An adaptive intelligent registration method for UAV remote sensing images

Abstract

The present invention discloses an adaptive unmanned aerial vehicle remote sensing image intelligent registration method, the method includes data collection, basic enhancement processing, registration accuracy optimization and adaptive intelligent registration. The present invention relates to the field of image registration technology, specifically refers to an adaptive unmanned aerial vehicle remote sensing image intelligent registration method, the present invention adopts two angles of basic enhancement and adaptive registration optimization to enhance adaptive performance, optimizes data quality through adaptive basic attribute enhancement, and improves the intelligence and adaptability of rigid transformation parameters through a registration accuracy optimization model combined with multiple parameter data training; adopts an adaptive image registration optimization method based on the combination of teaching learning and reinforcement learning, by combining reinforcement learning and teaching learning, it can not only intelligently optimize the optimal solution of parameters, but also adaptively adjust parameters according to the provided image flight parameter information, thereby optimizing the registration accuracy of the intelligent registration result.

Description

Technical Field

[0001] The present invention relates to the technical field of image registration, and in particular to an adaptive unmanned aerial vehicle remote sensing image intelligent registration method. Background Art

[0002] The adaptive UAV remote sensing image intelligent registration method is a technology that integrates deep learning and optimization algorithms to achieve accurate matching and alignment of remote sensing images. This method extracts multi-scale features, adapts to complex terrain, and uses an adaptive optimization strategy to dynamically adjust registration parameters to ensure registration accuracy in multiple scenes and multiple resolutions. Its main functions include improving the data fusion effect of UAV remote sensing images in disaster monitoring, land surveying and ecological assessment, and supporting intelligent analysis and decision-making of multi-source data.

[0003] However, in the existing intelligent registration methods for UAV remote sensing images, the intelligence of remote sensing image registration is mainly reflected in the automatic acquisition of rigid transformation parameters, while the adaptability depends on the intelligent understanding and acquisition of the image's own attributes. Although the existing technology has slightly improved in terms of intelligence, its performance in terms of adaptability is relatively general. Traditional methods are difficult to apply to the increasingly complex clusters of UAV remote sensing images, which leads to technical problems that both intelligence and adaptability need to be improved; in the existing adaptive intelligent registration methods for UAV remote sensing images, there is a technical problem that the computing performance of traditional teaching-based learning methods for large-scale problems cannot meet the computing resources they consume. At the same time, although the teaching-based learning method helps to intelligently calculate the optimal solution of rigid change parameters, it cannot meet the needs of adaptability. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an adaptive UAV remote sensing image intelligent registration method. In the existing UAV remote sensing image intelligent registration method, the intelligence of remote sensing image registration is mainly reflected in the automatic acquisition of rigid transformation parameters, while the adaptability depends on the intelligent understanding and acquisition of the image's own attributes. Although the prior art has slightly improved in terms of intelligence, its performance in terms of adaptability is relatively general. Traditional methods are difficult to apply to the increasingly complex UAV remote sensing image clusters, which leads to the technical problem that both intelligence and adaptability need to be improved. This solution creatively adopts two angles, basic enhancement and adaptive registration optimization, to enhance the adaptive performance. By enhancing the adaptive basic attributes, data quality is optimized. By combining the registration accuracy optimization model trained with multiple parameter data, the intelligence and adaptability of the rigid transformation parameters are improved. In view of the existing In the adaptive UAV remote sensing image intelligent registration method, there is a problem that the computing performance of the traditional teaching-based learning method for large-scale problems cannot meet the computing resources it consumes. At the same time, although the teaching-based learning method is helpful for the intelligent calculation of the optimal solution of rigidly changing parameters, it cannot meet the technical problem of adaptability. This scheme creatively adopts an adaptive image registration optimization method based on the combination of teaching-based learning and reinforcement learning. By combining reinforcement learning and teaching-based learning, it can not only intelligently optimize the optimal solution of parameters, but also adaptively adjust parameters according to the provided image flight parameter information, thereby optimizing the registration accuracy of the intelligent registration results, and is particularly suitable for remote sensing image registration tasks in complex scenes. It strengthens the adaptability of registration, provides powerful assistance for the intelligent registration of UAV remote sensing images, and also provides sufficient and accurate data support for subsequent in-depth UAV remote sensing image processing, recognition and analysis tasks.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides an adaptive UAV remote sensing image intelligent registration method, the method comprising the following steps:

[0006] Step S1: data collection;

[0007] Step S2: basic enhancement processing;

[0008] Step S3: optimizing registration accuracy;

[0009] Step S4: Adaptive intelligent registration.

