Human body infrared image small target detection method based on improved FGLCM features

Abstract

The invention discloses a human body infrared image small target detection method based on an improved FGLCM feature, and relates to the technical field of image processing and target detection, and the method comprises the following steps: building a boundary singularity auditing baseline under a unified time baseline, carrying out the multi-scale energy mapping of a curved surface fitting residual error of a human body infrared image, and generating a traction residual error distribution diagram; and constructing a reflection pseudo peak discriminator based on the traction residual distribution diagram, and extracting a pseudo peak kernel position by combining polarization sensitivity estimation and view angle transformation consistency constraint. According to the method, a closed-loop self-adaptive regulation and control mechanism is constructed through residual distribution auditing, pseudo peak identification, gradient registration, differential entropy enhancement and time sequence threshold adjustment, fitting abnormity and false highlight spots in the infrared image are inhibited, the accuracy and stability of small target detection are improved, and the method is suitable for a complex photo-thermal environment.

Description

Technical Field

[0001] This invention relates to the field of image processing and target detection technology, specifically to a method for detecting small targets in human infrared images based on improved FGLCM features. Background Technology

[0002] Small Target Detection in Human Infrared Images Based on Improved FGLCM Features is an intelligent detection method that utilizes improved gray-level co-occurrence matrix texture features (FGLCM, Facet-fitted gray-level co-occurrence matrix) to identify and locate subtle abnormal regions or small heat sources in human infrared images. This method first smoothly models local regions of the infrared image through multi-scale Facet surface fitting, based on bivariate cubic polynomials and Chebyshev orthogonal polynomials, effectively suppressing background noise and preserving the structure of target edges. Subsequently, the horizontal and vertical gradients of the fitting results are calculated separately, and the bidirectional gradients are fused to generate a high-response gradient map, significantly enhancing the contour features of small targets. Based on this, the algorithm performs improved GLCM calculations on the gradient image, constructing a local contrast index by selecting several pixels with the highest gray values ​​within a sliding window, improving sensitivity and noise resistance for low-contrast targets. Finally, an adaptive threshold segmentation strategy based on the mean and variance of the entire image is combined to dynamically discriminate and accurately extract candidate regions. This method can not only stably detect small-sized thermal anomalies (such as local inflammation, tumor hotspots, and abnormal body temperature areas) in human infrared images under complex backgrounds or lighting conditions, but also has low algorithm complexity, does not require a large number of training samples, and can be deployed in medical imaging equipment, security monitoring terminals, and mobile diagnostic systems to achieve efficient and low-false-alarm intelligent detection of human thermal features.

[0003] The existing technology has the following shortcomings: In existing technologies, infrared image-based facet fitting methods typically rely on local surface modeling to smooth image grayscale levels, achieving background suppression and target enhancement. However, in dynamic imaging or complex photothermal environments, if there are abrupt temperature changes, reflection spurious peaks, or local overexposure in the boundary regions of the infrared image, the polynomial surface model in the facet fitting stage will be strongly disturbed by abnormal grayscale points, causing distortion of the local fitting surface. This distortion is amplified into bright abnormal spots during subsequent gradient inversion, forming fake small target signals that do not exist in the real scene. When these false signals are input into the improved FGLCM feature extraction stage, the system misidentifies them as high-confidence targets, leading to serious misjudgments in the detection results, thus affecting the accuracy of human infrared small target detection and the reliability of diagnostic and security systems. This indicates that existing technologies lack a mechanism to suppress abnormal response in facet fitting under complex temperature change conditions at the boundary.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting small targets in human infrared images based on improved FGLCM features, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting small targets in human infrared images based on improved FGLCM features, comprising the following steps: A boundary singularity audit baseline is established under a unified time baseline, and multi-scale energy mapping is performed on the surface fitting residuals of human infrared images to generate a traction residual distribution map. A reflection pseudo-peak discriminator is constructed based on the traction residual distribution map. The positions of high-risk pseudo-peak nuclei are extracted by combining polarization sensitivity estimation and viewpoint transformation consistency constraints. A gradient orthogonal registration chain is established based on the pseudo-peak kernel position, and phase conjugate registration is performed on the horizontal and vertical gradients of the image to generate an edge evidence map. Based on the edge evidence map, a gray-level co-occurrence matrix feature structure is constructed, and a joint mechanism of extreme value triplet ranking statistics and dual neighborhood differential entropy is adopted to generate a confidence contrast spectrum. A cross-frame adaptive thresholding engine is built based on confidence contrast spectrum, and a dynamic balanced threshold is generated using a temporal thresholding mechanism based on mean and standard deviation. Based on the dynamic equilibrium threshold, multi-field coupling dynamic control is performed. The surface fitting weight and threshold gain are adjusted by the same frequency driving method of optical phase injection and thermal radiation perturbation to control the detection process.

