A remote sensing image target detection method based on RFTZN neural dynamics model

Abstract

The application discloses a remote sensing image target detection method based on an RFTZN neural dynamics model, and belongs to the technical field of remote sensing image target detection, which comprises the following steps: based on a to-be-detected remote sensing image and a detection target feature, a target detection task of the remote sensing image is constructed as a CEM optimization problem, and the CEM optimization problem is converted into an equivalent linear matrix equation; an RFTZN neural dynamics model is constructed; based on the RFTZN neural dynamics model, a solver is constructed, and the linear matrix equation is solved to obtain a linear filter; and the linear filter is applied to the to-be-detected remote sensing image to extract a detection target. The RFTZN neural dynamics model is constructed by innovatively introducing a feedback regulation mechanism and a dynamic self-adaptive scale factor, so that the solving speed and the anti-noise interference capability of the RFTZN neural dynamics model are far superior to those of a traditional model, and the dynamic optimization problem in a target detection process can be efficiently solved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image target detection technology, specifically relating to a remote sensing image target detection method based on the RFTZN (Robust Finite-Time Zeroing Neurodynamics) neurodynamics model. Background Technology

[0002] Target detection in remote sensing images is one of the core tasks of remote sensing information processing, and the Constrained Energy Minimization (CEM) algorithm is a widely used and effective technique in this field. The CEM algorithm achieves target detection by transforming the task into a large system of linear equations and solving it. In practical applications, the number of pixels in remote sensing images is enormous, leading to complex matrix operations. The dimension is extremely high, making solving this equation a huge challenge:

[0003] 1. Direct inversion method: When the matrix dimension in the linear equation system is extremely high, the computational cost of calculating its inverse matrix is ​​extremely large and time-consuming (e.g., several minutes or even hours), which cannot meet the application requirements of real-time or near real-time.

[0004] 2. Traditional Neural Dynamics Methods (e.g., ZNN / OZN): To accelerate the solution process, researchers use dynamic solvers such as ZN (Zero Neural Networks). These methods transform the solution process into a dynamic system, approximating the solution through time evolution. OZN, or Original ZN Neural Network model, is faster than direct inversion, but its design does not consider noise resistance mechanisms, making it highly sensitive to inherent sensor noise or data transmission noise in images, resulting in insufficient solution accuracy and robustness.

[0005] 3. Noise-resistant neurodynamic methods (such as NTZN): To address the noise resistance problem of OZN, models such as NTZN (Noise-resistant Nullified Neural Network) have been proposed. They suppress noise by introducing an integral term. However, this improvement significantly reduces computational speed. Although the NTZN model is robust, its convergence time (CT) is extremely long, which also fails to meet the requirements for rapid target detection in remote sensing images. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, the remote sensing image target detection method based on the RFTZN neurodynamics model provided by this invention unifies the three indicators of solution speed, solution accuracy, and noise robustness, achieving the effects of extremely fast solution, high robustness, and high-precision detection in the remote sensing image target detection process. To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a remote sensing image target detection method based on the RFTZN neurodynamics model, comprising the following steps:

[0007] S100. Based on the features of the remote sensing image to be detected and the target to be detected, the target detection task of the remote sensing image is constructed as a CEM optimization problem and transformed into an equivalent linear matrix equation.

[0008] S200, Constructing the RFTZN neurodynamic model;

[0009] S300: Based on the RFTZN neurodynamics model, a solver is constructed, and a linear filter is obtained by solving the linear matrix equations.

[0010] S400. Apply a linear filter to the image to be remotely sensed to extract the target.

[0011] Furthermore, in S100, the CEM optimization problem of the target detection task is expressed as:

[0012]

[0013] The linear matrix equation is expressed as:

[0014]

[0015]

[0016] ,

[0017] In the formula, This represents the coefficient vector of a linear filter. This represents the response of a linear filter. This represents the autocorrelation matrix of the pixels in the remote sensing image to be detected. Indicates the features of the detected target. Represents a constant matrix. Represents a linear filter. Represents a constant vector. Represents Lagrange multipliers, superscript This indicates the transpose operation.

[0018] Furthermore, in S200, the RFTZN neurodynamic model adopts a dual-channel nonlinear feedback structure and introduces a dynamic adaptive scaling factor and a complex-domain finite-time activation function.