[0010] Further, in step S1, the data collection is used to collect the original data set required for the adaptive UAV remote sensing image intelligent registration, specifically, to obtain the remote sensing image registration original data set from the original acquisition data set of the UAV remote sensing image through data collection;

[0011] The remote sensing image registration original data set specifically includes remote sensing image data, flight parameter data and initial registration transformation information.

[0012] Further, in step S2, the basic enhancement processing is used to perform adaptive basic enhancement on the remote sensing image, specifically to perform basic enhancement operation on the remote sensing image data in the remote sensing image registration original data set to obtain the remote sensing image registration enhanced data set, specifically including the following steps:

[0013] Step S21: image denoising, specifically, performing noise removal operations on the remote sensing image data using median filtering and Gaussian filtering in sequence to obtain denoised and optimized remote sensing image data;

[0014] Step S22: attribute enhancement, specifically enhancing the contrast, brightness and color spectrum of the denoised and optimized remote sensing image data to obtain attribute enhanced denoised and optimized remote sensing image data;

[0015] Step S23: sharpening and enhancement, specifically, using a Laplace operator to perform an edge enhancement operation on the attribute enhanced denoising and optimized remote sensing image data to obtain sharpened and enhanced image data;

[0016] Step S24: geometric correction and enhancement, specifically, performing geometric correction on the sharpened and enhanced image data according to the flight parameter data in the remote sensing image registration original data set to obtain geometrically optimized and corrected image data;

[0017] Step S25: Adaptive enhancement evaluation, specifically, performing manual quality evaluation on each image, and adaptively executing the operations of steps S21 to S24 in sequence according to the results of the manual quality evaluation, to obtain adaptive enhanced image data;

[0018] Step S26: basic enhancement processing, specifically obtaining a remote sensing image registration enhanced data set through the adaptive enhancement evaluation.

[0019] Further, in step S3, the registration accuracy optimization is used to adaptively and intelligently optimize the fine registration process, specifically, based on the remote sensing image registration enhanced data set, an adaptive image registration accuracy optimization method combining teaching learning and reinforcement learning is used to optimize the registration accuracy, and the optimal value of the rigid transformation parameter is obtained. Fine registration is performed by applying the optimal value of the rigid transformation parameter to obtain fine registration image data, which specifically includes the following steps:

[0020] Step S31: Modeling the rigid transformation problem, specifically constructing a registration problem model, and setting an objective function to obtain a combination of parameters to be optimized;

[0021] Step S32: Initializing the reinforcement learning agent, specifically constructing a reinforcement learning optimization agent, and specifically constructing an agent parameter combination, including action space parameters, state space parameters, and reward parameters;

[0022] The action space parameters are used to represent the rotation and translation parameters in the state of fine image registration;

[0023] The state space parameter is used to represent the combination of the action parameter and the registration error in the state of fine image registration;

[0024] The reward parameter is used to indicate that the reward parameter is increased as the registration error decreases;

[0025] The calculation formula of the agent parameter combination is:

[0026] ;

[0027] In the formula, A t is the agent parameter combination, a t is the action space parameter, s t is the state space parameter, r t is the reward parameter;

[0028] Step S33: constructing a teaching and learning model, specifically constructing a standard teaching and learning model, which is used as a basic model for rigid change optimization, and performing adaptive learning and updating of rigid transformation parameters by constructing a teacher role function and a student self-learning function;

[0029] The teacher role function is used to represent the optimal solution of the rigid transformation parameters;

[0030] The student self-learning function is used to represent a rigid transformation parameter combination for optimizing and updating toward an optimal solution;