[0007] Preferably, the steps for generating the traction residual distribution map are as follows: Perform grayscale consistency calibration and background drift correction, normalize the grayscale of the infrared image, and use the average grayscale of the image edge region as the background temperature reference to complete the grayscale difference compensation of the whole image. Perform local surface fitting and extract the fitting residual. Use a fixed-size sliding window to perform polynomial fitting, calculate the gray-level difference between each pixel and the fitted surface, and generate a fitting residual image. Multi-scale energy mapping is constructed based on the fitted residual image. Fine-scale, medium-scale and wide-scale residual energy maps are extracted and weighted and fused to identify energy anomaly regions and form a binary mask. Perform counterfactual playback on the energy anomaly region to generate two sets of perturbation images, extract stable residual pixels and complete clustering, and construct a traction residual distribution map.

[0008] The preferred method for extracting the high-risk pseudo-peak core position is as follows: Based on the traction residual distribution map, high traction residual blocks are extracted and regions with local brightness anomalies and multi-directional gradient abrupt changes are identified. Polarization sensitivity estimation is performed on the above-mentioned region using dual-view images, the gray-level difference between the two views is calculated and the region with strong polarization response is extracted; Perform viewpoint consistency constraint detection on the same region to identify regions that are inconsistent in grayscale difference and gradient direction; The polarization sensitivity image and the viewpoint consistency mask image are weighted and fused to generate a physical confidence spectrum. Gaussian smoothing and connected component clustering are then performed to extract the pixel with the lowest confidence as the location of the high-risk pseudo-peak kernel.

[0009] Preferably, the steps for generating the edge evidence map are as follows: Based on the location of high-risk pseudo-peak kernels, the infrared image fitted to the surface is subjected to initial extraction and orientation concentration of horizontal and vertical gradients, and the orientation angle and amplitude in the neighborhood are corrected. An orthogonal registration chain is constructed based on the entire image domain, and phase conjugate registration is performed on regions with inconsistent directions to unify the local gradient direction distribution; By fusing the corrected horizontal and vertical gradient information, non-maximum suppression and double-threshold edge extraction are performed to generate a stable edge evidence map for subsequent texture feature extraction and target localization.

[0010] Preferably, when performing phase conjugate registration, the gradient direction adjustment is based on the cosine of the angle between the principal direction angle and the original direction angle. When the angle exceeds a set threshold, the original direction is adjusted to be conjugate with the principal direction, and the difference in direction angle does not exceed a preset range, so as to ensure the continuity and consistency of the local gradient direction.

