Infrared thermal imaging water ripple noise frequency domain suppression method and system

Abstract

The present application relates to the field of infrared imaging technology, and discloses a kind of infrared thermal imaging water ripple noise frequency domain suppression method and system, the method includes the original infrared image sequence of non-refrigeration infrared core acquisition is carried out time domain accumulation and frequency domain transformation to obtain average power spectrum matrix and constructs water ripple noise frequency domain feature topology graph, based on topological graph executes noise region hierarchical clustering and filter kernel parameter distribution generation hierarchical adaptive trap wave filter parameter table, based on parameter table to the frequency domain representation of the infrared image to be processed executes hierarchical trap wave filtering and edge frequency domain protection, then through the effect of inhibition evaluation and filter parameter online correction obtains the infrared image after water ripple suppression;The present application is realized hierarchical clustering and adaptive trap wave filtering by constructing noise frequency domain feature topology graph, effectively suppresses water ripple noise and preserves image detail, realizes the balance of noise suppression and detail preservation, and can dynamically adapt noise characteristic change.

Description

Technical Field

[0001] This invention relates to the field of infrared imaging technology, and in particular to a method and system for frequency domain suppression of water ripple noise in infrared thermal imaging. Background Technology Uncooled infrared imaging technology senses the thermal radiation of a target using an infrared detector and converts it into an electrical signal. After signal processing, an infrared image is formed. It is widely used in scenarios such as surveillance, security, search and rescue, and vehicle-mounted systems. This technology features low cost, simple structure, and low power consumption. However, in practical applications, noise processing is required to improve image quality and enhance target observation and recognition.

[0002] In practical engineering applications, engineers have found that uncooled infrared sensors exhibit wavy, striped interference in images displaying uniform scenes. The root cause lies at the hardware level, including non-uniformity of the detector array, timing interference in the readout circuitry, and analog noise during signal transmission. This wavy noise reduces the uniformity and clarity of infrared images, leading to eye strain when observing uniform scenes. Furthermore, during target detection, it may obscure or misinterpret areas of abnormal temperature as actual locations, affecting the accurate identification of true targets. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, the present invention provides the following technical solution: A method for frequency domain suppression of water ripple noise in infrared thermal imaging includes: The raw infrared image sequence acquired by the uncooled infrared sensor is subjected to time-domain accumulation and frequency-domain transformation to obtain the average power spectrum matrix. Based on the average power spectrum matrix, a water ripple noise frequency domain feature topology map is constructed. Based on the water ripple noise frequency domain feature topology map, hierarchical clustering of noise regions and allocation of filter kernel parameters are performed to generate a hierarchical adaptive notch filter parameter table. Based on the hierarchical adaptive notch filter parameter table, hierarchical notch filtering and edge frequency domain protection are performed on the frequency domain representation of the infrared image to be processed to obtain the infrared image after water ripple suppression. Based on the infrared image after water ripple suppression and the infrared image to be processed, the suppression effect is evaluated and the filter parameters are corrected online. The correction results are written back to the hierarchical adaptive notch filter parameter table and the final display image is output.

[0004] Furthermore, the steps of performing time-domain accumulation and frequency-domain transformation on the raw infrared image sequence acquired by the uncooled infrared sensor to obtain the average power spectrum matrix, and constructing the frequency domain feature topology map of water ripple noise based on the average power spectrum matrix include: acquiring a K-frame raw infrared image sequence continuously acquired by the uncooled infrared sensor in a uniform scene; performing non-uniform correction and blind pixel replacement on each frame in the raw infrared image sequence to obtain a preprocessed infrared image sequence; performing a two-dimensional discrete Fourier transform on each frame in the preprocessed infrared image sequence to obtain the frequency domain complex matrix corresponding to each frame; calculating the square of the amplitude spectrum of the frequency domain complex matrix of each frame; and then averaging the squares of the amplitude spectra of all frames along the frame dimension to obtain the average power spectrum matrix; and constructing the frequency domain feature topology map of water ripple noise based on the average power spectrum matrix.

[0005] Further, the steps for obtaining a preprocessed infrared image sequence by acquiring a K-frame raw infrared image sequence continuously acquired by an uncooled infrared camera in a uniform scene, and performing non-uniform correction and blind pixel replacement on each frame in the raw infrared image sequence include: pointing the lens of the uncooled infrared camera at a uniform radiation temperature panel, continuously acquiring K frames of infrared images at a fixed frame rate in front of the uniform radiation temperature panel, arranging the K frames of infrared images in the order of acquisition time to form the raw infrared image sequence; performing two-point non-uniform correction on each frame in the raw infrared image sequence, performing blind pixel replacement on each corrected frame after completing the two-point non-uniform correction, and arranging the K frames of images after the two-point non-uniform correction and blind pixel replacement in the original time order to obtain the preprocessed infrared image sequence.

[0006] Furthermore, the steps for constructing a frequency domain feature topology map of water ripple noise based on the average power spectrum matrix include: performing a logarithmic transformation with the base of natural numbers on each element of the average power spectrum matrix to obtain a logarithmic power spectrum matrix; performing local median filtering on the logarithmic power spectrum matrix to obtain a background power spectrum matrix; subtracting the background power spectrum matrix from the logarithmic power spectrum matrix pixel by pixel to obtain a residual power spectrum matrix; setting an adaptive significance threshold on the residual power spectrum matrix; marking the frequency domain positions in the residual power spectrum matrix that are greater than the adaptive significance threshold as candidate noise frequency points; performing eight-connected domain analysis on all candidate noise frequency points in the frequency domain coordinate space; aggregating spatially adjacent candidate noise frequency points into noise frequency domain connected regions; calculating regional attribute parameters for each noise frequency domain connected region; using each noise frequency domain connected region as a topology node and the regional attribute parameters as node attributes; establishing topology edges between topology nodes whose frequency domain distance is less than a preset frequency domain adjacency radius; and using the ratio of the regional average energy values ​​of the two topology nodes carried by each topology edge as the edge weight to form a frequency domain feature topology map of water ripple noise.

[0007] Furthermore, based on the frequency domain feature topology map of water ripple noise, the steps of performing hierarchical clustering of noise regions and allocation of filter kernel parameters to generate a hierarchical adaptive notch filter parameter table include: extracting the regional attribute parameters of all topological nodes from the frequency domain feature topology map of water ripple noise; sorting all topological nodes in descending order according to the average energy value of the regions to obtain an energy-ranked node list; performing hierarchical clustering based on graph connectivity on the topological nodes in the energy-ranked node list to generate hierarchical clustering results of noise regions; and allocating corresponding notch filter kernel parameters to each noise frequency domain cluster according to the hierarchical clustering results of noise regions to generate a hierarchical adaptive notch filter parameter table.

[0008] Further, the steps for performing hierarchical clustering based on graph connectivity on the topological nodes in the energy sorting node list to generate hierarchical clustering results for noise regions include: setting the number of energy levels to 3, namely strong noise level, medium noise level, and weak noise level; calculating two energy boundary thresholds based on the distribution of the regional average energy values ​​of all topological nodes in the energy sorting node list using the Otsu's method; classifying topological nodes with regional average energy values ​​greater than the first energy boundary threshold into the strong noise level; classifying topological nodes with regional average energy values ​​greater than the second energy boundary threshold and less than or equal to the first energy boundary threshold into the medium noise level; and classifying topological nodes with regional average energy values ​​less than or equal to the second energy boundary threshold into the weak noise level; within the same noise level, traversing the topological nodes and their topological edges belonging to that level in the water ripple noise frequency domain feature topology graph; merging topological nodes directly connected by topological edges into the same noise frequency domain cluster; each noise frequency domain cluster inherits the noise level label to which it belongs; and after completing the above connected component search in the three levels of strong noise level, medium noise level, and weak noise level, all noise frequency domain clusters constitute the hierarchical clustering results for noise regions.

[0009] Furthermore, based on the hierarchical adaptive notch filter parameter table, the steps for performing hierarchical notch filtering and edge frequency domain protection on the frequency domain representation of the infrared image to be processed to obtain the infrared image after water ripple suppression include: acquiring the current frame infrared image acquired in real time by the uncooled infrared sensor as the infrared image to be processed; performing two-point non-uniform correction and blind pixel replacement on the infrared image to be processed; performing a two-dimensional discrete Fourier transform to obtain the current frame frequency domain complex matrix; constructing a comprehensive notch filter transfer function based on the hierarchical adaptive notch filter parameter table; performing edge frequency domain protection processing on the current frame frequency domain complex matrix and then applying the comprehensive notch filter transfer function to obtain the filtered frequency domain complex matrix; performing a two-dimensional discrete Fourier inverse transform on the filtered frequency domain complex matrix to obtain the spatial domain reconstructed image; and performing grayscale value truncation and normalization processing on the spatial domain reconstructed image to obtain the infrared image after water ripple suppression.

[0010] Furthermore, based on the hierarchical adaptive notch filter parameter table, the steps for constructing the comprehensive notch filter transfer function include: initializing a one matrix of the same size as the current frame's frequency domain complex matrix as the basic transfer function matrix; traversing each filter parameter record in the hierarchical adaptive notch filter parameter table, and for each filter parameter record, constructing a Gaussian notch attenuation kernel with its center frequency coordinate as the center and the notch bandwidth parameter as the bandwidth radius, while constructing the same Gaussian notch attenuation kernel at the conjugate symmetry position of the current frame's frequency domain complex matrix according to the frequency domain conjugate symmetry of the two-dimensional discrete Fourier transform; superimposing the Gaussian notch attenuation kernel corresponding to each filter parameter record onto the basic transfer function matrix, and after traversing all filter parameter records, updating the basic transfer function matrix to the comprehensive notch filter transfer function.

