Thermal infrared image driven microwave de-icing region segmentation method and system

Abstract

This invention relates to the field of image processing technology, specifically to a method and system for segmenting microwave de-icing areas driven by thermal infrared images. The method involves emitting a preset short-duration microwave pulse to the area to be de-iced, while simultaneously acquiring a thermal infrared image sequence covering the entire cycle of active temperature rise and passive thermal relaxation. The thermal relaxation time constant of each pixel on the time axis is then extracted to construct a feature matrix, which is further mapped to a pseudo-color image of thickness distribution. This matrix is ​​then logically fused with the transient temperature rise rate during the active temperature rise phase at the pixel level. This process determines a segmentation mask containing graded labels for the substrate area, thin ice area, and thick ice area, ensuring the scientific rigor and logical soundness of de-icing boundary identification. This effectively eliminates artifact interference introduced by the high reflectivity of the substrate material, significantly reducing the risk of accidental substrate damage or redundant energy consumption during graded de-icing operations, and improving the safety level and equipment lifespan of the operation.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a method and system for segmenting microwave de-icing regions driven by thermal infrared images. Background Technology

[0002] In the field of aircraft safety operations, icing is one of the major environmental risks threatening flight safety. The asymmetric accumulation of ice on critical aerodynamic surfaces such as wings and control surfaces directly damages the aircraft's shape, drastically increases weight, and disrupts laminar flow, ultimately leading to lift loss and reduced control effectiveness, posing a serious threat to flight safety. Meanwhile, microwave de-icing, as an active protection technology, uses electromagnetic energy to heat the ice layer, causing the ice-substrate interface to melt, providing a key technological approach for efficient and low-energy online de-icing.

[0003] It is worth noting that the depth of microwave interaction and the thermal effect on ice layers are strongly dependent on the ice thickness. Thinner ice layers heat up rapidly after absorbing microwave energy, while excessive energy may penetrate the ice and damage sensitive composite substrates; conversely, thicker ice regions require higher energy densities to achieve effective melting. This makes spatial perception of ice thickness crucial for power control in microwave de-icing. Especially for mixed ice formations of varying thicknesses that form on aircraft surfaces under complex weather conditions, accurate mapping of their thickness distribution directly determines the energy distribution strategy of the microwave array and the final de-icing efficiency.

[0004] However, existing infrared imaging-based de-icing monitoring technologies have significant limitations. Traditional single-frame infrared thermal images can only reflect the instantaneous temperature field of the surface and cannot distinguish the superficial similarities between "rapid heating of thin ice" and "low temperature in thick ice areas due to insufficient microwave reflection or absorption," essentially losing the temporal dimension of the heat conduction process. This results in the system only being able to roughly partition the ice based on preset, static temperature thresholds, making it difficult to cope with the continuously changing thermophysical characteristics of actual mixed ice layers in space. Consequently, microwave power allocation is often inaccurate: thin ice areas face the risk of substrate overheating damage due to energy overload, while thick ice areas suffer from incomplete de-icing due to insufficient energy. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides a method and system for segmenting microwave de-icing regions driven by thermal infrared images, in order to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention provides a method for segmenting microwave de-icing regions driven by thermal infrared images, the method comprising the following steps:

[0008] A preset short-duration microwave pulse is emitted to the area to be de-iced, and a continuous thermal infrared image sequence of the microwave-affected area in the time dimension is acquired simultaneously; the thermal infrared image sequence covers the active temperature rise stage during microwave heating and the passive thermal relaxation stage after microwave unloading.

[0009] For each pixel in the thermal infrared image sequence, a fitting operation is performed based on the temperature decay trend during the passive thermal relaxation stage to extract the thermal relaxation time constant; the thermal relaxation time constant is then used to construct a thermal relaxation feature matrix.

[0010] The thermal relaxation feature matrix is ​​mapped to a thickness distribution pseudo-color image. At the same time, the transient temperature rise rate during the active temperature rise stage is calculated and logically fused with the thickness distribution pseudo-color image at the pixel level. Based on the gradient difference of the fused thickness distribution pseudo-color image, a hierarchical segmentation mask containing classification labels for the base region, thin ice region, and thick ice region is output.

[0011] Another aspect of the present invention provides a thermal infrared image-driven microwave de-icing region segmentation system, comprising:

[0012] The infrared image acquisition module emits a preset short-time microwave pulse excitation to the area to be de-iced and simultaneously acquires a continuous thermal infrared image sequence of the microwave-affected area in the time dimension; the thermal infrared image sequence covers the active temperature rise stage during microwave heating and the passive thermal relaxation stage after microwave unloading.

[0013] The feature matrix construction module performs fitting calculations on each pixel in the thermal infrared image sequence based on the temperature decay trend during the passive thermal relaxation stage, extracts the thermal relaxation time constant, and constructs the thermal relaxation feature matrix using the thermal relaxation time constant.

[0014] The de-icing region classification module maps the thermal relaxation feature matrix into a thickness distribution pseudo-color image, and calculates the transient temperature rise rate during the active temperature rise stage. It then performs pixel-level logical fusion with the thickness distribution pseudo-color image. Based on the gradient difference of the fused thickness distribution pseudo-color image, it outputs a hierarchical segmentation mask containing classification labels for the base region, thin ice region, and thick ice region.

