominent global optimal solutions. Then, by referring to the exploring - wolf wandering mechanism in the wolf - pack algorithm, a parallel - search wolf - pack algorithm (PSWPA) is constructed through variable step - size and multi - dimensional search strategies. Using it to find the global optimal solution of GDW - NMI as the geometric - transformation parameters for image registration, the accurate registration of infrared images and visible - light images is realized. The specific steps are as follows: Step 1: Construct a GDW function to extract the salient - edge features of the infrared image and visible - light image to be registered respectively, obtaining two feature images; Step 2: Regard the two feature images as two random variables and construct a GDW - NMI function for measuring the similarity degree between the two images; Step 3: Initialize the PSWPA algorithm and randomly distribute artificial wolves into the optimization - solution space of GDW - NMI; Step 4: Execute the PSWPA algorithm. In the solution space of GDW - NMI, artificial wolves find all the extreme values through parallel search, continuously adjust the step - size, iteratively eliminate local extreme values, and screen out the global optimal solution; Step 5: Perform geometric transformation on the image parameters corresponding to the global optimal solution and apply them to the source images to be registered, obtaining the registered infrared and visible - light images.

2. The method for infrared and visible light image registration in a complex background according to claim 1, characterized in that: The similarity metric function GDW - NMI, based on the idea of region - based registration, uses GDW to extract the significant contours of images. By dividing the non - consistent - detail - information regions between images into flat regions and setting the corresponding gray - level values to zero, the interference of background - region information is excluded, and the types of gray - level value pairings in the overlapping region are reduced. The surface of the constructed GDW - NMI function is smooth and the main peak is sharp. When using this function for similarity measurement, it can avoid the interference of local extreme values in the search space, which is beneficial to improving the registration accuracy and registration correct rate. Its construction process is as follows: 1) Take the infrared and visible - light images to be registered as the floating image and reference image to be registered respectively, or vice versa. Draw circles with the center pixel points of the two images as the centers and a radius of (N - 1) / 2 respectively to create a window area GDW of size (N×N); 2) Sort the pixel values of all pixel points in the current window to obtain the maximum and minimum pixel values in the window. If the difference between the two is less than the set threshold, it is considered that the pixel points in the current window are in a flat region and the pixel value of its center point is set to 0, otherwise this region is regarded as an edge region. In this way, all pixel points in the current window are divided into three regions: low - pixel flat region, high - pixel flat region, and edge region. Then, update the pixel value according to the region where the center point is located. If it is in the edge region, set a new gray - level value by taking the average after removing the maximum and minimum values, otherwise set it to 0; 3) Traverse all pixel points of the entire image using the GDW window, perform the pixel value update operation in step 2), and thus obtain the feature image; consider these two feature images as two random variables, where the feature map of the reference image is denoted as A, and the feature map of the floating image is denoted as B, to form a similarity metric function GDW-NMI for measuring the similarity degree between the two images, and its normalized mutual information NMI(A,B) is expressed as: Among them, P A (a) represents the occurrence probability that the gray value in A is a, P A (a)log2P A (a) represents the amount of information that the gray value in A is a; P B (b) represents the occurrence probability that the gray value in B is b, P B (b)log2P B (b) represents the amount of information that the gray value in B is b; P AB (a, b) represents the probability that the pixel point with the gray value a in A has the gray value b in the corresponding pixel point in B; the problem of image registration is transformed into the problem of solving the optimal solution of the similarity measure GDW-NMI function.