[0011] Furthermore, the steps of evaluating the suppression effect and correcting the filtering parameters online based on the infrared image after water ripple suppression and the infrared image to be processed, writing the correction results back to the hierarchical adaptive notch filter parameter table, and outputting the final display image include: calculating the water ripple residual index and detail retention index based on the infrared image after water ripple suppression and the infrared image to be processed; determining the current filtering effect level and performing parameter correction based on the water ripple residual index and detail retention index to obtain the corrected hierarchical adaptive notch filter parameter table; storing the corrected hierarchical adaptive notch filter parameter table in the parameter register of the image preprocessing module of the uncooled infrared core; and outputting the infrared image after water ripple suppression to the display terminal as the final display image.

[0012] An infrared thermal imaging water ripple noise frequency domain suppression system is provided for implementing the aforementioned infrared thermal imaging water ripple noise frequency domain suppression method. The system includes: Water ripple noise frequency domain feature topology map construction module: used to perform time domain accumulation and frequency domain transformation on the raw infrared image sequence acquired by the uncooled infrared core to obtain the average power spectrum matrix, and construct the water ripple noise frequency domain feature topology map based on the average power spectrum matrix; Hierarchical adaptive notch filter parameter table generation module: Based on the frequency domain feature topology map of water ripple noise, it performs hierarchical clustering of noise regions and allocation of filter kernel parameters to generate hierarchical adaptive notch filter parameter tables. Water ripple suppression infrared image acquisition module: Based on the hierarchical adaptive notch filter parameter table, it performs hierarchical notch filtering and edge frequency domain protection on the frequency domain representation of the infrared image to be processed to obtain the water ripple suppression infrared image; Suppression effect evaluation and parameter correction module: Based on the infrared image after water ripple suppression and the infrared image to be processed, it performs suppression effect evaluation and online correction of filter parameters, writes the correction results back to the hierarchical adaptive notch filter parameter table and outputs the final display image.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the problem of water ripple noise suppression in uncooled infrared images. By combining time-domain accumulation and frequency-domain transformation, a topological map of the water ripple noise frequency domain features is constructed. This effectively solves the problems of confusion between the frequency domain features of random noise and water ripple noise, and inaccurate noise frequency point localization in traditional noise suppression methods. This topological map not only records the independent attributes of connected regions in the noise frequency domain but also reflects the spatial adjacency and energy correlation between regions through topological edges, providing crucial relational features for subsequent hierarchical clustering and avoiding the fragmentation of filter design caused by isolated frequency point analysis. Energy hierarchical and graph connectivity clustering based on the topological map achieves differentiated processing of noise regions. Strong noise levels are subjected to strong attenuation to ensure thorough suppression, while weak noise levels are subjected to weak attenuation to avoid excessive damage to effective information, resolving the contradiction of insufficient noise suppression or severe detail loss under uniform attenuation intensity. A Gaussian notch attenuation kernel and an edge frequency domain protection mechanism are employed. Frequency domain multiplication precisely controls the attenuation range and depth, while attenuation coefficients at edge-sensitive frequency domain locations are corrected retrospectively, effectively avoiding the contour blurring problem caused by overlap between the notch filter and the target edge frequency domain. By combining the dual-index evaluation of water ripple residual and detail retention with online parameter correction closed loop, it can dynamically adapt to the noise characteristic changes of uncooled infrared cores caused by factors such as temperature drift. It solves the problem of the degradation of the suppression effect of fixed filter parameters over time, and achieves long-term stable suppression of water ripple noise and effective preservation of infrared image details. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0015] Figure 1 This is a flowchart of a method for frequency domain suppression of water ripple noise in infrared thermal imaging according to the present invention; Figure 2 This is a schematic diagram of uniform scene acquisition in an embodiment of the present invention; Figure 3 This is a schematic diagram of time-domain cumulative averaging in an embodiment of the present invention; Figure 4 This is a schematic diagram of the frequency domain characteristics of water ripple noise in an embodiment of the present invention; Figure 5 This is a schematic diagram of hierarchical clustering of noise regions in an embodiment of the present invention; Figure 6This is a schematic diagram illustrating the construction of the integrated notch filter transfer function in an embodiment of the present invention; Figure 7 This is a schematic diagram of edge frequency domain protection processing in an embodiment of the present invention; Figure 8 This is a closed-loop diagram of online correction of filter parameters in an embodiment of the present invention; Figure 9 This is a functional block diagram of an infrared thermal imaging water ripple noise frequency domain suppression system according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Please see Figure 1 As shown, this embodiment provides a method for frequency domain suppression of water ripple noise in infrared thermal imaging, including: S1: Perform time-domain accumulation and frequency-domain transformation on the original infrared image sequence acquired by the uncooled infrared sensor to obtain the average power spectrum matrix, and construct the frequency domain feature topology map of water ripple noise based on the average power spectrum matrix.

[0018] Further, step S1 includes: S11: Obtain a sequence of K frames of raw infrared images continuously acquired by an uncooled infrared sensor in a uniform scene. Perform non-uniform correction and blind pixel replacement on each frame of the raw infrared image sequence to obtain a preprocessed infrared image sequence.

[0019] Further, step S11 includes: S111: Align the lens of the uncooled infrared sensor with a uniform radiation temperature panel, wherein the uniform radiation temperature panel refers to a standard blackbody radiation plate or a constant-temperature planar target with a uniform spatial temperature distribution over time. See also Figure 2This is a schematic diagram of uniform scene acquisition provided in an embodiment of this application. As shown in the figure, the lens of the uncooled infrared camera module faces the uniform radiation temperature panel, which emits infrared radiation towards the camera module. The camera module continuously acquires multiple frames of infrared images at a fixed frame rate to form the original infrared image sequence. In each acquired frame, due to the uniform spatial distribution of the surface temperature of the uniform radiation temperature panel, the image does not contain target texture and edge information. The image appears as a uniform background superimposed with water ripple noise, which is represented by a wave-like texture in the figure. In practical applications such as security monitoring and vehicle night vision, uncooled infrared cameras introduce water ripple noise into the image due to crosstalk in the readout circuit and inherent structural defects in the heat conduction path. This noise is mixed with the target signal and random noise, making it difficult to separate directly. The core function of using uniform scene acquisition is to eliminate the interference of the target signal on noise analysis, making the water ripple noise the only non-uniform feature in the image. This provides a clean data foundation for accurately locating the frequency distribution of the water ripple noise in the frequency domain, avoiding the obscuring and confusion of the noise spectrum by the target spectrum.

[0020] S112: Perform two-point non-uniformity correction on each frame of the original infrared image sequence. This involves acquiring calibration images using two uniform radiation calibration sources at different temperatures, calculating the gain correction coefficient and offset correction coefficient for each pixel based on the two sets of calibration images, multiplying the grayscale value of each pixel in each frame of the original infrared image sequence by the pixel's gain correction coefficient, and adding the pixel's offset correction coefficient to ensure all pixels output consistent grayscale values ​​under the same radiation input. After completing the two-point non-uniformity correction, perform blind pixel replacement on each corrected frame. This involves determining the coordinate position of the blind pixel in each frame according to a pre-calibrated blind pixel position mapping table, and replacing the blind pixel's grayscale value with the median of the effective pixel grayscale values ​​(excluding the blind pixel itself) within a 3x3 neighborhood window centered on the blind pixel's coordinate position. The K frames after the two-point non-uniformity correction and blind pixel replacement are arranged in their original chronological order to obtain the preprocessed infrared image sequence. Specifically, the two-point non-uniformity correction eliminates the spatial fixed pattern noise introduced by the response rate deviation of each pixel of the focal plane array detector, and the blind pixel replacement repairs the isolated anomalies caused by individual failed pixels. The above two operations ensure that the spatial non-uniformity features remaining in the preprocessed infrared image sequence are dominated only by water ripple noise, rather than by the difference in the pixel-level response of the detector. This lays the prerequisite for accurately extracting the frequency domain energy distribution of water ripple noise when performing frequency domain transformation on the preprocessed infrared image sequence in step S12.

[0021] S12: Perform a two-dimensional discrete Fourier transform on each frame of the preprocessed infrared image sequence to obtain the frequency domain complex matrix corresponding to each frame. Calculate the square of the amplitude spectrum of the frequency domain complex matrix of each frame, and then take the mean of the squares of the amplitude spectra of all frames along the frame dimension to obtain the average power spectrum matrix.

[0022] Further, step S12 includes: S121: Perform a two-dimensional discrete Fourier transform on the k-th frame of the preprocessed infrared image sequence, where k ranges from 1 to K, to obtain the frequency domain complex matrix corresponding to the k-th frame. The size of the frequency domain complex matrix is ​​consistent with the number of pixel rows and columns of the preprocessed infrared image. Each element in the frequency domain complex matrix is ​​a complex value, whose amplitude represents the energy of the corresponding spatial frequency component, and whose phase represents the displacement information of the corresponding spatial frequency component. Take the square of the complex modulus of each element in the frequency domain complex matrix corresponding to the k-th frame of the preprocessed infrared image to obtain the single-frame power spectrum matrix corresponding to the k-th frame.

[0023] S122: Calculate the arithmetic mean element-wise along the frame dimension of the K single-frame power spectrum matrices corresponding to the K preprocessed infrared images. That is, for each spatial frequency coordinate in the frequency domain, add the power values ​​of the K single-frame power spectrum matrices at that position and divide by K to obtain the average power spectrum matrix. The physical meaning of the average power spectrum matrix is ​​that, under multi-frame temporal accumulation, the power contribution of random noise is suppressed during the averaging process due to the lack of inter-frame correlation, and its power value approaches the expected level of the noise variance. See also... Figure 3 This is a schematic diagram of time-domain cumulative averaging provided in an embodiment of this application. As shown in the figure, frames 1, 2 to K are multiple frames in a preprocessed infrared image sequence. Each frame contains both water ripple noise represented by wavy textures and random noise represented by black scattered dots. The spatial location and shape of the water ripple noise remain consistent across frames, labeled as inter-frame correlated; the scattered dots of random noise vary across frames, labeled as inter-frame uncorrelated. After time-domain cumulative averaging, the water ripple noise texture is more clearly prominent in the average power spectrum matrix output below, while the scattered dots of random noise disappear. In the actual operation of uncooled infrared cores, the focal plane array detector is simultaneously affected by water ripple noise and random thermal noise, and their energy distribution in the frequency domain of a single frame may be confused. Time-domain cumulative averaging utilizes the key difference that water ripple noise is highly correlated between frames while random noise is statistically independent between frames. By averaging across multiple frames, the power contribution of random noise is effectively suppressed, while the power value of water ripple noise does not decay due to inter-frame consistency but instead stands out as a significant spike higher than the background, thus achieving reliable separation of the two types of noise in the frequency domain.