[0019] Furthermore, in S200, the RFTZN neurodynamic model is an error evolution dynamic equation. The error function of a linear filter, which accelerates the solution of linear matrix equations through dynamic evolution, converges to 0 from any initial value, as expressed by:

[0020]

[0021]

[0022]

[0023]

[0024] In the formula, Indicates time Evolving linear filters Corresponding error function, This represents the dynamic adaptive scaling factor. Describes the finite-time activation function in the complex field. Represents the time integral variable, express The corresponding dynamic adaptive institutional factor, express The corresponding error function, This represents the gain coefficient used to control the convergence speed of the direct error channel. This represents the gain coefficient used to control the adjustment level of the integral feedback channel. The dynamic factor parameter represents the adjustment of the slope of the inverse tangent function. This indicates the dynamic factor parameter that provides the base gain bias. Indicated by Let be the inverse cotangent function of the independent variable. This represents the magnitude scaling parameter that affects the strength of the control law. Indicates the index of the index term. Represents the exponent of the power term. Representation function The real part of the number.

[0025] Furthermore, in S300, the time... Evolving linear filters Corresponding error function derivative Combined with the RFTZN neurodynamic model, we obtain the time-varying results. Evolving linear filters derivative The dynamic equations, used as solvers for linear matrix equations, are expressed as follows:

[0026] .

[0027] Furthermore, S400 includes the following sub-steps:

[0028] S401. Obtaining a linear filter based on the solver. and extract the previous The elements are used to obtain the optimized linear filter coefficient vector. ;

[0029] S402, The optimized linear filter coefficient vector It is applied to the remote sensing image to be detected, calculates the filter response value of the spectral vector of each pixel, and then compresses the remote sensing image to be detected into a single-band energy map;

[0030] S403. Mark the pixels in the energy map whose energy values ​​are higher than the energy threshold as detection targets.

[0031] Furthermore, the energy threshold is either a fixed threshold set based on experience or a dynamic threshold determined by an adaptive threshold algorithm.

[0032] The beneficial effects of this invention are as follows:

[0033] 1. Convergence speed improved by 1-2 orders of magnitude: The RFTZN model of this invention (CT=0.12 seconds) is 13 times faster than the OZN model (CT=1.57 seconds) and 87 times faster than the NTZN model (CT=10.43 seconds), achieving millisecond-level solution speed for real-time processing of remote sensing images (such as video stream detection), which is completely impossible with existing technologies.

[0034] 2. Combining high robustness and high precision: This invention achieves strong noise resistance through a feedback adjustment mechanism (i.e., the integral term), while simultaneously achieving... It achieves the highest solution accuracy, overcoming the shortcomings of traditional target detection methods that use the NTZN model ("robust but extremely slow") and the OZN model ("fast but fragile").

[0035] 3. Achieves a balance between speed, accuracy, and robustness: This invention achieves the fastest speed, highest accuracy, and strongest robustness simultaneously through the collaborative design of dynamic scaling factors (i.e., acceleration) and feedback mechanisms (i.e., noise resistance), thus resolving the long-standing performance contradiction in traditional methods for solving CEM optimization problems. Attached Figure Description

[0036] Figure 1 The flowchart of the remote sensing image target detection method based on the RFTZN neurodynamic model provided by this invention is shown.

[0037] Figure 2 This is a diagram illustrating the target detection effect in remote sensing images provided by the present invention.

[0038] Figure 3 This is a comparison chart of the error convergence of the various models provided in this invention for solving the CEM optimization problem. Detailed Implementation

[0039] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0040] This invention provides a remote sensing image target detection method based on the RFTZN neurodynamics model. By proposing the Robust Finite-Time Zeroing Neurodynamics (RFTZN) model, and through innovative feedback adjustment mechanisms and dynamic adaptive scaling factors, it far surpasses traditional models in terms of solution speed and noise resistance, efficiently solving the dynamic optimization problem of the target detection process.

[0041] refer to Figure 1 The remote sensing image target detection method based on the RFTZN neurodynamic model includes the following steps:

[0042] S100. Based on the features of the remote sensing image to be detected and the target to be detected, the target detection task of the remote sensing image is constructed as a CEM optimization problem and transformed into an equivalent linear matrix equation.

[0043] S200, Constructing the RFTZN neurodynamic model;

[0044] S300: Based on the RFTZN neurodynamics model, a solver is constructed, and a linear filter is obtained by solving the linear matrix equations.

[0045] S400. Apply a linear filter to the image to be remotely sensed to extract the target.

[0046] In S100 of this embodiment, the CEM optimization problem of the target detection task is expressed as:

[0047]

[0048] Using the Lagrange multiplier method, it can be transformed into an equivalent linear matrix equation, which is expressed as:

[0049]

[0050]

[0051] ,

[0052] In the formula, This represents the coefficient vector of a linear filter. This represents the response of a linear filter. This represents the autocorrelation matrix of the pixels in the remote sensing image to be detected. Indicates the features of the detected target. Represents a constant matrix. Represents a linear filter. Represents a constant vector. Represents Lagrange multipliers, superscript This indicates the transpose operation.