[0031] Step S34: construct a reinforcement learning model, specifically, apply a deep Q network to update the agent parameter combination, and obtain the optimal parameter combination output of reinforcement learning by maximizing the Q value calculation. The calculation formula for updating the agent parameter combination is:

[0032] ;

[0033] In the formula, is the optimal parameter combination output of reinforcement learning, a t is the action space parameter, s t is the state space parameter, Q(·) is the Q value calculation function, is the learning rate parameter, r t is the reward parameter, is the discount factor, max is the maximum value function, a is the optimal action space parameter, st+1 is the state space parameter of the next iteration;

[0034] Step S35: constructing a teaching reinforcement learning combined model, specifically applying the reinforcement learning model in step S34, improving the teacher role function and the student self-learning function in the teaching learning model, obtaining a reinforcement teacher update function and a reinforcement student self-learning function, and performing parameter updates to obtain the optimal value of the rigid transformation parameter;

[0035] The reinforcement teacher update function is used to represent the optimal solution of the rigid transformation parameters after reinforcement learning optimization, and the calculation formula is:

[0036] ;

[0037] In the formula, It is the output of the enhanced teacher update function after enhanced teaching learning and updating, which is used to represent the optimal solution of the rigid transformation parameters obtained by the teacher update function guiding the student self-learning function. i is the i-th student parameter, used to represent the i-th rigid transformation parameter combination, rand is a random parameter, S teacher is the teacher parameter, which is used to represent the optimal solution of the rigid transformation parameters, Q(·) is the Q value calculation function, is the learning rate parameter, r t is the reward parameter, is the discount factor, max is the maximum value function, a is the optimal action space parameter, s t+1 is the state space parameter of the next iteration;

[0038] The reinforcement student self-learning function is used to represent the rigid transformation parameter combination for optimizing and updating the optimal solution in combination with reinforcement learning, and the calculation formula is:

[0039] ;

[0040] In the formula, is the output of the enhanced student self-learning function after enhanced teaching and learning, which is used to represent the optimal value of the rigid transformation parameter, S i is the i-th student parameter, used to represent the i-th rigid transformation parameter combination, rand is a random parameter, S j is related to the student parameter S i The closest combination of rigid transformation parameters;

[0041] Step S36: Registration accuracy optimization, specifically, training the registration accuracy optimization model through the rigid transformation problem modeling, the reinforcement learning agent initialization, the construction of the teaching learning model, the construction of the reinforcement learning model and the construction of the teaching reinforcement learning combined model to obtain the registration accuracy optimization model, and by using the registration accuracy optimization model, the registration accuracy is optimized to obtain the optimal value of the rigid transformation parameters.

[0042] Further, in step S4, the adaptive intelligent registration is used to perform adaptive intelligent registration according to the optimized parameter combination, specifically, to perform fine registration according to the optimal value of the rigid transformation parameter to obtain a registered image output, and to obtain matching feature data by performing accelerated robust feature extraction and cluster feature matching on the registered image output, and to eliminate abnormalities from the registered image by eliminating outliers, and to obtain an adaptive intelligent registered image output by performing an image combination operation;

[0043] The cluster feature matching specifically adopts a K nearest neighbor matching algorithm;

[0044] The outlier elimination operation specifically uses a random sample consensus algorithm to eliminate mismatches.

[0045] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0046] (1) In the existing intelligent registration methods for UAV remote sensing images, the intelligence of remote sensing image registration is mainly reflected in the automatic acquisition of rigid transformation parameters, while the adaptability depends on the intelligent understanding and acquisition of the image's own attributes. Although the existing technology has slightly improved in terms of intelligence, its adaptability performance is relatively general. Traditional methods are difficult to apply to the increasingly complex UAV remote sensing image clusters, which leads to technical problems that both intelligence and adaptability need to be improved. This solution creatively adopts two perspectives, basic enhancement and adaptive registration optimization, to enhance the adaptive performance. By enhancing the adaptive basic attributes, the data quality is optimized. By combining the registration accuracy optimization model trained with multiple parameter data, the intelligence and adaptability of the rigid transformation parameters are improved.