[0024] S13: Construct a frequency domain feature topology map of water ripple noise based on the average power spectrum matrix.

[0025] Further, step S13 includes: S131: Perform a logarithmic transformation with a base of natural numbers on each element of the average power spectrum matrix to obtain a logarithmic power spectrum matrix. The purpose of this logarithmic transformation is to compress the dynamic range of power values ​​in the average power spectrum matrix, transforming the power difference between the water ripple noise frequency point and the background frequency point from a multiplicative relationship to an additive relationship, facilitating subsequent separation of noise and background frequencies through difference operations. Perform local median filtering on the logarithmic power spectrum matrix. The filtering window size of the local median filter is a square window with an odd side length. The window side length is determined based on being greater than the principal axis length of the largest connected region occupied by the water ripple noise in the frequency domain and less than one-tenth of the frequency domain size. For example, the window side length can be set to 21 frequency pixels or 31 frequency pixels. For each frequency position in the logarithmic power spectrum matrix, the local median filter takes the median of all elements within its neighborhood window as the output value for that position, obtaining the background power spectrum matrix. The background power spectrum matrix represents a smooth background power distribution in the frequency domain that does not contain water ripple noise spikes. The residual power spectrum matrix is ​​obtained by subtracting the background power spectrum matrix pixel by pixel from the logarithmic power spectrum matrix. In the residual power spectrum matrix, the residual value at the water ripple noise frequency is positive and significantly greater than zero, while the residual value at the background frequency fluctuates around zero. Specifically, local median filtering is chosen instead of mean filtering as the background estimation method because water ripple noise manifests as local narrowband spikes in the frequency domain. Mean filtering does not completely suppress these spikes, resulting in residual information of some spikes in the background power spectrum matrix. This leads to an underestimation of the residual value at the water ripple noise frequency in the residual power spectrum matrix. Median filtering, on the other hand, has a natural suppression characteristic for local spikes, resulting in a more accurate background power estimate.

[0026] S132: An adaptive significance threshold is set for the residual power spectrum matrix, and frequency domain positions in the residual power spectrum matrix that are greater than the adaptive significance threshold are marked as candidate noise frequencies. The adaptive significance threshold is determined as follows: the arithmetic mean of all elements in the residual power spectrum matrix is ​​calculated and denoted as the residual mean; the standard deviation of all elements in the residual power spectrum matrix is ​​calculated and denoted as the residual standard deviation; the residual mean is added to the residual standard deviation by a preset multiple as the adaptive significance threshold. The preset multiple is determined based on the one-sided confidence probability level corresponding to the normal distribution assumption. The larger the preset multiple, the more stringent the screening of candidate noise frequencies, increasing the risk of missed detection but reducing the risk of false detection; the smaller the preset multiple, the more lenient the screening, increasing the risk of false detection but reducing the risk of missed detection. For example, the preset multiple can be set to 3, corresponding to a 99.7% one-sided confidence probability, or 2.5. For example, assuming the residual mean of the residual power spectrum matrix is ​​0.12, the residual standard deviation is 0.08, and the preset multiplier is 3, the adaptive significance threshold is calculated as 0.12 plus 3 multiplied by 0.08, which equals 0.36. Frequency domain positions in the residual power spectrum matrix with residual values ​​greater than 0.36 are marked as candidate noise frequencies. Specifically, the adaptive significance threshold is designed using a data-driven approach rather than a preset fixed threshold because the energy levels and frequency distributions of water ripple noise differ significantly between different models of uncooled infrared sensors and under different ambient temperatures; a fixed threshold cannot adapt to all operating conditions. Using a linear combination of the residual mean and residual standard deviation as the threshold automatically adjusts the screening sensitivity based on the actual statistical distribution of the residual power spectrum matrix under the current acquisition conditions, ensuring effective separation of water ripple noise frequencies from random noise background under different sensors and operating conditions.

[0027] S133: Perform an octet analysis on all candidate noise frequencies in the frequency domain coordinate space. This octet analysis involves checking if each candidate noise frequency's eight adjacent positions (up, down, left, right, upper left, upper right, lower left, and lower right) are also marked as candidate noise frequencies. If adjacent positions are also candidate noise frequencies, they are grouped into the same connected region. After traversing all candidate noise frequencies, spatially adjacent candidate noise frequencies are aggregated into noise frequency domain connected regions. For each noise frequency domain connected region, calculate region attribute parameters, including the region centroid frequency coordinates, region area, region major axis direction angle, and region average energy value. The region centroid frequency coordinates are the weighted average coordinates calculated using the residual values ​​of each frequency point in the residual power spectrum matrix as weights for all candidate noise frequencies in the noise frequency domain connected region. The region area is the total number of candidate noise frequencies contained in the noise frequency domain connected region. The major axis orientation angle of the region is the angle between the major axis of the smallest circumscribed ellipse of the two-dimensional point set formed by the frequency domain coordinates of all candidate noise frequency points in the connected region of the noise frequency domain and the horizontal axis of the frequency domain. Its physical meaning is the frequency domain distribution direction of the water ripple noise in the connected region. The average energy value of the region is the arithmetic mean of the power values ​​of all candidate noise frequency points in the connected region of the noise frequency domain at the corresponding positions in the average power spectrum matrix.

[0028] S134: Using each noise frequency domain connected region as a topology node and the region attribute parameter as the node attribute, establish topology edges between topology nodes whose frequency domain distance is less than a preset frequency domain adjacency radius. See also Figure 4 This is a schematic diagram of the frequency domain feature topology of water ripple noise provided in this application embodiment. As shown in the figure, the dashed rectangle represents the frequency domain plane, in which six elliptical noise frequency domain connected regions are distributed. The center of each region is marked with a solid circle indicating the centroid frequency coordinate of the region. Each region is divided into three levels according to the average energy value of the region: strong noise level, medium noise level, and weak noise level, which are distinguished by different shades of fill color. Regions with a frequency domain distance less than a preset frequency domain adjacency radius are connected by topological edges, and the edge weights on the topological edges reflect the degree of energy contrast between adjacent regions. In the water ripple noise suppression task of uncooled infrared core, the noise in the frequency domain is not concentrated at a single frequency point, but is dispersed into several connected regions with different energy levels and spatial locations, and some regions are spatially adjacent to each other because they originate from the same hardware root. Organizing the connected regions in the frequency domain of noise into an undirected weighted graph structure with attributes not only fully records the independent attribute characteristics of each noise region, but also preserves the spatial adjacency relationship and energy correlation information between regions through topological edges. This provides an irreplaceable relational input for the subsequent hierarchical clustering of noise regions that are close in frequency domain location and originate from the same root cause.

[0029] The overall function and technical effect of step S1 are analyzed as follows. Specifically, step S1 extracts the frequency spatial distribution information of water ripple noise from the original infrared image sequence acquired by the uncooled infrared sensor through a combination of time-domain accumulation and frequency-domain transformation, and organizes this information into an attributed undirected weighted graph structure, namely, the frequency domain feature topology map of water ripple noise. The combined strategy of uniform scene acquisition and multi-frame time-domain averaging has the core advantage of separating water ripple noise from the target signal and random noise, making it the only significant feature in the frequency domain, thus avoiding interference from the target signal spectrum and random noise spectrum during subsequent notch filter design. The background estimation strategy combining logarithmic transformation and local median filtering has the core advantage of accurately removing the frequency domain peak features of water ripple noise from the frequency domain background, and the natural suppression characteristic of median filtering for peaks ensures the accuracy of background estimation. The noise feature representation method, which combines connected component analysis and topology graph construction, has the core advantage of not only recording the independent attributes of each noise frequency domain connected region, but also recording the spatial adjacency and energy correlation between noise frequency domain connected regions through topological edges. This relational feature representation provides an irreplaceable input structure for performing hierarchical clustering based on graph connectivity in step S2. If step S1 only outputs an independent list of noise frequency points without constructing a topology graph structure, step S2 will be unable to utilize the spatial adjacency between noise frequency domain connected regions for clustering. It will then have to design a notch filter for each noise frequency point individually, resulting in an excessive number of filters, fragmented frequency domain coverage, and unstable filtering performance. The water ripple noise frequency domain feature topology graph, as the core output variable of step S1, provides complete frequency domain spatial relationship information and energy hierarchy information for hierarchical clustering and filter kernel parameter allocation in step S2, serving as a key data hub connecting frequency domain noise analysis and filter design.

[0030] S2: Based on the frequency domain feature topology map of water ripple noise, perform hierarchical clustering of noise regions and allocation of filter kernel parameters to generate a hierarchical adaptive notch filter parameter table.

[0031] Further, step S2 includes: S21: Extract the regional attribute parameters of all topological nodes from the frequency domain feature topology map of the water ripple noise. Sort all topological nodes in descending order according to the average regional energy value to obtain an energy-sorted node list. The energy-sorted node list is an ordered sequence structure, where each element is a reference to a topological node and its regional attribute parameters. The sequence is arranged in descending order of average regional energy value, and the topological node at the top of the sequence corresponds to the most energetic noise frequency domain connected region in the frequency domain.