[0053] To solve the above using neurodynamic methods Define a time Evolutionary Solution and the corresponding error function The objective of this invention is to design a dynamic system. , making Converge to 0 as quickly as possible from any initial value.

[0054] In S200 of this embodiment, the RFTZN neurodynamic model adopts a dual-channel nonlinear feedback structure and introduces a dynamic adaptive scaling factor and a complex-domain finite-time activation function; the resulting RFTZN neurodynamic model has strong robustness and finite-time convergence characteristics.

[0055] The RFTZN neurodynamic model in this embodiment is an error evolution dynamic equation. The error function of a linear filter, which accelerates the solution of linear matrix equations through dynamic evolution, converges to 0 from any initial value, as expressed by:

[0056]

[0057]

[0058]

[0059]

[0060] In the formula, Indicates time Evolving linear filters Corresponding error function, This represents the dynamic adaptive scaling factor. Describes the finite-time activation function in the complex field. Represents the time integral variable, express The corresponding dynamic adaptive institutional factor, express The corresponding error function, This represents the gain coefficient used to control the convergence speed of the direct error channel. This represents the gain coefficient used to control the adjustment level of the integral feedback channel. The dynamic factor parameter represents the adjustment of the slope of the inverse tangent function. This indicates the dynamic factor parameter that provides the base gain bias. Indicated by Let be the inverse cotangent function of the independent variable. This represents the magnitude scaling parameter that affects the strength of the control law. Indicates the index of the index term. Represents the exponent of the power term. Representation function The real part of the number.

[0061] Among them, the gain coefficient and All are positive real numbers, and These parameters are used to control the convergence speed of the direct error channel and the adjustment strength of the integral feedback channel, respectively, thus determining the overall convergence rate and feedback strength of the model; dynamic factor parameters and It is also a positive real number, among which The two work together to control the dynamic scaling factor. The amplitude and decay rate. Adjusting the slope of the inverse cotangent function, Provides the basic gain bias; in the complex domain activation function, the activation coefficients Must meet As an amplitude scaling parameter, it directly affects the strength of the control law; the exponential term is exponential. Must meet The growth rate of the control exponential part is designed to provide a significant acceleration when the error is large, while the power term exponent... It must be strictly controlled between 0 and 1, and all conditions must be met simultaneously. This is used to control the shape of the function near the origin, ensuring that the function has an infinite derivative when it approaches 0, thereby quickly eliminating residuals and achieving finite-time convergence.

[0062] In the process of constructing the RFTZN neurodynamic model mentioned above, specifically:

[0063] To overcome the noise sensitivity of traditional ZNN models, a basic framework incorporating an integral feedback term is first constructed. Unlike traditional... A dual-channel nonlinear feedback structure was introduced:

[0064]

[0065] Among them, the integral term It can accumulate historical error information. When the system is disturbed and produces a steady-state error, the integral term will continue to increase, thereby generating a reverse adjustment force to eliminate the error and ensure the robustness of the system.

[0066] Furthermore, to address the issue that the convergence speed of traditional models is limited by fixed parameters, a time-varying scaling factor is introduced into the aforementioned framework. The improved model evolved as follows:

[0067]

[0068] The specific dynamic scaling factor is defined as follows:

[0069]

[0070] This function has a "large first, small later" characteristic: hour, Larger It provides a huge initial gain, and the driving error decreases rapidly; with Increase, Smooth reduction avoids computational overflow and saves computational resources.

[0071] To address optimization problems in the complex domain and achieve finite-time convergence, a dedicated complex activation function is designed in this embodiment. To instantiate the formula above and For complex input The activation function is defined to process the real and imaginary parts separately: ; where, for the real part (Right now and The function expression is: This function incorporates an exponential term. (Provides superlinear acceleration when the error is large) and power terms (Maintain high gain as the error approaches zero to ensure convergence in finite time).

[0072] In S300 of this embodiment, the time will Evolving linear filters Corresponding error function derivative Combined with the RFTZN neurodynamic model, we obtain the time-varying results. Evolving linear filters derivative The dynamic equations, used as solvers for linear matrix equations, are expressed as follows:

[0073] .

[0074] Specifically, the derivation process of the above solver is as follows:

[0075] According to the defined error function Calculate its time derivative In the CEM detection task of a single remote sensing image, the autocorrelation matrix of the pixels in the remote sensing image to be detected... and detection target features It is a constant value calculated based on a static image, therefore the matrix sum vector For constant matrices and constant vectors, that is Differentiating both sides of the error function, we obtain a linear relationship between the rate of change of error and the rate of change of state: .