[0047] (2) In the existing adaptive UAV remote sensing image intelligent registration methods, there is a problem that the computing performance of the traditional teaching-based learning method for large-scale problems cannot meet the computing resources it consumes. At the same time, although the teaching-based learning method is helpful for the intelligent calculation of the optimal solution of rigidly changing parameters, it cannot meet the technical problem of adaptability. This scheme creatively adopts an adaptive image registration optimization method based on the combination of teaching-based learning and reinforcement learning. By combining reinforcement learning and teaching-based learning, it can not only intelligently optimize the optimal solution of parameters, but also adaptively adjust the parameters according to the provided image flight parameter information, thereby optimizing the registration accuracy of the intelligent registration results. It is also particularly suitable for remote sensing image registration tasks in complex scenes, enhancing the adaptability of registration, providing strong assistance for the intelligent registration of UAV remote sensing images, and providing sufficient and accurate data support for subsequent deep-level UAV remote sensing image processing, recognition and analysis tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a process flow of an adaptive UAV remote sensing image intelligent registration method provided by the present invention;

[0049] Figure 2 This is a schematic diagram of the basic enhancement process in step S2;

[0050] Figure 3 Schematic diagram of the process of optimizing the registration accuracy in step S3.

[0051] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0054] Example 1, see Figure 1The present invention provides an adaptive UAV remote sensing image intelligent registration method, which comprises the following steps:

[0055] Step S1: data collection;

[0056] Step S2: basic enhancement processing;

[0057] Step S3: optimizing registration accuracy;

[0058] Step S4: Adaptive intelligent registration.

[0059] Example 2, see Figure 1 In step S1, the data collection is used to collect the original data set required for the adaptive UAV remote sensing image intelligent registration, specifically, to obtain the remote sensing image registration original data set from the original acquisition data set of the UAV remote sensing image through data collection;

[0060] The remote sensing image registration original data set specifically includes remote sensing image data, flight parameter data and initial registration transformation information;

[0061] The remote sensing image data specifically includes remote sensing image data collected by drones at different times, different flight altitudes and different viewing angles;

[0062] The flight parameter data specifically includes flight path parameters, timestamp parameters and camera parameters;

[0063] The flight path parameters specifically include flight route, flight altitude and parting angle parameters;

[0064] The timestamp parameter specifically refers to the acquisition time parameter of each remote sensing image;

[0065] The camera parameters specifically include camera focal length parameters, distortion parameters and resolution parameters;

[0066] The initial registration transformation information specifically includes a rough registration initial result and a rough registration parameter;

[0067] The coarse registration parameters specifically refer to transformation matrix parameters used in coarse registration, including rotation matrix, translation vector and scale factor parameters.

[0068] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, the basic enhancement processing is used to perform adaptive basic enhancement on the remote sensing image, specifically to perform basic enhancement operation on the remote sensing image data in the remote sensing image registration original data set to obtain a remote sensing image registration enhanced data set, specifically including the following steps:

[0069] Step S21: image denoising, specifically, performing noise removal operations on the remote sensing image data using median filtering and Gaussian filtering in sequence to obtain denoised and optimized remote sensing image data;

[0070] Step S22: attribute enhancement, specifically enhancing the contrast, brightness and color spectrum of the denoised and optimized remote sensing image data to obtain attribute enhanced denoised and optimized remote sensing image data;

[0071] The contrast is specifically enhanced by histogram equalization;

[0072] The brightness is specifically enhanced by gamma correction;

[0073] The color spectrum is specifically enhanced by adjusting the band ratio to enhance the multi-spectral information;

[0074] Step S23: sharpening and enhancement, specifically, using a Laplace operator to perform an edge enhancement operation on the attribute enhanced denoising and optimized remote sensing image data to obtain sharpened and enhanced image data;

[0075] Step S24: geometric correction and enhancement, specifically, performing geometric correction on the sharpened and enhanced image data according to the flight parameter data in the remote sensing image registration original data set to obtain geometrically optimized and corrected image data;

[0076] Step S25: Adaptive enhancement evaluation, specifically, performing manual quality evaluation on each image, and adaptively executing the operations of steps S21 to S24 in sequence according to the results of the manual quality evaluation, to obtain adaptive enhanced image data;

[0077] Step S26: basic enhancement processing, specifically obtaining a remote sensing image registration enhanced data set through the adaptive enhancement evaluation.