[0015] As described above, the thermal infrared image-driven microwave de-icing region segmentation method and system provided by the present invention have at least the following beneficial effects:

[0016] 1. The thermal infrared image-driven microwave de-icing area segmentation method and system provided by this invention, by emitting a preset short-duration microwave pulse excitation to the area to be de-iced, and simultaneously acquiring a continuous thermal infrared image sequence covering the entire cycle of active temperature rise and passive thermal relaxation, extracts the thermal relaxation time constant of each pixel on the time axis to construct a feature matrix. This effectively solves the technical bottleneck of traditional monitoring methods that cannot distinguish between uneven energy radiation and differences in physical thickness. It greatly improves the sensitivity of identifying hidden ice layers and avoids the risk of misjudgment caused by microwave standing waves or power hotspots from a physical perspective, thereby significantly improving the system's adaptability to complex operating environments. While ensuring detection accuracy, it provides a high-confidence physical basis for subsequent refined on-demand de-icing operations.

[0017] 2. Based on the obtained thermal relaxation feature matrix, this invention further maps it into a thickness distribution pseudo-color image and performs pixel-level logical fusion with the transient temperature rise rate during the active temperature rise stage. This determines a segmentation mask containing graded labels for the substrate region, thin ice region, and thick ice region, ensuring the scientific rigor and logical soundness of de-icing boundary identification. It effectively eliminates artifact interference introduced by the high reflectivity of the substrate material, significantly reducing the risk of accidental substrate damage or redundant energy consumption during graded de-icing operations, and improving the safety level and equipment lifespan of the operation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.

[0020] Figure 2 This is a schematic diagram of the logic connection for solving the surface heat dissipation interference component in this invention.

[0021] Figure 3 This is a schematic diagram illustrating the logical connection for determining the optimal estimate of the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant in this invention.

[0022] Figure 4 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0023] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0024] In traditional de-icing monitoring systems, fixed calibration curves and static mapping rules cannot adapt to variations in ice types, ambient temperatures, and microwave heating conditions. When impurities or uneven thickness distribution exist within the ice layer, the system cannot establish a dynamic correlation between the thermal relaxation time constant and the ice thickness, leading to a mismatch between the characteristic mapping relationship and the actual physical state. This static mapping mechanism reduces the accuracy of thickness inversion, resulting in pseudo-color images of the thickness distribution containing significant errors, ultimately affecting the reliability of de-icing decisions.

[0025] For example, in low-temperature environments, the thermal conductivity of ice changes, causing the thermal relaxation time constant of ice layers of the same thickness to drift compared to the standard calibration curve. If the system still uses a preset fixed calibration curve to map the thermal relaxation feature matrix into a pseudo-color image of the thickness distribution, it may estimate an ice layer with an actual thickness of 5 mm as 3 mm, and misjudge a 2 mm thin ice layer as ice-free. This misjudgment will lead to improper allocation of de-icing resources, with thin ice areas being ignored and thick ice areas being overheated.

[0026] If the above problems are not addressed, errors in thickness estimation will lead to a misalignment between the de-icing plan and the actual situation, resulting in energy waste or incomplete de-icing. Rigid feature mapping relationships will hinder the system from capturing rapid spatial changes in ice thickness, delaying the identification of dangerous thin ice areas. Static processing of multimodal data (thermal relaxation and temperature rise rates) will also cause phase deviations between thickness estimation and real-time heating effects, reducing the reliability of thickness segmentation and classification, ultimately creating a negative feedback loop that affects de-icing efficiency and safety.

[0027] To address the aforementioned issues, this application first considers establishing a dynamic correlation mechanism between the thermal relaxation time constant and ice thickness. Traditional systems use fixed calibration curves for thickness mapping, leading to misjudgments of key thickness features and failing to reflect the actual state of the ice layer. To resolve this, this application attempts to dynamically couple the transient temperature rise rate with the thickness mapping, adjusting the thickness inversion model in real time by adjusting the coefficients. Further analysis reveals that relying solely on the thermal relaxation time constant or static mapping relationship is insufficient to capture the spatial non-uniformity of the ice layer's distribution. It is necessary to introduce temperature rise rate data from the active heating phase to collaboratively calculate the thickness estimate and generate dynamic calibration parameters based on its changing trends. By designing a linkage mechanism between the adjustment coefficient mapping module and the thickness inversion module, the thickness estimate adaptively adjusts with the actual state of the ice layer, thereby resolving the mapping lag and feature mismatch problems.

[0028] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] Example 1:

[0030] Please see Figure 1-3 As shown, a method for segmenting microwave de-icing regions driven by thermal infrared images includes the following steps:

[0031] A preset short-duration microwave pulse is emitted to the area to be de-iced, and a continuous thermal infrared image sequence of the microwave-affected area is acquired in the time dimension simultaneously; the thermal infrared image sequence covers the active temperature rise stage during microwave heating and the passive thermal relaxation stage after microwave unloading.

[0032] Furthermore, a preset short-duration microwave pulse is emitted to the area to be de-iced, and a continuous sequence of thermal infrared images of the microwave-affected area over time is simultaneously acquired, including:

[0033] A synchronous trigger command is generated based on a preset sampling frequency, and the transmitting end of the microwave array and the acquisition end of the thermal infrared imaging device are activated simultaneously using the synchronous trigger command.

[0034] Within a first preset duration of maintaining power output by the microwave array, continuous frame images of the microwave action area are acquired in real time, an active temperature rise image sequence is generated, and the end time of the first preset duration is marked as the heat source cutoff point.

[0035] Starting from the heat source cutoff point, within the second preset time period after the microwave array power output is stopped, continuous frame images of the microwave action area are continuously acquired to generate a passive thermal relaxation image sequence.

[0036] Based on the time reference determined by the synchronous trigger command, the active temperature rise image sequence and the passive thermal relaxation image sequence are spliced ​​in time and spatially registered to construct a continuous thermal infrared image sequence of the microwave action area in the time dimension.