[0032] S22: Perform hierarchical clustering based on graph connectivity on the topological nodes in the energy sorting node list to generate hierarchical clustering results for noisy regions.

[0033] Further, step S22 includes: S221: Set the energy levels to 3: strong noise, medium noise, and weak noise. Based on the distribution of the regional average energy values ​​of all topological nodes in the energy ranking node list, calculate two energy boundary thresholds using the maximum inter-class variance method. The input to the maximum inter-class variance method is a one-dimensional numerical sequence consisting of the regional average energy values ​​of all topological nodes in the energy ranking node list. The output is the two energy boundary thresholds that maximize the sum of the inter-class variances among the three categories, denoted as the first energy boundary threshold and the second energy boundary threshold, where the first energy boundary threshold is greater than the second energy boundary threshold. Topological nodes with a regional average energy value greater than the first energy boundary threshold are classified as strong noise level; topological nodes with a regional average energy value greater than the second energy boundary threshold and less than or equal to the first energy boundary threshold are classified as medium noise level; and topological nodes with a regional average energy value less than or equal to the second energy boundary threshold are classified as weak noise level. For example, suppose the energy sorting node list contains 12 topological nodes, and the regional average energy value distribution ranges from 500 to 50000. The first energy boundary threshold is calculated to be 28000 and the second energy boundary threshold is 8000 using the Otsu's method. Then, topological nodes with a regional average energy value greater than 28000 are classified as strong noise level, topological nodes with a regional average energy value greater than 8000 and less than or equal to 28000 are classified as medium noise level, and topological nodes with a regional average energy value less than or equal to 8000 are classified as weak noise level. Specifically, the Otsu's method is used instead of equal-interval grading or manually preset fixed thresholds because the energy distribution of water ripple noise varies significantly among different uncooled infrared cores. Some cores may have a small number of extremely strong noise frequency domain connected regions coexisting with a large number of weak noise frequency domain connected regions. Under such skewed distribution, equal-interval grading would classify a large number of weak noise frequency domain connected regions into the same level and lose its ability to distinguish them. Otsu's method can automatically find the optimal dividing point based on the actual distribution of the average energy value of the region, so as to maximize the energy difference between different levels.

[0034] S222: Within the same noise level, traverse the topological nodes and their edges belonging to that level in the water ripple noise frequency domain feature topology graph, and merge the topological nodes directly connected by topological edges into the same noise frequency domain cluster. See also Figure 5This is a schematic diagram of hierarchical clustering of noise regions provided in an embodiment of this application. As shown in the figure, the noise frequency domain connected regions in the frequency domain plane are divided into three levels according to energy level: strong noise level, medium noise level, and weak noise level, represented by grayscale blocks from dark to light. Within the same noise level, regions directly connected by topological edges are merged into the same noise frequency domain cluster, with the boundaries of the clusters marked by dashed boxes. A total of 6 clusters are formed in the figure, where cluster 1 contains two strong noise level regions connected by topological edges, cluster 2 contains two connected medium noise level regions, and clusters 3 to 6 each contain an independent region. In the practical application of uncooled infrared cores, the same crosstalk path of the readout circuit or the same thermal conduction structure defect often generates multiple spatially adjacent and energy-similar noise regions in the frequency domain. If these regions are split into independent filtering targets and notch filters are designed separately, frequency domain coverage gaps may occur between the filters, resulting in incomplete noise suppression. By clustering at the same level using the connectivity of topological edges, adjacent regions originating from the same hardware root cause are grouped into a unified noise frequency domain cluster, ensuring that the notch filters subsequently assigned to each cluster can continuously and completely cover the actual occupied range of the noise.

[0035] S23: Based on the hierarchical clustering results of the noise region, assign corresponding notch filter kernel parameters to each noise frequency domain cluster and generate a hierarchical adaptive notch filter parameter table.

[0036] Further, step S23 includes: S231: For each noise frequency domain cluster, calculate the weighted average of the centroid frequency coordinates of all topological nodes within the cluster as the center frequency coordinate of the cluster, with the weight being the area of ​​each topological node. The weighted average is calculated as follows: multiply the horizontal component of the centroid frequency coordinate of each topological node within the cluster by the area of ​​that node, sum the results, and then divide by the sum of the areas of all topological nodes within the cluster to obtain the horizontal component of the center frequency coordinate. The same operation is performed on the vertical component to obtain the vertical component of the center frequency coordinate. Simultaneously, calculate the frequency domain range covered by all topological nodes within the cluster as the frequency domain span of the cluster. The frequency domain span is calculated as follows: project the frequency domain coordinates of all candidate noise frequency points among all topological nodes within the cluster onto the horizontal and vertical axes, calculate the difference between the maximum and minimum coordinate values ​​in the horizontal direction and the difference between the maximum and minimum coordinate values ​​in the vertical direction, and take the larger of the two as the frequency domain span. For example, suppose a noise frequency domain cluster contains 3 topological nodes with area of ​​15, 25 and 10 frequency points respectively, and the centroid frequency coordinates of the regions are 120 horizontal and 80 vertical, 130 horizontal and 85 vertical, and 125 horizontal and 78 vertical respectively. Then the horizontal component of the center frequency coordinate is calculated as the sum of 120 x 15 + 130 x 25 + 125 x 10 divided by the sum of 15 x 25 + 10, which equals 1800 + 3250 + 1250 divided by 50, which equals 126. The vertical component is calculated as the sum of 80 x 15 + 85 x 25 + 78 x 10 divided by 50, which equals 1200 + 2125 + 780 divided by 50, which equals 82.1.

[0037] S232: Determine the notch attenuation depth coefficient based on the noise level marking. The notch attenuation depth coefficient is a real number between 0 and 1. The closer the value is to 0, the stronger the attenuation, i.e., the more thorough the energy suppression at that frequency position; the closer the value is to 1, the weaker the attenuation. The notch attenuation depth coefficient is the smallest for strong noise level, i.e., the strongest attenuation; the next strongest is for medium noise level, and the largest is for weak noise level, i.e., the weakest attenuation. The notch attenuation depth coefficient is determined to ensure that the ripple noise is effectively suppressed while avoiding excessive attenuation that introduces ringing artifacts. For example, the notch attenuation depth coefficient for strong noise level can be set to 0.05, for medium noise level to 0.20, and for weak noise level to 0.50. Determine the notch bandwidth parameter based on the frequency domain span. The notch bandwidth parameter is a positive real number that characterizes the attenuation coverage radius of the notch filter around the center frequency coordinate. The larger the frequency domain span, the larger the notch bandwidth parameter, to ensure complete coverage of the actual occupied range of the noise frequency domain cluster. The notch bandwidth parameter is determined by multiplying the frequency domain span by a bandwidth expansion factor. The bandwidth expansion factor is a constant greater than 1, and its determination is based on reserving an appropriate frequency domain margin to cope with slight drift of the noise frequency position while fully covering the noise area. For example, the bandwidth expansion factor can be set to 1.2 or 1.5.

[0038] S233: Combine the center frequency coordinates, notch bandwidth parameters, notch attenuation depth coefficient, and noise level marker of each noise frequency domain cluster into a single filter parameter record. All filter parameter records for all noise frequency domain clusters are arranged hierarchically according to the noise level marker, with strong noise level records at the top, followed by medium noise level records, and weak noise level records at the bottom, forming a hierarchical adaptive notch filter parameter table. This hierarchical adaptive notch filter parameter table is a hierarchically indexed structured parameter table. Each record corresponds to the complete filter control parameters for a noise frequency domain cluster to be suppressed. The number of rows in the hierarchical adaptive notch filter parameter table equals the total number of noise frequency domain clusters, and the number of columns is four, corresponding to the center frequency coordinates, notch bandwidth parameters, notch attenuation depth coefficient, and noise level marker, respectively.

[0039] The overall function and technical effect of step S2 are analyzed as follows. Specifically, step S2 transforms the water ripple noise frequency domain feature topology map output from step S1 into a structured parameter table that can directly drive the construction of notch filters, completing the mapping from noise feature description to filter design parameters. A hierarchical clustering strategy combining energy grading and graph connectivity clustering is employed. Its core advantage lies in achieving differentiated processing of water ripple noise. Strong attenuation is applied to strong noise regions to ensure sufficient noise suppression, while weak attenuation is applied to weak noise regions to avoid over-processing and damaging effective image information that may be adjacent to weak noise frequencies. This hierarchical processing strategy achieves a better balance between noise suppression and detail preservation compared to the scheme of applying uniform intensity attenuation to all noise regions. The center frequency coordinates and frequency domain span of the noise frequency domain clusters are calculated using a weighted area center and frequency domain span method. Its core advantage is that area weighting causes the center frequency coordinates to be biased towards the topology node region with a larger area, i.e., a denser noise frequency point, ensuring that the center of the notch filter is aligned with the location where the noise energy is most concentrated. The output of step S2, namely the hierarchical adaptive notch filter parameter table, provides complete parameter input for constructing the comprehensive notch filter transfer function in step S3. Without the hierarchical clustering and parameter allocation in step S2, step S3 will not be able to obtain the center frequency, bandwidth, and attenuation depth information of each noise region to be suppressed, and will not be able to construct a targeted notch filter. In step S2, the topological edge information of the water ripple noise frequency domain feature topology graph is used to search for connected components within the same level. This operation directly depends on the topology graph structure constructed in step S1. If step S1 only outputs a list of isolated noise frequency domain connected regions without establishing topological edges, step S2 will degenerate into a simple hierarchical classification based solely on the energy threshold, and will not be able to use spatial adjacency relationships for reasonable clustering. This will result in noise regions with similar frequency domain locations and originating from the same hardware root cause being split into multiple independent noise frequency domain clusters, each assigned an independent notch filter. Frequency domain coverage gaps may occur between the filters, resulting in incomplete noise suppression.