[0078] By performing the above operations, in the existing UAV remote sensing image intelligent registration method, the intelligence of remote sensing image registration is mainly reflected in the automatic acquisition of rigid transformation parameters, while the adaptability depends on the intelligent understanding and acquisition of the image's own attributes. Although the existing technology has slightly improved in intelligence, the performance in adaptability is relatively general. Traditional methods are difficult to apply to the increasingly complex UAV remote sensing image clusters, which leads to the technical problem that both intelligence and adaptability need to be improved. This solution creatively adopts two angles of basic enhancement and adaptive registration optimization to enhance the adaptive performance. By enhancing the basic attributes adaptively, the data quality is optimized. By combining the registration accuracy optimization model trained with multiple parameter data, the intelligence and adaptability of the rigid transformation parameters are improved.

[0079] Example 4, see Figure 1 and Figure 3This embodiment is based on the above embodiment. In step S3, the registration accuracy optimization is used to perform adaptive intelligent optimization on the fine registration process. Specifically, based on the remote sensing image registration enhanced data set, an adaptive image registration accuracy optimization method combining teaching learning and reinforcement learning is used to optimize the registration accuracy, and the optimal value of the rigid transformation parameter is obtained. Then, fine registration is performed by applying the optimal value of the rigid transformation parameter to obtain fine registration image data. Specifically, the following steps are included:

[0080] Step S31: Modeling the rigid transformation problem, specifically constructing a registration problem model, and setting an objective function to obtain a combination of parameters to be optimized;

[0081] The calculation formula of the parameter combination to be optimized is:

[0082] ;

[0083] In the formula, O is the parameter combination to be optimized, R0 is the initial value of the rotation matrix parameter, is the initial value of the horizontal coordinate index of the translation vector, is the initial value of the vertical coordinate index of the translation vector;

[0084] The registration problem model specifically refers to taking the rigid transformation model as the registration target problem model and defining the rigid transformation model. The calculation formula is:

[0085] ;

[0086] Where T(x,y) is the coordinate data after rigid transformation, R is the rotation matrix, L is the translation vector, (x,y) is the pixel coordinate in the original image used for registration, x is the horizontal coordinate index, and y is the vertical coordinate index;

[0087] The objective function is used to minimize the registration error between the original image and the registered image, and specifically adopts a mutual information calculation function as the objective function;

[0088] Step S32: Initializing the reinforcement learning agent, specifically constructing a reinforcement learning optimization agent, and specifically constructing an agent parameter combination, including action space parameters, state space parameters, and reward parameters;

[0089] The action space parameters are used to represent the rotation and translation parameters in the state of fine image registration, and the calculation formula is:

[0090] ;

[0091] In the formula, a t is the action space parameter, t is the iteration index, R t are the rotation matrix parameters, is the horizontal coordinate parameter of the translation vector, is the vertical coordinate parameter of the translation vector;

[0092] The state space parameter is used to represent the combination of the action parameter and the registration error in the state of fine image registration, and the calculation formula is:

[0093] ;

[0094] In the formula, s t is the state space parameter, E t is the registration error, specifically the mutual information error, a t is the action space parameter;

[0095] The reward parameter is used to indicate the increase of the reward parameter as the registration error decreases. The calculation formula is:

[0096] ;

[0097] In the formula, r t is the reward parameter, E t is the registration error, MI(·) is the mutual information error calculation function, T source is the pixel coordinate data in the original image, R t is the rotation matrix parameter, L t is the translation vector parameter, T target (R t , L t ) is the rigid transformation parameter (R t ,L t ) pixel coordinate data after rigid transformation;

[0098] The calculation formula of the agent parameter combination is:

[0099] ;

[0100] In the formula, A t is the agent parameter combination, a t is the action space parameter, s t is the state space parameter, r t is the reward parameter;

[0101] Step S33: constructing a teaching and learning model, specifically constructing a standard teaching and learning model, which is used as a basic model for rigid change optimization, and performing adaptive learning and updating of rigid transformation parameters by constructing a teacher role function and a student self-learning function;