[0037] It should be added that, in the process of temporally stitching and spatially registering the active temperature rise image sequence and the passive thermal relaxation image sequence, to address the perspective distortion that may exist due to the infrared camera's installation angle, a homography matrix is ​​used for spatial pixel registration to construct a continuous thermal infrared image sequence of the microwave-affected area in the time dimension. The registration calculation formula is as follows: , where (u,v) are the pixel coordinates of the original infrared image in pixels; (x',y') are the corrected physical plane projection coordinates in millimeters; H is a 3×3 homography transformation matrix, obtained by solving the four corner points of the pre-calibrated de-icing area.

[0038] For each pixel in the thermal infrared image sequence, a fitting operation is performed based on the temperature decay trend during the passive thermal relaxation stage to extract the thermal relaxation time constant; the thermal relaxation feature matrix is ​​constructed using the thermal relaxation time constant, including:

[0039] Segmenting passive thermal relaxation image sequences from thermal infrared image sequences;

[0040] Pixel-level temporal extraction is performed on the passive thermal relaxation image sequence. Each pixel in the image plane is traversed, and the temperature values ​​of the corresponding multi-frames on the time axis are extracted to form a single-point transient temperature decay curve.

[0041] The single-point transient temperature decay curve is mapped to convert the original temperature value into a natural logarithm value, and the logarithmic difference between adjacent time steps is calculated to generate a logarithmic difference decay sequence.

[0042] The logarithmic difference decay sequence and the standard exponential decay theoretical model are approximated by a preset iterative algorithm to obtain the thermal relaxation time constant of the pixel.

[0043] Based on the planar coordinates of each pixel, the thermal relaxation time constant calculated for each pixel is filled into the corresponding two-dimensional array space to generate a thermal relaxation feature matrix.

[0044] The working principle of this application is as follows: Utilizing the exponential decay characteristic of the thermal diffusion equation, the temperature-time relationship is decoupled into a pure decay rate characteristic independent of the initial energy input through a nonlinear logarithmic transformation. By constructing a logarithmic difference sequence, not only is the complex nonlinear heat flow process linearized, but the amplitude background noise introduced by uneven microwave heating power distribution or differences in emissivity on the ice surface is also effectively eliminated. The finally extracted thermal relaxation time constant depends only on the thermal resistance and thermal capacity characteristics of the microwave-treated region (i.e., material thickness and structure), thus achieving a precise mapping from temperature data to the intrinsic properties of the material.

[0045] Furthermore, a pre-defined iterative algorithm is used to approximate the logarithmic difference decay sequence with the standard exponential decay theoretical model, including:

[0046] For the logarithmic difference decay sequence, the discrete curvature values ​​of each sampling point are calculated using the central difference method; at the same time, local maxima are retrieved in the discrete curvature value sequence, and the sampling time of the sampling point corresponding to the local maxima is identified as the critical time frame; the logarithmic difference decay sequence is divided into transient dominant segments using the critical time frame as the time axis division boundary.

[0047] The single exponential model with the first preset weight is prefitted to the transient dominant section, the surface heat dissipation interference component is calculated, and the surface heat dissipation interference component is subtracted from the original logarithmic difference decay sequence to generate the residual decay sequence.

[0048] The residual decay sequence is imported into the hyperbolic tangent nonlinear regression model for a second-order fine-tuning iterative approximation. The tail decay rate when the residual decay sequence approaches zero is locked, and the reciprocal of the tail decay rate is defined as the thermal relaxation time constant.

[0049] Specifically, for the acquired log-difference decay sequence We perform discretized second-order curvature analysis and calculate the curvature value at each point in the sequence using the central difference method. The calculation formula is: ,in , and The first , and The logarithmic temperature difference at time t, in units of △t represents the sampling time interval in seconds; by iterating through the calculation results, the following is identified: The moment when a local maximum is reached is marked as the critical time frame where the decay rate undergoes a sudden inflection point. This inflection point physically corresponds to the phase transition boundary where heat flow changes from two-dimensional surface diffusion to one-dimensional longitudinal conduction, thereby rigidly dividing the sequence along the time axis. The transient dominant region characterized by rapid heat dissipation on the surface, and The non-transient dominant region characterized by internal thermal capacity hysteresis.

[0050] Simultaneously, a single-exponential benchmark function is invoked to extrapolate the data characteristics of the transient dominant segment, solving for the surface heat dissipation interference component present across the entire time axis. Then, a vector subtraction operation is performed, that is, the surface heat dissipation interference component is subtracted point by point from the original logarithmic difference decay sequence to generate a residual decay sequence. .

[0051] Finally, the residual decay sequence The data is imported into a hyperbolic tangent nonlinear regression model for secondary refined iterative approximation. This model aims to fit the process by which the residual signal eventually tends towards thermal equilibrium through the saturation characteristics of the hyperbolic tangent function. The model formula is defined as follows: , where R(t) represents the mathematical model prediction of the remaining residual signal after the initial exponential decay fitting; is the equivalent initial amplitude of the residual signal, and is the dimensionless logarithm; This is the phase delay parameter, in seconds; The tail decay rate factor to be identified, in reciprocals of seconds, characterizes how quickly the residuals approach zero; the mean square error between the observed residual sequence and the model predictions is locked after iterative convergence. The optimal value is determined, and the reciprocal of the tail decay rate is defined as the thermal relaxation time constant. The unit is seconds.

[0052] Furthermore, the solution logic for the surface heat dissipation interference component is as follows:

[0053] Based on the critical time frame, early data points located before the critical time frame are extracted from the logarithmic difference decay sequence to form a transient dominant data subset;

[0054] A single exponential benchmark function is constructed, and the nonlinear least squares method is used to fit the single exponential benchmark function to the transient dominant data subset. During the fitting process, the data points are assigned a first preset weight that decreases with time, thereby determining the best estimated values ​​of the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant.