[0040] S3: Based on the hierarchical adaptive notch filter parameter table, hierarchical notch filtering and edge frequency domain protection are performed on the frequency domain representation of the infrared image to be processed to obtain the infrared image after water ripple suppression.

[0041] Further, step S3 includes: S31: Acquire the current frame infrared image captured in real time by the uncooled infrared camera module as the infrared image to be processed. After performing two-point non-uniform correction and blind pixel replacement on the infrared image to be processed, perform a two-dimensional discrete Fourier transform to obtain the current frame frequency domain complex matrix. The infrared image to be processed comes from the real-time video stream of the uncooled infrared camera module in actual application scenarios such as monitoring, security, search and rescue, or vehicle-mounted applications. Unlike the uniform scene used to construct the average power spectrum matrix in step S11, the infrared image to be processed contains effective image information such as targets, backgrounds, and edges. The size of the current frame frequency domain complex matrix is ​​consistent with the number of pixel rows and columns of the infrared image to be processed.

[0042] S32: Construct a comprehensive notch filter transfer function based on the hierarchical adaptive notch filter parameter table.

[0043] Further, step S32 includes: S321: Initialize an all-one matrix of the same size as the current frame's frequency domain complex matrix as the basic transfer function matrix. The initial value of each element in the basic transfer function matrix is ​​1, indicating that all frequency components pass through without attenuation before any notch filtering is applied.

[0044] S322: Traverse each filter parameter record in the hierarchical adaptive notch filter parameter table. For each filter parameter record, construct a Gaussian notch attenuation kernel with its center frequency coordinate as the center and the notch bandwidth parameter as the bandwidth radius. See also Figure 6 This is a schematic diagram of the integrated notch filter transfer function provided in this application embodiment. As shown in the figure, multiple Gaussian notch attenuation kernels are distributed in the frequency domain plane. Each attenuation kernel is represented by a concentric circle, with the solid dot at the center marked as the center frequency coordinate. The attenuation intensity gradually decreases from the center outwards, with the innermost dark area representing the strongest attenuation and the outermost dashed circle representing the attenuation boundary. The three attenuation kernels in the figure correspond to strong noise level, medium noise level, and weak noise level, respectively. The attenuation kernel for the strong noise level has the deepest central region, indicating the strongest attenuation, while the attenuation kernel for the weak noise level has the shallowest central region, indicating the weakest attenuation. Simultaneously, based on the frequency domain conjugate symmetry, identical attenuation kernels are constructed at symmetrical positions at the center point in the frequency domain, with dashed lines connecting them to indicate the symmetry. In infrared image processing of uncooled infrared cores, the frequency domain energy distribution of water ripple noise exhibits significant hierarchical differences. Applying a uniform attenuation intensity to all noise regions will result in insufficient suppression of strong noise regions or excessive suppression of weak noise regions. The integrated notch filter transfer function superimposes attenuation kernels of different noise levels in the frequency domain, achieving differentiated attenuation control for strong, medium, and weak noise regions. This ensures that each noise region is effectively suppressed while avoiding excessive attenuation that could harm adjacent effective frequency components.

[0045] S323: The Gaussian notch attenuation kernel corresponding to each filter parameter record is superimposed onto the basic transfer function matrix. Specifically, at the frequency domain position covered by the Gaussian notch attenuation kernel, the corresponding value of the basic transfer function matrix is ​​multiplied by the attenuation coefficient value of the Gaussian notch attenuation kernel. When the same frequency domain position is simultaneously covered by Gaussian notch attenuation kernels corresponding to multiple filter parameter records, the attenuation effect at that position is the product of the attenuation coefficient values ​​of each Gaussian notch attenuation kernel, achieving superimposed attenuation of multiple notch filters in the overlapping region of the frequency domain. After traversing all filter parameter records, the basic transfer function matrix is ​​updated to the composite notch filter transfer function. The composite notch filter transfer function is a real number matrix of the same size as the current frame's frequency domain complex matrix. Each element takes a value from 0 to 1. A value of 1 indicates that the frequency component passes through without attenuation, while a value close to 0 indicates that the frequency component is strongly suppressed.

[0046] S33: Perform edge frequency domain protection processing on the current frame frequency domain complex matrix and then apply the integrated notch filter transfer function to obtain the filtered frequency domain complex matrix.

[0047] Further, step S33 includes: S331: Perform spatial edge detection on the infrared image to be processed. The spatial edge detection uses the Sobel operator to calculate the gradient components of the infrared image in the horizontal and vertical directions respectively. Take the square root of the sum of the squares of the gradient components in the two directions to obtain the gradient magnitude map. Set an edge detection threshold for the gradient magnitude map, and mark the pixel positions in the gradient magnitude map that are greater than the edge detection threshold as edge pixels to obtain an edge binary mask. The edge detection threshold is determined by calculating the sum of the mean and standard deviation of the gradient magnitude of all pixels in the gradient magnitude map. For example, if the mean of the gradient magnitude map is 12.5 and the standard deviation is 8.3, then the edge detection threshold is 20.8. Perform a two-dimensional discrete Fourier transform on the edge binary mask to obtain an edge frequency domain mask matrix. Calculate the amplitude value of each frequency position in the edge frequency domain mask matrix, and form a set of edge-sensitive frequency domain positions with amplitude values ​​greater than a preset edge frequency energy threshold. The preset edge frequency energy threshold is determined by taking the 95th percentile value of the amplitude values ​​at all frequency positions in the edge frequency domain mask matrix. For example, the preset edge frequency energy threshold can be set as the value corresponding to the 95th percentile after sorting the amplitude values ​​of the edge frequency domain mask matrix. The physical meaning of the set of edge-sensitive frequency domain positions is that these frequency domain positions carry the main frequency energy of the target contour and scene edges in the infrared image to be processed. If notch attenuation is applied at these positions, the edge information will be weakened, resulting in blurred target contours and weakened scene edges.

[0048] S332: Check the synthetic notch filter transfer function for frequency domain locations that overlap with the set of edge-sensitive frequency domain locations. See also Figure 7This is a schematic diagram of edge frequency domain protection processing provided in an embodiment of this application. As shown in the figure, the infrared image to be processed on the left contains an infrared target represented by a simplified human silhouette. After spatial edge detection, an edge binary mask is obtained, which retains only the edge pixels of the target silhouette. After frequency domain transformation, the edge binary mask is projected onto the frequency domain plane on the right. The elliptical area filled with diagonal shading represents the edge-sensitive frequency domain position, and the gray circular area represents the notch attenuation area. The two partially overlap, and the overlapping area is marked with intersecting lines and labeled as corrected attenuation. In practical application scenarios such as security and search and rescue, operators rely on the clarity of the target silhouette in the infrared image for rapid identification and situation assessment. If the attenuation area of ​​the notch filter happens to cover the frequency position occupied by the target edge information, it will cause the target silhouette to be blurred, seriously affecting the real-time interpretation effect. The edge frequency domain protection mechanism detects the overlap between the notch attenuation region and the edge sensitive frequency domain position, and corrects the attenuation coefficient value at the overlapping position by adjusting the noise suppression strength at that position in exchange for the preservation of edge information. This achieves a controllable trade-off between suppressing water ripple noise and preserving the target edge.

[0049] S333: Multiply the current frame frequency domain complex matrix element by element with the modified integrated notch filter transfer function, that is, multiply the complex value at each frequency position in the current frame frequency domain complex matrix by the attenuation coefficient value of the modified integrated notch filter transfer function at the corresponding frequency position to obtain the filtered frequency domain complex matrix.

[0050] S34: Perform a two-dimensional discrete Fourier inverse transform on the filtered frequency domain complex matrix to obtain the spatial domain reconstructed image. The spatial domain reconstructed image is then truncated and normalized. Truncating involves setting pixel gray values ​​less than 0 to 0 and pixel gray values ​​greater than the upper limit of gray quantization for the uncooled infrared sensor to the upper limit. This upper limit is determined by the number of bits in the analog-to-digital converter (ADC) of the uncooled infrared sensor; for example, a 14-bit ADC corresponds to an upper limit of 16383. Normalization involves linearly mapping the gray values ​​of the truncated spatial domain reconstructed image to the grayscale display range of the display terminal; for example, mapping to an 8-bit grayscale range of 0 to 255. After truncating and normalizing, a water ripple-suppressed infrared image is obtained.

[0051] The overall function and technical effect of step S3 are analyzed as follows. Specifically, step S3 transforms the hierarchical adaptive notch filter parameter table output from step S2 into a comprehensive notch filter transfer function and applies it to the frequency domain representation of the infrared image to be processed, thus achieving frequency-domain directional suppression of water ripple noise. The filtering is performed using a Gaussian notch attenuation kernel with frequency domain multiplication instead of spatial domain convolution. Its core advantage lies in the fact that the Gaussian notch attenuation kernel acts directly on the noise frequency location in the frequency domain, enabling precise control of the attenuation frequency range and attenuation depth. This avoids frequency domain leakage caused by the finite length truncation of the spatial domain convolution filter, which leads to the erroneous attenuation of non-target frequencies. An edge frequency domain protection mechanism is employed. Its core advantage is that it solves the edge blurring problem that may occur when the attenuation region of the notch filter overlaps with the frequency domain distribution of the target edge information. By correcting the attenuation coefficient value at the overlapping location, the suppression of water ripple noise at that location is appropriately reduced in exchange for the preservation of edge information, achieving a controllable trade-off between noise suppression and edge preservation. If edge frequency domain protection processing is missing in step S3, the integrated notch filter transfer function will apply a preset attenuation depth to all noise frequency locations. When the noise frequency location overlaps with the frequency distribution of the target edge, the frequency energy of the target edge will be attenuated to the same extent, causing the target outline in the infrared image to become blurred after water ripple suppression, reducing the effectiveness of the infrared image for human observation and target detection. This is particularly critical for applications such as security and search and rescue that require rapid identification of target outlines. Step S3 directly relies on the hierarchical adaptive notch filter parameter table output in step S2 to obtain the center frequency coordinates, notch bandwidth parameters, and notch attenuation depth coefficients of each noise frequency domain cluster. If the hierarchical parameter allocation in step S2 is missing, step S3 will be unable to distinguish the energy intensity differences of different noise regions and can only apply a uniform attenuation depth to all noise regions, resulting in insufficient suppression of strong noise regions or excessive suppression of weak noise regions.