[0102] The teacher role function is used to represent the optimal solution of the rigid transformation parameters;

[0103] The student self-learning function is used to represent a rigid transformation parameter combination for optimizing and updating toward an optimal solution;

[0104] Step S34: construct a reinforcement learning model, specifically, apply a deep Q network to update the agent parameter combination, and obtain the optimal parameter combination output of reinforcement learning by maximizing the Q value calculation. The calculation formula for updating the agent parameter combination is:

[0105] ;

[0106] In the formula, is the optimal parameter combination output of reinforcement learning, a t is the action space parameter, s t is the state space parameter, Q(·) is the Q value calculation function, is the learning rate parameter, r t is the reward parameter, is the discount factor, max is the maximum value function, a is the optimal action space parameter, s t+1 is the state space parameter of the next iteration;

[0107] Step S35: constructing a teaching reinforcement learning combined model, specifically applying the reinforcement learning model in step S34, improving the teacher role function and the student self-learning function in the teaching learning model, obtaining a reinforcement teacher update function and a reinforcement student self-learning function, and performing parameter updates to obtain the optimal value of the rigid transformation parameter;

[0108] The reinforcement teacher update function is used to represent the optimal solution of the rigid transformation parameters after reinforcement learning optimization, and the calculation formula is:

[0109] ;

[0110] In the formula, It is the output of the enhanced teacher update function after enhanced teaching learning and updating, which is used to represent the optimal solution of the rigid transformation parameters obtained by the teacher update function guiding the student self-learning function. i is the i-th student parameter, used to represent the i-th rigid transformation parameter combination, rand is a random parameter, S teacher is the teacher parameter, which is used to represent the optimal solution of the rigid transformation parameters, Q(·) is the Q value calculation function, is the learning rate parameter, r t is the reward parameter, is the discount factor, max is the maximum value function, a is the optimal action space parameter, s t+1 is the state space parameter of the next iteration;

[0111] The reinforcement student self-learning function is used to represent the rigid transformation parameter combination for optimizing and updating the optimal solution in combination with reinforcement learning, and the calculation formula is:

[0112] ;

[0113] In the formula, is the output of the enhanced student self-learning function after enhanced teaching and learning, which is used to represent the optimal value of the rigid transformation parameter, S i is the i-th student parameter, used to represent the i-th rigid transformation parameter combination, rand is a random parameter, S j is related to the student parameter S i The closest combination of rigid transformation parameters;

[0114] Step S36: Registration accuracy optimization, specifically, training the registration accuracy optimization model through the rigid transformation problem modeling, the reinforcement learning agent initialization, the construction of the teaching learning model, the construction of the reinforcement learning model and the construction of the teaching reinforcement learning combined model to obtain the registration accuracy optimization model, and by using the registration accuracy optimization model, the registration accuracy is optimized to obtain the optimal value of the rigid transformation parameters.

[0115] By performing the above operations, in view of the fact that in the existing adaptive UAV remote sensing image intelligent registration method, the computing performance of the traditional teaching-based learning method for large-scale problems cannot meet the computing resources it consumes, and although the teaching-based learning method is helpful for the intelligent calculation of the optimal solution of rigidly changing parameters, it cannot meet the technical problem of adaptability, this scheme creatively adopts an adaptive image registration optimization method based on the combination of teaching-based learning and reinforcement learning. By combining reinforcement learning and teaching-based learning, it can not only intelligently optimize the optimal solution of parameters, but also adaptively adjust parameters according to the provided image flight parameter information, thereby optimizing the registration accuracy of the intelligent registration result, and being particularly suitable for remote sensing image registration tasks in complex scenes, enhancing the adaptability of registration, providing powerful assistance for the intelligent registration of UAV remote sensing images, and also providing sufficient and accurate data support for subsequent in-depth UAV remote sensing image processing, recognition and analysis tasks.