[0076] Based on the previously constructed RFTZN neurodynamic model, the expected error is... follow The nonlinear robust evolution law is used to achieve fast convergence and noise resistance.

[0077] Combining the linear relationship between the rate of change of error and the rate of change of state with the expected evolution law, we obtain the result over time... Evolving linear filters derivative The dynamic equations are: .

[0078] The above equation is a first-order ordinary differential equation (ODE) concerning a linear filter. In practical calculations, it is rearranged as follows: The form is used to facilitate numerical integration, and the final solver RFTZN-CEM is expressed as:

[0079]

[0080] S400 of this embodiment includes the following sub-steps:

[0081] S401. Obtaining a linear filter based on the solver. and extract the previous The elements are used to obtain the optimized linear filter coefficient vector. ;

[0082] Wherein, the linear filter is expressed as ,forward Each element corresponds to the number of bands in the remote sensing image to be detected.

[0083] S402. Apply the optimized linear filter coefficient vector to the remote sensing image to be detected, calculate the filter response value of the spectral vector of each pixel, and then compress the remote sensing image to be detected into a single-band energy map.

[0084] Specifically, the spectral vector of each pixel in the remote sensing image to be detected (corresponding matrix) (a column), calculate its filter response value. This step compresses a multi-band high-dimensional remote sensing image into a single-band two-dimensional grayscale image, commonly referred to as an "energy map" or "abundance map"; in this map, the target features... Areas with high matching scores will appear as high energy values ​​(bright colors), while background areas will be suppressed to low energy values ​​(dark colors).

[0085] S403. Mark the pixels in the energy map whose energy values ​​are higher than the energy threshold as detection targets;

[0086] Among them, the energy threshold is either a fixed threshold set based on experience or a dynamic threshold determined by an adaptive threshold algorithm;

[0087] Specifically, to obtain the final target location, the generated energy map undergoes thresholding. An adaptive thresholding algorithm or an empirically set energy threshold can be used to mark pixels with energy values ​​higher than the threshold as targets (corresponding to subsequent steps). Figure 2 The red box area in the image shows that areas below the energy threshold are marked as background, thus completing target detection and outputting the final result.

[0088] It should be noted that the target detection method provided by this invention can be applied not only to target detection in remote sensing images, but also widely to satellite data processing, precision agriculture, urban planning, and maritime monitoring. This invention only specifically describes its implementation process in remote sensing image target detection; the target detection process in other fields is similar to that of this invention. Specifically, this invention can be used in real-time or near-real-time analysis systems that require rapid and accurate identification of specific targets (such as aircraft, ships, vehicles, specific buildings, or crops) from complex backgrounds. For example, in precision agriculture, it can be used to locate pest and disease areas; in maritime monitoring, it can be used to quickly detect ships on a vast sea surface.

[0089] In this embodiment of the invention, a simulation verification example of the above-described remote sensing image target detection method is provided.

[0090] In this embodiment, the RFTZN-CEM method for constructing the solver was simulated and verified, and compared with four existing technologies, namely OZN, OGN, NTGON, and NTZN.

[0091] like Figure 2This paper demonstrates the detection performance of the method of this invention on four different sets of remote sensing images, specifically the original RGB images of (a) Port A, (b) Park B, (c) Airport C, and (d) Airport D. The middle row (e) to (h) shows the "energy map" obtained after applying the linear filter solved using the RFTZN neurodynamic model of this invention to the original image. It is clearly visible in the image that the targets (ships, sports fields, information signs) are greatly enhanced and displayed as bright white, while the background is effectively suppressed and displayed as black. The bottom row (i) to (l) shows the final detection results obtained after thresholding the energy map; all targets are accurately marked with red boxes. This set of images strongly demonstrates that the RFTZN method of this invention can effectively and accurately extract targets from complex backgrounds.

[0092] like Figure 3 As shown, this demonstrates the effectiveness of different species models (RFTZN, OGN, NTGON, OZN, NTZN) in solving the CEM equation. At that time, its error Over time The curve changing in seconds, represented by the solid red line of the RFTZN invention, shows the changes over time. The time error is the largest, but it converges rapidly within 0.12 seconds. The following extremely low levels demonstrate the fastest convergence speed and the highest solution accuracy. In contrast, the light blue curve representing OZN converges in approximately 1.57 seconds; the tan dashed line representing NTZN converges extremely slowly, with a still high error within the 5-second simulation time; while the green and purple curves representing OGN and NTGON, although initially decreasing rapidly, ultimately fail to converge to high accuracy, with steady-state errors remaining at a certain level. The lowest level of accuracy. This figure visually demonstrates that the RFTZN model of this invention achieves both the fastest convergence speed and the highest solution accuracy when solving the CEM problem.