[0052] S4: Based on the infrared image after water ripple suppression and the infrared image to be processed, perform suppression effect evaluation and online correction of filter parameters, write the correction results back to the hierarchical adaptive notch filter parameter table and output the final display image.

[0053] Further, step S4 includes: S41: Based on the infrared image after water ripple suppression and the infrared image to be processed, calculate the water ripple residual index and the detail retention index.

[0054] Further, step S41 includes: S411: Perform a two-dimensional discrete Fourier transform on the infrared image after water ripple suppression to obtain the suppressed frequency domain complex matrix. Extract the amplitude value of the suppressed frequency domain complex matrix at the center frequency coordinate position corresponding to all filter parameter records in the hierarchical adaptive notch filter parameter table, and record this amplitude value as the residual amplitude value. Simultaneously, extract the amplitude value at the same center frequency coordinate position in the current frame's frequency domain complex matrix, and record this amplitude value as the original amplitude value. For each filter parameter record, calculate the quotient of the residual amplitude value at the corresponding center frequency coordinate position divided by the original amplitude value to obtain the single-point residual ratio of that filter parameter record. Calculate the arithmetic mean of the single-point residual ratios of all filter parameter records to obtain the water ripple residual index. The water ripple residual index ranges from 0 to 1, where a value closer to 0 indicates that the energy of the water ripple noise at the center frequency coordinate position is more thoroughly suppressed, and a value closer to 1 indicates that the residual energy of the water ripple noise is higher. For example, suppose the hierarchical adaptive notch filter parameter table contains 8 filter parameter records. The residual amplitude values ​​at the center frequency coordinates of each filter parameter record are 120, 85, 95, 200, 150, 70, 110, and 60, respectively, and the original amplitude values ​​are 2400, 1700, 1900, 4000, 3000, 1400, 2200, and 1200, respectively. Then the single-point residual ratios of each filter parameter record are 0.050, 0.050, 0.050, 0.050, 0.050, 0.050, and 0.050, respectively, and the water ripple residual index is 0.050.

[0055] S412: Calculate the structural similarity metric between the infrared image after water ripple suppression and the infrared image to be processed. The structural similarity metric is calculated as follows: Divide both the infrared image after water ripple suppression and the infrared image to be processed into sliding window regions of equal size, each sliding window region being 11 x 11 pixels with a sliding step of 1 pixel. For each sliding window position, calculate the pixel mean and pixel variance of the infrared image after water ripple suppression within that window, as well as the pixel mean and pixel variance of the infrared image to be processed within that window. Simultaneously, calculate the covariance of corresponding pixels in the two images within that window. Substitute the above mean, variance, and covariance into the structural similarity metric function, which is the product of the brightness comparison component, the contrast comparison component, and the structural comparison component. The brightness comparison component is determined by the relationship between the means of the two image windows, the contrast comparison component is determined by the relationship between the variances of the two image windows, and the structural comparison component is determined by the relationship between the covariance and variance of the two image windows. The arithmetic mean of the output values ​​of the structural similarity metrics function for all sliding window positions is used to obtain the global structural similarity metric, which is then used as the detail preservation index. The detail preservation index is a real number between 0 and 1. The closer the value is to 1, the higher the structural similarity between the infrared image after water ripple suppression and the infrared image to be processed, meaning that the filtering process causes less damage to the effective image details.

[0056] S42: Based on the water ripple residual index and detail retention index, determine the current filtering effect level and perform parameter correction to obtain the corrected graded adaptive notch filter parameter table. (See also...) Figure 8 This is a closed-loop diagram of online correction of filter parameters provided in this application embodiment. As shown in the figure, after the current frame is processed by notch filtering, the system calculates two evaluation quantities: water ripple residual index and detail preservation index. Both indices are input into the effect judgment stage. The effect judgment generates three branches based on the comparison relationship between the two indices and preset thresholds: the undersuppression branch performs enhanced attenuation operation, the oversuppression branch performs weakened attenuation and bandwidth reduction operation, and the qualified branch keeps the parameters unchanged. The correction results of the three branches are all stored in the parameter register. The updated parameters in the parameter register are fed back to the notch filtering stage of the current frame through the left closed loop for direct use in the next frame. In monitoring and security scenarios where uncooled infrared cores operate continuously for a long time, the core temperature will drift due to changes in ambient temperature and its own power consumption. Temperature drift will cause the energy level and frequency position of water ripple noise to change slowly. If the filter parameters remain fixed after initialization, they will gradually deviate from the optimal operating point, resulting in degradation of noise suppression effect or excessive loss of image details. This closed-loop correction mechanism, through frame-by-frame evaluation and parameter fine-tuning, enables the filter's attenuation depth and bandwidth to automatically adapt to dynamic changes in noise characteristics, maintaining stable filtering performance during continuous operation.

[0057] Further, step S42 includes: S421: Compare the water ripple residual index with a preset upper limit threshold, and simultaneously compare the detail retention index with a preset lower limit threshold. The upper limit threshold is determined based on the human eye's perception limit of water ripple noise in an infrared image. When the water ripple residual index is below this threshold, the human eye cannot perceive the water ripple noise at normal viewing distances. For example, the upper limit threshold can be set to 0.10 or 0.15. The lower limit threshold is determined based on the minimum acceptable structural fidelity of the infrared image in a target detection task. When the detail retention index is below this threshold, the detail loss introduced by filtering will affect the recognizability of the target contour. For example, the lower limit threshold can be set to 0.85 or 0.90.

[0058] S422: If the water ripple residual index is greater than the upper limit threshold and the detail retention index is greater than or equal to the lower limit threshold, the current filtering effect level is determined to be under-suppressed. Under-suppression means that the water ripple noise is not sufficiently suppressed, but the effective image details are well preserved, and the attenuation strength of the filter is insufficient. The notch attenuation depth coefficient of all filter parameters recorded in the hierarchical adaptive notch filter parameter table is increased by a preset depth adjustment step size. Specifically, this increase operation involves subtracting the preset depth adjustment step size from the notch attenuation depth coefficient in each filter parameter record to obtain a new notch attenuation depth coefficient. Since a smaller notch attenuation depth coefficient indicates stronger attenuation, reducing the notch attenuation depth coefficient is equivalent to increasing the attenuation depth. The preset depth adjustment step size is determined based on selecting the minimum adjustment amount that can eliminate the under-suppression state within a finite number of frames, while ensuring parameter convergence. For example, the preset depth adjustment step size can be set to 0.02 or 0.05. The increased notch attenuation depth coefficient must not be less than 0.01. If the calculated result is less than 0.01, it should be set to 0.01 to prevent the complete elimination of this frequency component from causing a frequency domain hole.

[0059] S423: If the water ripple residual index is less than or equal to the upper limit threshold of residual index and the detail retention index is less than the lower limit threshold of detail retention index, then the current filtering effect level is determined to be oversuppressed. Oversuppression means that the water ripple noise has been sufficiently suppressed, but the effective image details are excessively damaged, and the attenuation range or depth of the filter is too large. Reduce the notch attenuation depth coefficient of all filter parameters recorded in the hierarchical adaptive notch filter parameter table by a preset depth adjustment step. Specifically, the reduction operation involves adding the preset depth adjustment step to the notch attenuation depth coefficient in each filter parameter record to obtain a new notch attenuation depth coefficient. Increasing the notch attenuation depth coefficient is equivalent to reducing the attenuation depth. The increased notch attenuation depth coefficient must not exceed 0.95. If the calculated result is greater than 0.95, it is set to 0.95 to ensure that the notch filter still maintains a minimum attenuation effect on the noise frequency position. Simultaneously, reduce the notch bandwidth parameter of all filter parameters recorded in the hierarchical adaptive notch filter parameter table by a preset bandwidth adjustment step. Specifically, the reduction operation involves subtracting the preset bandwidth adjustment step from the notch bandwidth parameter in each filter parameter record to obtain a new notch bandwidth parameter. The preset bandwidth adjustment step size is determined based on selecting an adjustment amount that effectively reduces the range of accidental damage to edge frequencies, while ensuring that the notch filter can still cover the core region of the noise frequency domain cluster. For example, the preset bandwidth adjustment step size can be set to 1 frequency pixel or 2 frequency pixels. The reduced notch bandwidth parameter must not be less than half of the frequency domain span calculated in step S231. If the calculation result is less than half of the frequency domain span, it is set to half of the frequency domain span to ensure that the notch filter at least covers the core region of the noise frequency domain cluster.

[0060] S424: If the water ripple residual index is less than or equal to the upper limit threshold of residual index and the detail retention index is greater than or equal to the lower limit threshold of detail retention index, then the current filtering effect level is determined to be qualified. "Qualified" indicates that water ripple noise has been sufficiently suppressed and effective image details have been well preserved, and the current filtering parameters are at their optimal operating point. The hierarchical adaptive notch filter parameter table remains unchanged.