[0055] Using the determined optimal estimate, the single exponential baseline function is extended to the time axis of the entire passive thermal relaxation stage. The theoretical value of the function at each moment is calculated, and a simulation curve is generated. This simulation curve is defined as the surface heat dissipation disturbance component.

[0056] Specifically, based on critical time frames The entire passive thermal relaxation sequence is divided into time domain segments, and time intervals are extracted. The sampled data within constitutes the transient dominant data subset. ,in Indicates a point in time The surface temperature value collected at that location, The determination is based on the zero point of the second derivative of the logarithmic decay curve. The data in this interval are mainly dominated by the rapid convection heat transfer between the ice surface and the air, and have not yet been significantly affected by the heat capacity.

[0057] Constructing a single-exponential benchmark function As the fitting kernel, where This is the surface heat dissipation amplitude coefficient; The thermal decay time constant is characterized; a weighted objective function is constructed using the nonlinear least squares method; the objective function J is minimized through an iterative algorithm until the parameters converge, thereby determining the optimal estimates of the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant. These optimal estimates are then substituted back into the single-exponential reference function, which is then extended to the time axis of the entire passive thermal relaxation stage. The theoretical value of this function at each time step is calculated. This generates a simulated curve.

[0058] It should be added that the single-exponential benchmark function originates from the first-order thermal response model based on Newton's law of cooling, and its physical basis is as follows:

[0059] When heat dissipation from an object's surface is dominated by convection, the rate of temperature change is directly proportional to the temperature difference between the object and its environment:

[0060] ;

[0061] Where k is a constant related to the convective heat transfer coefficient and heat capacity.

[0062] Solving this differential equation yields the single exponential decay form:

[0063] ;

[0064] Let A = τ=1 / k (heat dissipation time constant), then the reference function is obtained. ;

[0065] The initial values ​​are set based on the physical characteristics of the transient dominant subset of data:

[0066] 1. Initial value of amplitude coefficient A:

[0067] ;

[0068] Where T (t=0) is the measured temperature at the instant the heat source is cut off. The ambient temperature is known or the stable temperature at the end of the relaxation period.

[0069] Initial value of the heat dissipation time constant τ:

[0070] Based on experience, the heat dissipation time constant τ during the early convection-dominated phase is typically in the range of 0.1s to 0.3s.

[0071] Set the objective function F(A,τ)= The objective function is to minimize the weighted sum of squared biases between the model predictions and the experimental data; weights The function is set to an exponential function that decreases over time to enhance the contribution of early data. Using the LM method, the partial derivatives of the parameters A and τ in the baseline function are calculated to construct the normal equation. The parameters are updated iteratively with damping until the convergence condition is met, where F(A,τ) is the objective function value to be minimized.

[0072] Furthermore, determining the optimal estimates of the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant includes:

[0073] Based on the transient dominant data subset and the first preset weights, the weighted residual between the actual observed logarithmic temperature value at each moment and the theoretical predicted value of the single exponential benchmark function under the current parameter settings is calculated, and the weighted residuals at all moments are squared and summed to construct a global objective function that characterizes the degree of fitting error.

[0074] The global objective function is subjected to partial differential operations with respect to the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant to generate a sensitivity gradient vector. The parameter update step size is dynamically calculated based on the magnitude of the sensitivity gradient vector. When the magnitude of the sensitivity gradient vector is greater than a preset threshold, the upper limit of the parameter update step size is matched. When the magnitude of the sensitivity gradient vector is less than the preset threshold, the lower limit of the parameter update step size is matched.

[0075] Using the sensitivity gradient vector and parameter update step size, the surface heat dissipation amplitude coefficient and surface heat dissipation time constant are iteratively corrected in the parameter space. After each correction, the value of the global objective function is recalculated until the rate of change of the global objective function value obtained by two adjacent iterations is lower than the preset convergence accuracy threshold. The parameter values ​​at this convergence state are locked as the best estimates of the surface heat dissipation amplitude coefficient and surface heat dissipation time constant.

[0076] Specifically, based on transient dominant data subsets ,in For a point in time, Given the logarithmically converted temperature rise data, and M representing the total number of data points, a global objective function is constructed to characterize the degree of fitting error. This function introduces weights To enhance the data contribution during the early, rapid decay phase, the calculation formula is as follows: ,in Let be the surface heat dissipation amplitude coefficient to be determined, in Kelvin; Let be the surface heat dissipation time constant to be determined, in seconds; As the first preset weight, an inverse decay strategy is adopted and set as follows. To offset the uncertainty caused by the decrease in signal-to-noise ratio of later data points, and to ensure that the center of gravity of the fitting is locked in the initial stage of intense energy exchange, δ is a small constant to prevent the denominator from being zero. The logic of this weight setting is: the closer the data point is to t=0, the higher its signal-to-noise ratio and the less it is affected by internal heat conduction interference, so it is given a higher weight.

[0077] Next, regarding the global objective function Perform partial differential operations with respect to the two undetermined parameters to construct the sensitivity gradient vector. To improve convergence stability, an improved strategy using the Gauss-Newton method is employed to dynamically calculate the parameter update step size ΔP. The calculation formula is as follows: ,in This is an approximation of the Hessian matrix. It is the identity matrix. As a damping factor, the algorithm automatically reduces it when the magnitude of the sensitivity gradient vector is large (the error decrease is greater than the threshold). To accelerate the approximation, the step size is made closer to the Gaussian-Newton step size; when the modulus is small (close to the extreme point or on flat terrain), it is increased. To suppress oscillations, the step size should be oriented towards the gradient descent direction. The initial value is recommended to be set between 0.01 and 0.1.