[0093] Furthermore, as shown in Table 1, it quantifies in detail... Figure 3 The comparison results show that ASSRE (mean steady-state error) and MSSRE (maximum steady-state error) are used to measure accuracy, while CT (convergence time) is used to measure speed. Data shows that the convergence time CT of the RFTZN model in this invention is only 0.12 seconds, and the ASSRE reaches... In comparison, OZN's CT is 1.57 seconds (13 times slower than this invention), and ASSRE is... (About 8 times lower accuracy); NTZN's CT scan takes 10.43 seconds (87 times slower than this invention), and ASSRE is... OGN and NTGON's ASSRE is as high as The accuracy and level of the solution are completely substandard. These data clearly show that the RFTZN model of this invention surpasses all existing technical solutions in both key indicators of solution speed and solution accuracy.

[0094] Table 1: Comparison Results of Different Models

[0095]

[0096] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0097] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A remote sensing image target detection method based on the RFTZN neurodynamic model, characterized in that, Includes the following steps: S100. Based on the features of the remote sensing image to be detected and the target to be detected, the target detection task of the remote sensing image is constructed as a CEM optimization problem and transformed into an equivalent linear matrix equation. S200, Constructing the RFTZN neurodynamic model; S300: Based on the RFTZN neurodynamics model, a solver is constructed, and a linear filter is obtained by solving the linear matrix equations. S400. Apply a linear filter to the image to be remotely sensed to extract the target. In S200, the RFTZN neurodynamic model is an error evolution dynamic equation. The error function of a linear filter, which accelerates the solution of linear matrix equations through dynamic evolution, converges to 0 from any initial value, as expressed by: In the formula, Indicates time Evolving linear filters Corresponding error function, This represents the dynamic adaptive scaling factor. Describes the finite-time activation function in the complex field. Represents the time integral variable, express The corresponding dynamic adaptive institutional factor, express The corresponding error function, This represents the gain coefficient used to control the convergence speed of the direct error channel. This represents the gain coefficient used to control the adjustment level of the integral feedback channel. The dynamic factor parameter represents the adjustment of the slope of the inverse tangent function. This indicates the dynamic factor parameter that provides the base gain bias. Indicated by Let be the inverse cotangent function of the independent variable. This represents the magnitude scaling parameter that affects the strength of the control law. Indicates the index of the index term. Represents the exponent of the power term. Representation function The real part of the number; In S300, the time will Evolving linear filters Corresponding error function derivative Combined with the RFTZN neurodynamic model, we obtain the time-varying results. Evolving linear filters derivative The dynamic equations, used as solvers for linear matrix equations, are expressed as follows: 。 2. The remote sensing image target detection method based on the RFTZN neurodynamic model according to claim 1, characterized in that, In S100, the CEM optimization problem for the target detection task is expressed as: The linear matrix equation is expressed as: ， In the formula, This represents the coefficient vector of a linear filter. This represents the response of a linear filter. This represents the autocorrelation matrix of the pixels in the remote sensing image to be detected. Indicates the features of the detected target. Represents a constant matrix. Represents a linear filter. Represents a constant vector. Represents Lagrange multipliers, superscript This indicates the transpose operation.

3. The remote sensing image target detection method based on the RFTZN neurodynamic model according to claim 1, characterized in that, In S200, the RFTZN neurodynamic model adopts a dual-channel nonlinear feedback structure and introduces a dynamic adaptive scaling factor and a complex-domain finite-time activation function.

4. The remote sensing image target detection method based on the RFTZN neurodynamic model according to claim 1, characterized in that, S400 includes the following steps: S401. Obtaining a linear filter based on the solver. and extract the previous The optimized linear filter coefficient vector is obtained by considering the elements. ; S402, The optimized linear filter coefficient vector It is applied to the remote sensing image to be detected, calculates the filter response value of the spectral vector of each pixel, and then compresses the remote sensing image to be detected into a single-band energy map; S403. Mark the pixels in the energy map whose energy values ​​are higher than the energy threshold as detection targets.

5. The remote sensing image target detection method based on the RFTZN neurodynamic model according to claim 4, characterized in that, The energy threshold is either a fixed threshold set based on experience or a dynamic threshold determined by an adaptive threshold algorithm.

Images