[0061] S43: The corrected hierarchical adaptive notch filter parameter table is stored in the parameter register of the image preprocessing module of the uncooled infrared sensor, for direct use in subsequent frames for water ripple noise frequency domain suppression. The parameter register is a designated address area in the on-chip memory of the uncooled infrared sensor, and each filter parameter record is written into the continuous address space of the parameter register according to a fixed byte format. When the uncooled infrared sensor acquires the next frame of infrared image, the processing flow in step S3 directly reads the corrected hierarchical adaptive notch filter parameter table from the parameter register for notch filtering, without repeating the offline analysis processes of steps S1 and S2. The infrared image after water ripple suppression is output to the display terminal as the final display image. The display terminal is a monitoring display screen, a vehicle-mounted infrared display screen, or a search and rescue handheld terminal display module. The final display image is presented to the operator for target observation and situation assessment via the display terminal.

[0062] The overall function and technical effect of step S4 are analyzed as follows. Specifically, step S4 constructs an online evaluation and parameter self-calibration closed loop for the water ripple noise suppression effect, enabling the hierarchical adaptive notch filter parameter table to be dynamically adjusted according to the actual filtering effect of each frame. The core advantage of using a dual-index joint evaluation method—water ripple residuality and detail preservation—lies in the fact that a single index cannot fully reflect the multidimensional characteristics of the filtering effect: using only the water ripple residuality index can only determine whether noise is sufficiently suppressed but cannot determine whether effective image information is excessively damaged; using only the detail preservation index can only determine whether image details are preserved but cannot determine whether noise still exists. The dual-index joint evaluation divides the filtering effect into three levels: under-suppression, over-suppression, and acceptable, covering all possible non-optimal operating states in the filter parameter space. Each state corresponds to a clear parameter correction direction, avoiding blind parameter tuning or parameter oscillation. The core advantage of employing parameter register write-back and direct call to subsequent frames lies in removing the offline analysis costs of steps S1 and S2 from the real-time processing chain. The uncooled infrared camera only needs to execute steps S1 and S2 once during initialization or periodic maintenance to obtain the initial hierarchical adaptive notch filter parameter table. In subsequent continuous video stream processing, only steps S3 and S4 for real-time processing and parameter fine-tuning are performed, significantly reducing the computational load of real-time processing. This allows the method to adapt to the performance constraints of miniaturization and low power consumption of uncooled infrared cameras. The parameter correction results in step S4 directly update the notch attenuation depth coefficient and notch bandwidth parameters in the hierarchical adaptive notch filter parameter table. These updated parameters are read by step S3 in the processing of the next frame of infrared image and used to construct a new comprehensive notch filter transfer function, forming a closed-loop feedback path between steps S3 and S4. This closed-loop feedback mechanism can address the slow changes in water ripple noise characteristics caused by temperature drift and other factors during long-term operation of the uncooled infrared sensor. When the water ripple noise energy increases due to temperature drift, step S4 detects an undersuppression state and increases the notch attenuation depth coefficient, allowing the filtering intensity of subsequent frames to automatically adapt to the new noise level. When the water ripple noise energy decreases, step S4 detects an oversuppression state and decreases the notch attenuation depth coefficient and notch bandwidth parameter, avoiding unnecessary damage to effective image details. Without the closed-loop correction mechanism in step S4, the hierarchical adaptive notch filter parameter table will remain fixed after initialization. As the uncooled infrared sensor operates for longer and the ambient temperature changes, the fixed filter parameters will gradually deviate from the optimal operating point, leading to a degradation in water ripple noise suppression or excessive loss of image details. Ultimately, this will affect the real-time interpretation capability of infrared images by operators in applications such as monitoring, security, and search and rescue.

[0063] For example, the following provides a complete workflow example to illustrate the actual execution process of steps S1 to S4 above. Assume the uncooled infrared sensor is a vanadium oxide uncooled focal plane array detector with a pixel size of 640 columns by 512 rows, an analog-to-digital conversion bit depth of 14 bits (i.e., a grayscale upper limit of 16383), and a frame rate of 25 frames per second. During the initialization phase, the lens of the uncooled infrared sensor is aimed at a uniform radiation temperature panel at 35 degrees Celsius, and 128 frames of raw infrared image sequence are continuously acquired at a frame rate of 25 frames per second. After performing two-point non-uniformity correction and blind pixel replacement on each of the 128 frames of raw infrared image sequence, a preprocessed infrared image sequence is obtained. Two-dimensional discrete Fourier transforms were performed on 128 frames of the preprocessed infrared image sequence to obtain 128 frequency domain complex matrices. The square of the amplitude spectrum of each frequency domain complex matrix was calculated to obtain 128 single-frame power spectrum matrices. The average power spectrum matrix (640 columns x 512 rows) was obtained by averaging these 128 single-frame power spectrum matrices along the frame dimension. Logarithmic transforms were then performed on the average power spectrum matrix to obtain the logarithmic power spectrum matrix. Local median filtering was then performed on the logarithmic power spectrum matrix using a 31x31 pixel window to obtain the background power spectrum matrix. The background power spectrum matrix was then subtracted pixel-by-pixel from the logarithmic power spectrum matrix to obtain the residual power spectrum matrix. The residual mean of the residual power spectrum matrix was calculated to be 0.15, and the standard deviation was 0.09. The adaptive significance threshold was 0.15 + 3 x 0.09 = 0.42. Frequency domain positions in the residual power spectrum matrix with a value greater than 0.42 were marked as candidate noise frequencies, resulting in approximately 350 candidate noise frequencies. Eight-connected-domain analysis was performed on 350 candidate noise frequency points to obtain 9 noise frequency domain connected regions. The centroid frequency coordinates, area, major axis direction angle, and average energy value of each noise frequency domain connected region were calculated. Using the 9 noise frequency domain connected regions as topological nodes, with a preset frequency domain adjacency radius of 40 frequency pixels, the Euclidean distance between all pairs of topological nodes was checked. Among them, 4 pairs had a distance of less than 40 frequency pixels, and 4 topological edges were established to form a water ripple noise frequency domain feature topology map. The average energy value of the 9 topological nodes was classified using the Otsu's method, resulting in a first energy boundary threshold of 25000 and a second energy boundary threshold of 7000. Two topological nodes were classified into the strong noise level, three into the medium noise level, and four into the weak noise level. Connected component search was performed within each level to obtain a total of 7 noise frequency domain clusters. For each noise frequency domain cluster, the center frequency coordinates and frequency domain span are calculated. Notch attenuation depth coefficients and notch bandwidth parameters are assigned according to the noise level, generating a hierarchical adaptive notch filter parameter table containing 7 filter parameter records. Upon entering the real-time processing stage, the uncooled infrared sensor acquires the current frame infrared image in the security scenario as the infrared image to be processed. After performing two-point non-uniform correction and blind pixel replacement on the infrared image to be processed, a two-dimensional discrete Fourier transform is performed to obtain the current frame's frequency domain complex matrix.A 640-column by 512-row all-one basic transfer function matrix is ​​initialized. A Gaussian notch attenuation kernel is constructed by traversing seven filter parameter records and then superimposed to obtain the composite notch filter transfer function. Sobel edge detection is performed on the infrared image to be processed to obtain a binary edge mask. A two-dimensional discrete Fourier transform is performed on the binary edge mask to obtain an edge frequency domain mask matrix. The set of edge-sensitive frequency domain positions is extracted, and the overlap with the attenuation region of the composite notch filter transfer function is checked and corrected with an edge protection callback ratio of 0.6 to obtain the corrected composite notch filter transfer function. The current frame's frequency domain complex matrix is ​​multiplied element-wise with the corrected composite notch filter transfer function to obtain the filtered frequency domain complex matrix. A two-dimensional inverse discrete Fourier transform and grayscale truncation and normalization are performed to obtain the infrared image after water ripple suppression. After suppressing the water ripples, frequency domain analysis was performed on the infrared image to calculate the water ripple residual index, which was 0.08. The structural similarity metric, i.e., the detail retention index, was calculated to be 0.92. The water ripple residual index of 0.08 was compared with the upper limit threshold of 0.10, and 0.08 was less than 0.10. The detail retention index of 0.92 was compared with the lower limit threshold of 0.85, and 0.92 was greater than 0.85. The current filtering effect level was determined to be qualified. The hierarchical adaptive notch filter parameter table was kept unchanged and written into the parameter register. The infrared image after water ripple suppression was then output to the security monitoring display screen as the final display image.

[0064] Example 2: This embodiment, based on Embodiment 1, provides an infrared thermal imaging water ripple noise frequency domain suppression system, such as... Figure 9 As shown, it includes: Water ripple noise frequency domain feature topology map construction module: used to perform time domain accumulation and frequency domain transformation on the raw infrared image sequence acquired by the uncooled infrared core to obtain the average power spectrum matrix, and construct the water ripple noise frequency domain feature topology map based on the average power spectrum matrix; Hierarchical adaptive notch filter parameter table generation module: Based on the frequency domain feature topology map of water ripple noise, it performs hierarchical clustering of noise regions and allocation of filter kernel parameters to generate hierarchical adaptive notch filter parameter tables. Water ripple suppression infrared image acquisition module: Based on the hierarchical adaptive notch filter parameter table, it performs hierarchical notch filtering and edge frequency domain protection on the frequency domain representation of the infrared image to be processed to obtain the water ripple suppression infrared image; Suppression effect evaluation and parameter correction module: Based on the infrared image after water ripple suppression and the infrared image to be processed, it performs suppression effect evaluation and online correction of filter parameters, writes the correction results back to the hierarchical adaptive notch filter parameter table and outputs the final display image.