[0078] Finally, using the calculated parameters, the step size ΔP is updated, and iterative corrections are performed in the parameter space. The update formula is as follows: ,in Let be the parameter vector for the k-th iteration; after each correction, the global objective function is recalculated. The new value is determined, and the rate of change between two adjacent iterations is judged. Whether the value is below the preset convergence accuracy threshold, once this condition is met, the algorithm is considered to have found the global optimum, and the parameter value in the convergence state is locked as the surface heat dissipation amplitude coefficient. With surface heat dissipation time constant The best estimate.

[0079] The thermal relaxation feature matrix is ​​mapped to a thickness distribution pseudo-color image, and the transient temperature rise rate during the active temperature rise phase is calculated. This rate is then logically fused with the thickness distribution pseudo-color image at the pixel level. Based on the gradient differences in the fused thickness distribution pseudo-color image, a hierarchical segmentation mask containing classification labels for the base region, thin ice region, and thick ice region is output, specifically including:

[0080] The thermal relaxation feature matrix is ​​retrieved, and the thermal relaxation time constant in the matrix is ​​numerically mapped using a preset thickness-time constant nonlinear calibration curve. The relaxation parameters are converted into ice layer thickness estimates, and the ice layer thickness estimates are projected onto a preset single-channel grayscale space to generate a pseudo-color image of the thickness distribution.

[0081] The active temperature rise image sequence during microwave heating is obtained from the thermal infrared image sequence. The temperature data of the first and last frames of the sequence are extracted and the difference is calculated. The difference is divided by the duration of microwave heating to obtain the transient temperature rise rate matrix.

[0082] Pixel-level fusion rules are constructed, and the thickness distribution pseudo-color image is corrected by confidence weighting using the transient temperature rise rate matrix, thereby generating the fused thickness distribution pseudo-color image.

[0083] The Laplacian gradient operation is performed on the fused thickness distribution pseudo-color image to extract edge texture features of different thickness regions. The calculated gray values ​​are then compared stepwise with preset base determination thresholds and preset thick ice determination thresholds. Regions below the base determination threshold are marked as base regions, regions between the base determination threshold and the thick ice determination threshold are marked as thin ice regions, and regions above the thick ice determination threshold are marked as thick ice regions. Finally, a hierarchical segmentation mask carrying classification labels for the base region, the thin ice region, and the thick ice region is output.

[0084] The core logic of this application's design is as follows: Utilizing the thermal relaxation time constant during the cooling period, which is essentially a measure of the material's thermal inertia, it exhibits a robust nonlinear relationship with the ice layer thickness. Generally, the greater the thickness, the slower the heat dissipation, and the longer the time constant. This time constant is mapped to a pseudo-color map of the thickness distribution using a pre-calibrated curve, providing a preliminary spatial estimate of the thickness. However, a single feature is susceptible to surface contamination, micropores, or noise interference. Therefore, the transient temperature rise rate during the heating period is extracted simultaneously. Under the same heating power, the thinner the ice layer, the smaller the heat capacity, and the faster the temperature rise. The temperature rise rate matrix reflects the agility of the thermal response in different regions.

[0085] The thickness distribution pseudocolor map is corrected for reliability using a temperature rise rate matrix. The physical principle is that if the temperature rise rate of a region closely matches the theoretical response that its predicted thickness should have (for example, a region predicted as thin ice actually shows a rapid temperature rise), then the reliability of the thickness estimate is high; conversely, if there is a significant deviation (such as a region predicted as thick ice but with a rapid temperature rise), then the thickness estimate at that point may be unreliable, and its weight should be reduced.

[0086] After obtaining a more reliable thickness estimate, direct thresholding can be performed, but the boundaries may be coarse. Therefore, a Laplacian gradient operation is executed. The principle behind this operation is that abrupt changes in ice thickness (such as the edge from thin ice to thick ice) create grayscale steps in the thickness map, and the Laplacian operator can keenly enhance the texture features of these edges. This allows subsequent thresholding to be based not only on absolute thickness values ​​but also on enhanced perception of boundary locations, resulting in more accurate and clearer partitions.

[0087] The generation logic of the thickness distribution pseudo-color image specifically includes:

[0088] The thermal relaxation feature matrix is ​​retrieved, and each thermal relaxation time constant in the matrix is ​​numerically mapped using a preset thickness-time constant nonlinear calibration curve. The relaxation parameters are converted into ice layer thickness estimates, which are then projected onto a preset single-channel grayscale space to generate a pseudo-color image of the thickness distribution, including:

[0089] Traverse each pixel coordinate in the thermal relaxation feature matrix, obtain the thermal relaxation time constant corresponding to the current coordinate, and substitute it as an independent variable into the pre-constructed thickness-time constant nonlinear calibration curve model. This model is fitted by multinomial regression of experimental data of multiple sets of standard thickness ice samples. The physical thickness value corresponding to the pixel coordinate is output through the mapping calculation of the model, thereby constructing the physical matrix of ice layer thickness.

[0090] The maximum and minimum thickness values ​​in the physical matrix of ice layer thickness are statistically analyzed, and the difference between the two is calculated to determine the global dynamic range of ice layer thickness. For the physical thickness value of each pixel coordinate, the relative proportion of the difference between it and the minimum thickness value in the global dynamic range is calculated to generate a normalized relative thickness coefficient.

[0091] The thickness relative coefficient is multiplied by the preset maximum bit depth of the single-channel image, and the result is rounded down. The resulting integer value is then used as the pixel gray value and filled into the corresponding coordinate position in the preset single-channel gray space to generate a pseudo-color image with thickness distribution.