3. The method for registering infrared and visible light images in a complex background according to claim 1, characterized in that: The global optimal solution of GDW-NMI is obtained by using the Parallel Search Wolf Pack Algorithm (PSWPA). The algorithm utilizes multi-dimensional parallel search and variable step size, with a high degree of global search freedom, ensuring that artificial wolves do not miss extreme points during the summoning and besieging processes. All extreme values in the similarity measure search space are searched in parallel by artificial wolves, and the global optimal solution is selected as the geometric transformation parameter for image registration, achieving accurate registration of infrared and visible images. In the PSWPA algorithm, the initial position of artificial wolf i is set as X i , and the value of the GDW-NMI objective function at this position is called the prey odor concentration Y in the PSWPA algorithm i . The artificial wolf at the position with the highest prey odor concentration is set as the alpha wolf. The number of search space dimensions of artificial wolves is D, which represents different directions in the GDW-NMI geometric space, respectively indicating the horizontal displacement, vertical displacement, scaling ratio, and rotation angle in image geometric transformation. For each search dimension d ∈ [1, D], the closest distance between the beta wolf and the alpha wolf in this dimension is set as d near . The execution processes of the three behaviors of artificial wolves in this dimension are as follows: 1) Exploration Wolf Roaming: The exploration wolf i has a search step size at the current iteration number T cur as follows is: Among them, S u and S l respectively represent the search upper and lower limits of this dimension; during the exploration wolf's wandering process, it continuously moves towards the position with a higher or better prey odor concentration. If it encounters a better position, it updates the current position, and the update formula is: where k is the number of updates, and X i (k) is the original position of Wolf Scout i, and X i (k + 1) is its updated position; λ0 is a random number between [-1, 1], which is used to adjust the search step size of the Wolf Scout; during the wandering process of the Wolf Scout, the increment of the prey odor concentration, that is, the prey odor concentration at the new position minus the prey odor concentration at the previous position, and the ratio to the prey odor concentration at the previous position is called the prey odor concentration change rate; the Wolf Scout dynamically adjusts its wandering step size according to the prey odor concentration change rate, which is expressed as: Among them, C represents the number of times that the change rate of the prey odor concentration of scout wolf i ranks among the top in the entire wolf pack during the current wandering process. The fact that the change rate of the prey odor concentration ranks among the top indicates a sudden increase in the GDW-NMI value. At this time, quickly reduce the wandering step length of this scout wolf through equation (4), and only conduct a detailed search in a relatively small area, while other scout wolves continue to maintain the step length of the previous wandering. In this way, all scout wolves can perform parallel searches without interfering with each other; 2) Alpha Wolf Summoning and Fierce Wolf Siege: When the Scout Wolf roams to reach T scout , update the Alpha Wolf and perform two independent actions: Alpha Wolf Summoning and Fierce Wolf Siege; when the Fierce Wolf hears the Alpha Wolf Summoning, it quickly advances towards the position of the Alpha Wolf. When the distance between the Fierce Wolf and the Alpha Wolf is less than parameter d near , perform the siege action; the transition of the Fierce Wolf from the summoning action to the siege action is also determined by parameter d near The position update formula for Fierce Wolf j is: X j (k + 1) = X j (k) + λ1 * (X lead - X j (k)) (5) Among them, X j (k) is the original position of the fierce wolf, X j (k + 1) is its updated position, X lead is the position of the current leading wolf; when X j (k) ≥ d near , take λ1 as a random number between [0, 1] to ensure that the fierce wolf moves quickly towards the leading wolf; when X j (k) < d near , take λ1 as a random number between [-1, 1] to ensure that the fierce wolf searches back and forth near the leading wolf; if the leading wolf changes during this process, the new leading wolf re-executes the summoning process; otherwise, when the number of loops reaches T calling , stop the summoning and siege behaviors, start the survival of the fittest mechanism, and re-initialize the artificial wolf with the worst position (the lowest prey odor concentration) among the artificial wolves; According to the processes of 1) and 2) above, each time the artificial wolf moves one step, the parameters in all its dimensions are updated once; 1) and 2) are repeatedly executed until the maximum number of iterations T is reached. max At this time, the position where the leading wolf is located is the global optimal solution T of GDW-NMI. best The geometric position within the search space where the leading wolf is located is the optimal geometric transformation parameter for image registration.

Images