[0116] Example 5, see Figure 1 , this embodiment is based on the above embodiment, in step S4, the adaptive intelligent registration is used to perform adaptive intelligent registration according to the optimized parameter combination, specifically, to perform fine registration according to the optimal value of the rigid transformation parameter, to obtain the registered image output, and to obtain matching feature data by performing accelerated robust feature extraction and cluster feature matching on the registered image output, and to eliminate the abnormalities of the registered image by eliminating the outliers, and to obtain the adaptive intelligent registered image output by image combination operation;

[0117] The cluster feature matching specifically adopts a K nearest neighbor matching algorithm;

[0118] The outlier elimination operation specifically uses a random sample consensus algorithm to eliminate mismatches.

[0119] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0120] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0121] The present invention and its embodiments are described above, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it and design structural methods and embodiments similar to the technical solution without creativity without departing from the purpose of the invention, they should all fall within the protection scope of the present invention.

Claims

1. An adaptive UAV remote sensing image intelligent registration method, characterized by: The method comprises the following steps: Step S1: Data collection, from the original acquisition data set of the UAV remote sensing image, through data collection, the original data set of remote sensing image registration is obtained; Step S2: basic enhancement processing, performing basic enhancement operations on the remote sensing image data in the remote sensing image registration original data set to obtain a remote sensing image registration enhanced data set; Step S3: Registration accuracy optimization, which is used to perform adaptive intelligent optimization on the fine registration process. Specifically, based on the remote sensing image registration enhanced data set, an adaptive image registration accuracy optimization method combining teaching learning and reinforcement learning is used to optimize the registration accuracy, and the optimal value of the rigid transformation parameter is obtained. Fine registration is performed by applying the optimal value of the rigid transformation parameter to obtain fine registration image data, which specifically includes the following steps: Step S31: Modeling of rigid transformation problem; Step S32: Initialization of reinforcement learning agent; Step S33: Construction of teaching learning model; Step S34: Construction of reinforcement learning model; Step S35: Construction of teaching reinforcement learning combined model; Step S36: Optimization of registration accuracy; In step S35, the teaching reinforcement learning combined model is constructed, specifically, the reinforcement learning model in step S34 is applied to improve the teacher role function and the student self-learning function in the teaching learning model, to obtain the reinforcement teacher update function and the reinforcement student self-learning function, and to update the parameters to obtain the optimal value of the rigid transformation parameter; The reinforcement teacher update function is used to represent the optimal solution of the rigid transformation parameters after reinforcement learning optimization, and the calculation formula is: In the formula, is the output of the reinforcement teacher update function after reinforcement teaching and learning update, which is used to represent the optimal solution of the rigid transformation parameters obtained by the reinforcement teacher update function to guide the student self-learning function. i is the i-th student parameter, used to represent the i-th rigid transformation parameter combination, rand is a random parameter, S teacher is the teacher parameter, which is used to represent the optimal solution of the rigid transformation parameters, Q(·) is the Q value calculation function, α is the learning rate parameter, r t is the reward parameter, γ is the discount factor, max is the maximum value function, a is the optimal action space parameter, s t+1 is the state space parameter of the next iteration; The reinforcement student self-learning function is used to represent the rigid transformation parameter combination optimized and updated to the optimal solution in combination with reinforcement learning, and the calculation formula is: In the formula, is the output of the enhanced student self-learning function after enhanced teaching and learning, which is used to represent the optimal value of the rigid transformation parameter, S i is the i-th student parameter, used to represent the i-th rigid transformation parameter combination, rand is a random parameter, S j is related to the student parameter S i The closest rigid transformation parameter combination; In step S36, the registration accuracy optimization is specifically to train the registration accuracy optimization model through the rigid transformation problem modeling, the reinforcement learning agent initialization, the construction of the teaching learning model, the construction of the reinforcement learning model and the construction of the teaching reinforcement learning combined model to obtain the registration accuracy optimization model, and to optimize the registration accuracy by using the registration accuracy optimization model to obtain the optimal value of the rigid transformation parameter; Step S4: Adaptive intelligent registration to obtain an adaptive intelligent registered image output.

2. The adaptive UAV remote sensing image intelligent registration method according to claim 1, characterized in that: In step S1, the data collection is used to collect the original data set required for adaptive UAV remote sensing image intelligent registration; the remote sensing image registration original data set specifically includes remote sensing image data, flight parameter data and initial registration transformation information.