[0092] Specifically, the pre-defined logic of the thickness-time constant nonlinear calibration curve model is based on laboratory calibration experiments: a set of standard ice samples with gradient thickness differences are selected, and thermal relaxation tests are conducted under controlled ambient temperature and standard microwave excitation power. The thermal relaxation time constant corresponding to each standard thickness is calculated. Considering that the propagation of heat waves in the medium follows the diffusion equation, its characteristic time is often proportional to the square of the thickness. Therefore, a quadratic polynomial regression algorithm is used to evaluate the experimental data. The calibration model formula was determined by fitting the model. ,in This is a predicted value for ice thickness, in millimeters. is the thermal relaxation time constant, in seconds; , , These are the coefficients of the quadratic term (unit: ). ), linear term coefficient (unit: mm / s) and intercept constant (unit: mm).

[0093] It should be added that, Indicates the thickness of the standard parts in the Arctic Ocean. This indicates that for a known thickness of Thermal relaxation tests were conducted on standard parts from the Arctic Ocean, and the actual calculated thermal relaxation time constant was obtained.

[0094] The logic for determining the calibration model formula is as follows: several sets of experimental data were obtained through experiments. According to physical principles, in the thermal diffusion equation, the characteristic time is proportional to the square of the diffusion length. Therefore, a quadratic polynomial is pre-selected. As the mathematical form of the calibration model, d represents the predicted thickness. The input is the time constant; several sets of experimental data are imported into a quadratic polynomial, and the least squares method is used to iteratively find the optimal coefficients. , , .

[0095] During the real-time processing phase, the system retrieves the thermal relaxation feature matrix and iterates through the thermal relaxation time constant at each pixel coordinate (i,j) in the matrix. Substitute this value into the aforementioned preset nonlinear calibration model to calculate the corresponding physical thickness value. This allows us to construct the physical matrix of ice thickness. Next, we traverse the physical matrix to extract the maximum and minimum thickness values ​​across the entire region. Calculate the global dynamic range of the ice layer thickness, i.e., the range between the maximum and minimum thickness values; then, calculate the physical thickness for each pixel. Using the formula Calculate its grayscale value, where N is the preset bit depth of a single-channel image. This indicates a floor operation, which ultimately rounds down the grayscale value of each pixel. Enter the corresponding coordinates to generate a single-channel grayscale image. You can further use a lookup table to apply a preset pseudo-color spectrum to map the grayscale values ​​to RGB colors, thereby outputting a thickness distribution pseudo-color image.

[0096] The method involves using the transient temperature rise rate matrix to perform confidence-weighted correction on the pseudo-color image of the thickness distribution, including:

[0097] The transient temperature rise rate matrix is ​​numerically normalized to generate a standardized temperature rise rate feature matrix.

[0098] Traverse each pixel in the thickness distribution pseudo-color image, and look up the corresponding theoretical temperature rise rate reference value in the preset thickness-temperature rise inverse relationship table according to the gray value of the pixel. Calculate the absolute value of the deviation between the actual temperature rise rate value and the theoretical temperature rise rate reference value under the same coordinate in the standardized temperature rise rate feature matrix, and define the absolute value of the deviation as the physical conflict index.

[0099] The physical conflict index is substituted into the negative exponential decay function for calculation to generate the confidence weight coefficient. The confidence weight coefficient is then used to perform multiplication correction on the original grayscale values ​​of the corresponding pixels in the thickness distribution pseudocolor image, and finally the fused thickness distribution pseudocolor image is output.

[0100] Specifically, the transient temperature rise rate matrix is ​​numerically normalized, and the normalization calculation formula is as follows: ,in Let be the initial temperature rise rate at coordinate (i,j) in the matrix. and These are the global minimum and global maximum values ​​in the matrix, respectively. The dimensionless values ​​in the standardized temperature rise rate feature matrix output are strictly limited to the range [0,1].

[0101] Iterate through each pixel (i,j) in the thickness distribution pseudo-color image and extract its original grayscale value. (representing the physical thickness of the ice layer), and based on the principles of specific heat capacity and thermal resistance in physics, utilizing... Find the corresponding theoretical temperature rise rate reference value in the preset thickness-temperature rise inverse ratio table. It is important to clarify the underlying logic of this inverse proportionality table: This table is constructed based on extensive offline control experiments. In these experiments, ice samples of different standard thicknesses were heated under constant microwave power, and their temperature rise rate per unit time was recorded. Because thinner ice layers have lower heat capacity and are significantly affected by substrate reflection, their temperature rise rate is typically faster. Conversely, thicker ice layers have higher heat capacity and more uniform heat distribution due to volume diffusion, resulting in a relatively slower temperature rise rate. This is achieved by fitting the experimental data with an inverse proportional function (e.g., ...). The system discretizes and stores the fitted curves, thus establishing a unique mapping table from thickness grayscale values ​​to the theoretical normalized temperature rise rate; the system calculates the standardized actual temperature rise rate under the same coordinate system. Compared with the theoretical reference value obtained from the table The absolute value of the deviation between them, i.e. and will Defined as the physical conflict index, this index objectively quantifies the degree of contradiction between the thickness determination result of the current pixel and the actual heating dynamic characteristics.

[0102] Subsequently, in order to convert physical conflicts into usable corrective weights, the physical conflict index was... Substituting into the preset negative exponential decay function for calculation, the formula is as follows: ,in For the generated confidence weight coefficients, This is the preset severity factor (recommended value range: 2.0 to 5.0). Regarding the preset logic of the negative exponential decay function: the negative exponential function has the mathematical properties of monotonically decreasing and being extremely sensitive to large errors. When the conflict exponent... When approaching 0, the weight The weights approach 1; as the conflict exponent increases, the weights rapidly decay, approaching 0. This design can non-linearly suppress outliers. The value of is usually determined iteratively during the model calibration phase by maximizing the segmentation accuracy of the labeled samples.