Claims

1. A method for frequency domain suppression of water ripple noise in infrared thermal imaging, characterized in that, The method includes: S1: Obtain a sequence of K frames of raw infrared images continuously acquired by an uncooled infrared sensor in a uniform scene. Perform non-uniform correction and blind pixel replacement on each frame of the raw infrared image sequence to obtain a preprocessed infrared image sequence. Perform a two-dimensional discrete Fourier transform on each frame of the preprocessed infrared image sequence to obtain the frequency domain complex matrix corresponding to each frame. Calculate the square of the amplitude spectrum of the frequency domain complex matrix of each frame. Then, calculate the mean of the squares of the amplitude spectra of all frames along the frame dimension to obtain the average power spectrum matrix. Constructing a frequency domain feature topology map of water ripple noise based on the average power spectrum matrix includes: performing a logarithmic transformation to the base of natural numbers on each element of the average power spectrum matrix to obtain a logarithmic power spectrum matrix; performing local median filtering on the logarithmic power spectrum matrix to obtain a background power spectrum matrix; subtracting the background power spectrum matrix pixel-by-pixel from the logarithmic power spectrum matrix to obtain a residual power spectrum matrix; setting an adaptive significance threshold on the residual power spectrum matrix; marking frequency domain positions in the residual power spectrum matrix that are greater than the adaptive significance threshold as candidate noise frequencies; and performing operations on all candidate noise frequencies in the frequency domain coordinate space. Eight-connected domain analysis aggregates spatially adjacent candidate noise frequency points into noise frequency domain connected regions. For each noise frequency domain connected region, region attribute parameters are calculated, including the region centroid frequency coordinates, region area, region major axis direction angle, and region average energy value. Each noise frequency domain connected region is used as a topological node, and the region attribute parameters are used as node attributes. Topological edges are established between topological nodes whose frequency domain distance is less than a preset frequency domain adjacency radius. Each topological edge carries the ratio of the region average energy values ​​of the topological nodes at both ends as the edge weight, forming a water ripple noise frequency domain feature topological map. S2: Based on the frequency domain feature topology of water ripple noise, perform hierarchical clustering of noise regions and allocation of filter kernel parameters to generate a hierarchical adaptive notch filter parameter table; S3: Based on the hierarchical adaptive notch filter parameter table, hierarchical notch filtering and edge frequency domain protection are performed on the frequency domain representation of the infrared image to be processed to obtain the infrared image after water ripple suppression. S4: Based on the infrared image after water ripple suppression and the infrared image to be processed, perform suppression effect evaluation and online correction of filter parameters, write the correction results back to the hierarchical adaptive notch filter parameter table and output the final display image.

2. The infrared thermal imaging water ripple noise frequency domain suppression method according to claim 1, characterized in that, The step of obtaining the preprocessed infrared image sequence includes: The lens of the uncooled infrared sensor is aimed at the uniform radiation temperature panel. The uncooled infrared sensor continuously acquires K frames of infrared images at a fixed frame rate in front of the uniform radiation temperature panel. The K frames of infrared images are arranged in chronological order of acquisition to form the original infrared image sequence. Two-point non-uniformity correction is performed on each frame of the original infrared image sequence. After the two-point non-uniformity correction is completed, blind pixel replacement is performed on each corrected frame. The K frames after the two-point non-uniformity correction and blind pixel replacement are arranged in the original time order to obtain the preprocessed infrared image sequence.

3. The infrared thermal imaging water ripple noise frequency domain suppression method according to claim 1, characterized in that, The step of generating the hierarchical adaptive notch filter parameter table includes: Extract the regional attribute parameters of all topological nodes from the frequency domain feature topology map of water ripple noise, and sort all topological nodes in descending order according to the regional average energy value to obtain an energy sorted node list; Perform hierarchical clustering based on graph connectivity on the topological nodes in the energy sorting node list to generate hierarchical clustering results for noisy regions; Based on the hierarchical clustering results of the noise regions, corresponding notch filter kernel parameters are assigned to each noise frequency domain cluster, generating a hierarchical adaptive notch filter parameter table.

4. The infrared thermal imaging water ripple noise frequency domain suppression method according to claim 3, characterized in that, The steps for generating hierarchical clustering results for noisy regions include: The energy levels are set to three: strong noise, medium noise, and weak noise. Based on the distribution of the regional average energy values ​​of all topological nodes in the energy sorting node list, two energy boundary thresholds are calculated using the maximum inter-class variance method. Topological nodes with regional average energy values ​​greater than the first energy boundary threshold are classified as strong noise, topological nodes with regional average energy values ​​greater than the second energy boundary threshold and less than or equal to the first energy boundary threshold are classified as medium noise, and topological nodes with regional average energy values ​​less than or equal to the second energy boundary threshold are classified as weak noise. Within the same noise level, traverse the topological nodes and their edges belonging to that level in the water ripple noise frequency domain feature topology graph. Merge the topological nodes directly connected by the topological edges into the same noise frequency domain cluster. Each noise frequency domain cluster inherits the noise level label to which it belongs. After completing the above connected component search in the three levels of strong noise, medium noise and weak noise respectively, all noise frequency domain clusters constitute the hierarchical clustering result of the noise region.

5. The infrared thermal imaging water ripple noise frequency domain suppression method according to claim 1, characterized in that, The steps for obtaining the infrared image after water ripple suppression include: The current frame infrared image acquired in real time by the uncooled infrared sensor is used as the infrared image to be processed. After performing two-point non-uniform correction and blind element replacement on the infrared image to be processed, a two-dimensional discrete Fourier transform is performed to obtain the current frame frequency domain complex matrix. Based on the hierarchical adaptive notch filter parameter table, a comprehensive notch filter transfer function is constructed. After performing edge frequency domain protection processing on the current frame frequency domain complex matrix, a comprehensive notch filter transfer function is applied to obtain the filtered frequency domain complex matrix. A two-dimensional discrete Fourier inverse transform is performed on the filtered frequency domain complex matrix to obtain a spatial domain reconstructed image. The spatial domain reconstructed image is then subjected to grayscale value truncation and normalization to obtain an infrared image with water ripple suppression.

6. The infrared thermal imaging water ripple noise frequency domain suppression method according to claim 5, characterized in that, The steps for constructing the synthetic notch filter transfer function include: Initialize an all-one matrix of the same size as the current frame's frequency domain complex matrix as the base transfer function matrix; Traverse each filter parameter record in the hierarchical adaptive notch filter parameter table. For each filter parameter record, construct a Gaussian notch attenuation kernel with its center frequency coordinate as the center and the notch bandwidth parameter as the bandwidth radius. At the same time, based on the frequency domain conjugate symmetry of the two-dimensional discrete Fourier transform, construct the same Gaussian notch attenuation kernel at the conjugate symmetry position of the complex matrix in the frequency domain of the current frame. The Gaussian notch attenuation kernel corresponding to each filter parameter record is superimposed onto the basic transfer function matrix. After traversing all filter parameter records, the basic transfer function matrix is ​​updated to the comprehensive notch filter transfer function.

7. The infrared thermal imaging water ripple noise frequency domain suppression method according to claim 1, characterized in that, The step of writing the correction result back to the hierarchical adaptive notch filter parameter table and outputting the final display image includes: Based on the infrared image after water ripple suppression and the infrared image to be processed, the water ripple residual index and the detail retention index are calculated. Based on the water ripple residual index and detail retention index, the current filtering effect level is determined and parameter correction is performed to obtain the corrected graded adaptive notch filter parameter table. The corrected hierarchical adaptive notch filter parameter table is stored in the parameter register of the image preprocessing module of the uncooled infrared core, and the infrared image after water ripple suppression is output to the display terminal as the final display image.

8. An infrared thermal imaging water ripple noise frequency domain suppression system, used to implement the infrared thermal imaging water ripple noise frequency domain suppression method according to any one of claims 1-7, characterized in that, The system includes: The water ripple noise frequency domain feature topology map construction module is used to acquire a K-frame raw infrared image sequence continuously collected by an uncooled infrared sensor in a uniform scene. Non-uniform correction and blind pixel replacement are performed on each frame of the raw infrared image sequence to obtain a preprocessed infrared image sequence. A two-dimensional discrete Fourier transform is performed on each frame of the preprocessed infrared image sequence to obtain the corresponding frequency domain complex matrix. The square of the amplitude spectrum of the frequency domain complex matrix of each frame is calculated, and then the mean of the squares of the amplitude spectra of all frames along the frame dimension is calculated to obtain the average power spectrum matrix. A logarithmic transformation with a base of natural numbers is performed on each element of the average power spectrum matrix to obtain the logarithmic power spectrum matrix. Local median filtering is performed on the logarithmic power spectrum matrix to obtain the background power spectrum matrix. The background power spectrum matrix is ​​subtracted from the logarithmic power spectrum matrix pixel by pixel to obtain the residual. Power spectrum matrix; an adaptive significance threshold is set for the residual power spectrum matrix, and the frequency domain positions in the residual power spectrum matrix that are greater than the adaptive significance threshold are marked as candidate noise frequency points; an eight-connected domain analysis is performed on all candidate noise frequency points in the frequency domain coordinate space, and spatially adjacent candidate noise frequency points are aggregated into noise frequency domain connected regions. For each noise frequency domain connected region, regional attribute parameters are calculated, including the region centroid frequency coordinates, region area, region major axis direction angle, and region average energy value; each noise frequency domain connected region is used as a topological node, and the regional attribute parameters are used as node attributes. Topological edges are established between topological nodes whose frequency domain distance is less than the preset frequency domain adjacency radius. Each topological edge carries the ratio of the regional average energy values ​​of the topological nodes at both ends as the edge weight, forming a water ripple noise frequency domain feature topological map; Hierarchical adaptive notch filter parameter table generation module: Based on the frequency domain feature topology map of water ripple noise, it performs hierarchical clustering of noise regions and allocation of filter kernel parameters to generate hierarchical adaptive notch filter parameter tables. Water ripple suppression infrared image acquisition module: Based on the hierarchical adaptive notch filter parameter table, it performs hierarchical notch filtering and edge frequency domain protection on the frequency domain representation of the infrared image to be processed to obtain the water ripple suppression infrared image; Suppression effect evaluation and parameter correction module: Based on the infrared image after water ripple suppression and the infrared image to be processed, it performs suppression effect evaluation and online correction of filter parameters, writes the correction results back to the hierarchical adaptive notch filter parameter table and outputs the final display image.

Images