3. The adaptive UAV remote sensing image intelligent registration method according to claim 2 is characterized by: In step S2, the basic enhancement processing is used to perform adaptive basic enhancement on the remote sensing image, and specifically includes the following steps: step S21: image denoising; step S22: attribute enhancement; step S23: sharpening enhancement; step S24: geometric correction enhancement; step S25: adaptive enhancement evaluation; step S26: basic enhancement processing.

4. The adaptive UAV remote sensing image intelligent registration method according to claim 3 is characterized by: In step S21, the image denoising is specifically to sequentially perform noise removal operations on the remote sensing image data using median filtering and Gaussian filtering to obtain denoised and optimized remote sensing image data; In step S22, the attribute enhancement is specifically to enhance the contrast, brightness and color spectrum of the denoised and optimized remote sensing image data to obtain attribute enhanced denoised and optimized remote sensing image data; In step S23, the sharpening enhancement is specifically to use the Laplace operator to perform edge enhancement operation on the attribute enhanced denoising optimized remote sensing image data to obtain sharpened enhanced image data; In step S24, the geometric correction enhancement is specifically to perform geometric correction on the sharpened enhanced image data according to the flight parameter data in the remote sensing image registration original data set to obtain geometrically optimized corrected image data; In step S25, the adaptive enhancement evaluation is specifically to perform manual quality evaluation on each image, and adaptively perform the operations of steps S21 to S24 in sequence according to the results of the manual quality evaluation, so as to obtain adaptive enhanced image data; In step S26, the basic enhancement processing is specifically to obtain a remote sensing image registration enhanced data set through the adaptive enhancement evaluation.

5. The adaptive UAV remote sensing image intelligent registration method according to claim 4 is characterized by: In step S31, the rigid transformation problem is modeled, specifically, a registration problem model is constructed, and an objective function is set to obtain a parameter combination to be optimized; In step S32, the reinforcement learning agent is initialized, specifically, a reinforcement learning optimization agent is constructed, and a combination of agent parameters is specifically constructed, including action space parameters, state space parameters, and reward parameters; The action space parameters are used to represent the rotation and translation parameters in the state of fine image registration; The state space parameter is used to represent the combination of the action space parameter and the registration error in the state of fine image registration; The reward parameter is used to indicate that the reward parameter is increased as the registration error decreases; The calculation formula of the agent parameter combination is: A t =[a t ,s t ,r t ]; In the formula, A t is the agent parameter combination, a t is the action space parameter, s t is the state space parameter, r t is the reward parameter; In step S33, the teaching and learning model is constructed, specifically, a standard teaching and learning model is constructed as a basic model for rigid change optimization, and the rigid transformation parameters are adaptively learned and updated by constructing a teacher role function and a student self-learning function; The teacher role function is used to represent the optimal solution of the rigid transformation parameters; The student self-learning function is used to represent a rigid transformation parameter combination for optimizing and updating toward an optimal solution; In step S34, the reinforcement learning model is constructed by applying a deep Q network to update the agent parameter combination, and obtaining the optimal parameter combination output of reinforcement learning by maximizing the Q value. The calculation formula for updating the agent parameter combination is: In the formula, Q′ is the optimal parameter combination output of reinforcement learning, a t is the action space parameter, s t is the state space parameter, Q(·) is the Q value calculation function, α is the learning rate parameter, r t is the reward parameter, γ is the discount factor, max is the maximum value function, a is the optimal action space parameter, s t+1 are the state space parameters for the next iteration.

6. The adaptive UAV remote sensing image intelligent registration method according to claim 5, characterized in that: In step S4, the adaptive intelligent registration is used to perform adaptive intelligent registration based on the optimized parameter combination, specifically, to perform fine registration based on the optimal value of the rigid transformation parameter to obtain a registered image output, and to obtain matching feature data by performing accelerated robust feature extraction and cluster feature matching on the registered image output, and to eliminate abnormal registered images through an outlier elimination operation, and to obtain an adaptive intelligent registered image output through an image combination operation.

Images