[0103] Finally, using the calculated reliability weight coefficients The original grayscale values ​​of the corresponding pixels in the thickness distribution pseudocolor image Perform pixel-level multiplication correction operations, i.e. The result is then rounded to produce a final output of a pseudo-color image showing the thickness distribution after fusion.

[0104] Example 2:

[0105] like Figure 4 As shown, the thermal infrared image-driven microwave de-icing region segmentation system includes:

[0106] The infrared image acquisition module emits a preset short-time microwave pulse excitation to the area to be de-iced and simultaneously acquires a continuous thermal infrared image sequence of the microwave-affected area in the time dimension; the thermal infrared image sequence covers the active temperature rise stage during microwave heating and the passive thermal relaxation stage after microwave unloading.

[0107] The feature matrix construction module performs fitting calculations on each pixel in the thermal infrared image sequence based on the temperature decay trend during the passive thermal relaxation stage, extracts the thermal relaxation time constant, and constructs the thermal relaxation feature matrix using the thermal relaxation time constant.

[0108] The de-icing region classification module maps the thermal relaxation feature matrix into a thickness distribution pseudo-color image, and calculates the transient temperature rise rate during the active temperature rise stage. It then performs pixel-level logical fusion with the thickness distribution pseudo-color image. Based on the gradient difference of the fused thickness distribution pseudo-color image, it outputs a hierarchical segmentation mask containing classification labels for the base region, thin ice region, and thick ice region.

[0109] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0110] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0111] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0112] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for segmenting microwave de-icing regions driven by thermal infrared images, characterized in that, The method includes: A preset short-duration microwave pulse is emitted to the area to be de-iced, and a continuous thermal infrared image sequence of the microwave-affected area in the time dimension is acquired simultaneously; the thermal infrared image sequence covers the active temperature rise stage during microwave heating and the passive thermal relaxation stage after microwave unloading. For each pixel in the thermal infrared image sequence, a fitting operation is performed based on the temperature decay trend during the passive thermal relaxation stage to extract the thermal relaxation time constant; and the thermal relaxation time constant is used to construct a thermal relaxation feature matrix. The thermal relaxation feature matrix is ​​mapped to a pseudo-color image of thickness distribution. The active temperature rise image sequence during microwave heating is obtained from the thermal infrared image sequence. The temperature data of the first and last frames of the sequence are extracted and the difference is calculated. The difference is divided by the duration of microwave heating to obtain the transient temperature rise rate matrix. The transient temperature rise rate matrix is ​​numerically normalized to generate a standardized temperature rise rate feature matrix. Traverse each pixel in the thickness distribution pseudo-color image, and look up the corresponding theoretical temperature rise rate reference value in the preset thickness-temperature rise inverse relationship table according to the gray value of the pixel. Calculate the absolute value of the deviation between the actual temperature rise rate value under the same coordinate in the standardized temperature rise rate feature matrix and the theoretical temperature rise rate reference value, and define the absolute value of the deviation as the physical conflict index. The physical conflict index is substituted into the negative exponential decay function for calculation to generate the confidence weight coefficient; the confidence weight coefficient is used to perform a multiplication correction operation on the original gray value of the corresponding pixel in the thickness distribution pseudo-color image to obtain the fused thickness distribution pseudo-color image. The Laplacian gradient operation is performed on the fused thickness distribution pseudo-color image to extract edge texture features of different thickness regions. The calculated gray values ​​are then compared stepwise with preset base determination thresholds and preset thick ice determination thresholds. Regions below the base determination threshold are marked as base regions, regions between the base determination threshold and the thick ice determination threshold are marked as thin ice regions, and regions above the thick ice determination threshold are marked as thick ice regions. Finally, a hierarchical segmentation mask carrying classification labels for the base region, the thin ice region, and the thick ice region is output.

2. The thermal infrared image-driven microwave de-icing region segmentation method according to claim 1, characterized in that, Extract the thermal relaxation time constant and construct the thermal relaxation feature matrix using the thermal relaxation time constant, including: Segmenting passive thermal relaxation image sequences from thermal infrared image sequences; Pixel-level temporal extraction is performed on the passive thermal relaxation image sequence. Each pixel in the image plane is traversed, and the temperature values ​​of the corresponding multi-frames on the time axis are extracted to form a single-point transient temperature decay curve. The single-point transient temperature decay curve is mapped to convert the original temperature value into a natural logarithm value, and the logarithmic difference between adjacent time steps is calculated to generate a logarithmic difference decay sequence. The logarithmic difference decay sequence is approximated by the standard exponential decay theoretical model using a preset iterative algorithm, and the thermal relaxation time constant of the pixel is obtained by convergence. Based on the planar coordinates of each pixel, the thermal relaxation time constant calculated for each pixel is filled into the corresponding two-dimensional array space to generate a thermal relaxation feature matrix.

3. The method of claim 2, wherein the thermal image is a thermal infrared image. The logarithmic difference decay sequence is approximated to the standard exponential decay theoretical model using a pre-defined iterative algorithm, including: For the logarithmic difference decay sequence, the discrete curvature values ​​of each sampling point are calculated using the central difference method; at the same time, local maxima are retrieved in the discrete curvature value sequence, and the sampling time of the sampling point corresponding to the local maxima is identified as the critical time frame; the logarithmic difference decay sequence is segmented using the critical time frame as the time axis dividing boundary to obtain the transient dominant segment. The single exponential model with a first preset weight is prefitted to the transient dominant segment to calculate the surface heat dissipation interference component, and the surface heat dissipation interference component is subtracted from the original logarithmic difference decay sequence to generate a residual decay sequence. The residual decay sequence is imported into the hyperbolic tangent nonlinear regression model for a second refined iterative approximation. The tail decay rate when the residual decay sequence approaches zero is locked, and the reciprocal of the tail decay rate is defined as the thermal relaxation time constant.

4. The method of claim 3, wherein the thermal image is a thermal infrared image. The solution logic for the surface heat dissipation interference component is as follows: Based on the critical time frame, early data points located before the critical time frame are extracted from the logarithmic difference decay sequence to form a transient dominant data subset; A single exponential benchmark function is constructed, and the nonlinear least squares method is used to fit the single exponential benchmark function to the transient dominant data subset. During the fitting process, the data points are assigned the first preset weight that decreases with time, thereby determining the best estimated values ​​of the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant. Using the determined optimal estimate, the single exponential benchmark function is extended to the time axis of the entire passive thermal relaxation stage. The theoretical value of the function at each moment is calculated, and a simulation curve is generated. This simulation curve is defined as the surface heat dissipation disturbance component.

5. The method of claim 4, wherein, The logic for determining the optimal estimates of the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant is as follows: Based on the transient dominant data subset and the first preset weights, the weighted residual between the actual observed logarithmic temperature value at each moment and the theoretical predicted value of the single exponential benchmark function under the current parameter settings is calculated, and the weighted residuals at all moments are squared and summed to construct a global objective function. The global objective function is subjected to partial differential operations with respect to the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant to generate a sensitivity gradient vector. The parameter update step size is dynamically calculated based on the magnitude of the sensitivity gradient vector. When the magnitude of the sensitivity gradient vector is greater than a preset threshold, the upper limit of the parameter update step size is matched. When the magnitude of the sensitivity gradient vector is less than the preset threshold, the lower limit of the parameter update step size is matched. Using the sensitivity gradient vector and the parameter update step size, the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant are iteratively corrected in the parameter space. After each correction, the value of the global objective function is recalculated until the rate of change of the global objective function value obtained from two adjacent iterations is lower than the preset convergence accuracy threshold. The parameter values ​​in the convergence state at this time are locked as the best estimates of the surface heat dissipation amplitude coefficient and the surface heat dissipation time constant.

6. The method of claim 1, wherein, The thermal relaxation feature matrix is ​​mapped to a pseudo-color image of thickness distribution. The specific mapping logic is as follows: Traverse each pixel coordinate in the thermal relaxation feature matrix, obtain the thermal relaxation time constant corresponding to the current coordinate, and substitute it as an independent variable into the pre-constructed thickness-time constant nonlinear calibration curve model. Calculate and output the physical thickness value corresponding to the pixel coordinate through the mapping of the model, thereby constructing the ice layer thickness physical matrix. The maximum and minimum thickness values ​​in the physical matrix of ice layer thickness are statistically analyzed, and the difference between the two is calculated to determine the global dynamic range of ice layer thickness. For the physical thickness value of each pixel coordinate, the relative proportion of the difference between it and the minimum thickness value in the global dynamic range is calculated to generate a normalized thickness relative coefficient. The thickness relative coefficient is multiplied by the preset maximum bit depth of the single-channel image, and the result is rounded down. The resulting integer value is then used as the pixel gray value and filled into the corresponding coordinate position in the preset single-channel gray space to generate a thickness distribution pseudo-color image.

7. A thermal infrared image driven microwave de-icing zone segmentation system, characterized by, It is implemented based on the thermal infrared image-driven microwave de-icing region segmentation method according to any one of claims 1-6, and includes: The infrared image acquisition module emits a preset short-time microwave pulse excitation to the area to be de-iced and simultaneously acquires a continuous thermal infrared image sequence of the microwave-affected area in the time dimension; the thermal infrared image sequence covers the active temperature rise stage during microwave heating and the passive thermal relaxation stage after microwave unloading. The feature matrix construction module performs fitting calculations on each pixel in the thermal infrared image sequence based on the temperature decay trend during the passive thermal relaxation stage, extracts the thermal relaxation time constant, and constructs the thermal relaxation feature matrix using the thermal relaxation time constant. The de-icing area classification module maps the thermal relaxation feature matrix to a thickness distribution pseudo-color image; it obtains the active temperature rise image sequence during microwave heating from the thermal infrared image sequence, extracts the temperature data of the first and last frames of the sequence and calculates the difference, divides the difference by the microwave heating duration to obtain the transient temperature rise rate matrix; it performs numerical normalization on the transient temperature rise rate matrix to generate a standardized temperature rise rate feature matrix; it iterates through each pixel in the thickness distribution pseudo-color image, looks up the corresponding theoretical temperature rise rate reference value in a preset thickness-temperature rise inverse relationship table based on the gray value of the pixel, calculates the absolute value of the deviation between the actual temperature rise rate value at the same coordinate in the standardized temperature rise rate feature matrix and the theoretical temperature rise rate reference value, and defines the absolute value of the deviation as the physical conflict index; it then calculates the physical conflict index. The burst index is substituted into the negative exponential decay function for calculation to generate the confidence weight coefficient. The confidence weight coefficient is used to perform a multiplication correction operation on the original gray value of the corresponding pixel in the thickness distribution pseudo-color image to obtain the fused thickness distribution pseudo-color image. The Laplacian gradient operation is performed on the fused thickness distribution pseudo-color image to extract the edge texture features of different thickness regions. The calculated gray value is compared stepwise with the preset basal determination threshold and the preset thick ice determination threshold. Regions below the basal determination threshold are marked as basal regions, regions between the basal determination threshold and the thick ice determination threshold are marked as thin ice regions, and regions above the thick ice determination threshold are marked as thick ice regions. Finally, a hierarchical segmentation mask carrying the classification labels of the basal region, the thin ice region, and the thick ice region is output.

Images