[0011] Preferably, the confidence contrast spectrum generation steps are as follows: Based on pixels with high edge intensity in the stable edge evidence map, a direction-aware gray-level co-occurrence matrix structure is established in the original gray-level image. The gray-level co-occurrence relationship is statistically analyzed in multiple fixed directions, and directional weights are assigned according to gradient intensity. Local extreme gray-level pairs are extracted from the direction-weighted gray-level co-occurrence matrix, extreme triplet ranking statistics are performed, the consistency of sorting in each direction is analyzed, and the structural consistency index is calculated to characterize the stability of regional texture. A dual neighborhood differential entropy joint mechanism is established based on the structural consistency index. The gray-level differential entropy in the horizontal and vertical directions is calculated separately, and the direction sensitivity and consistency index are fused to generate a confidence contrast spectrum. After smoothing and normalization, the output is used as the input basis for subsequent threshold discrimination.

[0012] Preferably, the dynamic equilibrium threshold generation steps are as follows: Based on the confidence contrast spectrum, the confidence contrast values ​​in the current frame and multiple historical frames are extracted, and the temporal mean and standard deviation of each pixel are calculated to form a temporal mean image and a standard deviation image. Construct a sensitivity feedback structure related to cross-frame residual fluctuations, calculate the residual fluctuation index based on the confidence contrast difference between the current frame and the previous frame of the candidate target region, and adjust the detection threshold of the current frame according to the fluctuation changes. By fusing temporal statistics and residual feedback results, a dynamic equilibrium threshold map of the current frame is generated. The confidence contrast spectrum map and the dynamic threshold map are compared pixel by pixel. Combined with temporal consistency verification and connected component analysis, a high-confidence target mask image is output.

[0013] Preferably, multi-field coupling dynamic control is performed based on a dynamic equilibrium threshold. The surface fitting weight and threshold gain are adjusted by a synchronous driving method of optical phase injection and thermal radiation perturbation. The detection process steps are as follows: Based on the dynamic equilibrium threshold map and confidence contrast spectrum, false bright regions without temporal continuity in the current frame are identified, and high-risk false bright spots are determined through neighborhood structure analysis. A phase compensation map is established based on the optical incident angle and phase perturbation. Position-specific weights are injected into the fitted surface to suppress abnormal curvature response. Extract the thermal intensity trajectory and calculate the thermal perturbation residual. Inject perturbation terms into abnormal regions to reduce the reliability of their fitted response. A three-variable coupling matrix consisting of optical phase perturbation amplitude, thermal response residual, and threshold gain is constructed, and the fitting curvature and threshold gain strategies are adjusted synchronously. The fitting adjustment plot and the threshold adjustment plot are used together to update the detection process input, complete the update of the detection results and store the adjustment parameters, and construct a closed-loop self-healing control structure.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention achieves accurate identification and structural mapping of fitting anomaly regions by introducing a traction residual distribution map; it establishes a false peak identification mechanism by combining polarization sensitivity and viewpoint consistency constraints, improving the resolution of false bright spots; it ensures the consistency of edge features and the robustness of texture expression through gradient orthogonal registration and differential entropy enhancement strategies; it introduces a cross-frame temporal threshold mechanism and a residual sensitive feedback loop to achieve temporal dynamic stability of detection conditions; and it constructs a dynamic coupling control strategy through optical phase injection and thermal radiation perturbation driving, completing the synchronous adaptive adjustment of surface fitting weights and detection threshold gain, enabling the detection process to have automatic identification, dynamic response, self-repair, and closed-loop control capabilities, effectively suppressing false targets and improving detection accuracy and robustness, especially suitable for infrared image environments with significant photothermal interference, complex structures, and weak targets. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method for detecting small targets in human infrared images based on improved FGLCM features according to the present invention.

[0017] Figure 2 This is a statistical chart of the small target detection results of the FGLCM of this invention. Detailed Implementation

[0018] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0019] This invention provides, for example Figure 1 The method for detecting small targets in human infrared images based on improved FGLCM features, as shown, includes the following steps: A boundary singularity audit baseline is established under a unified time baseline. Multi-scale energy mapping is performed on the surface fitting residuals of human infrared images to obtain regions of anomalous energy concentration. A traction residual distribution map is generated based on counterfactual playback. To address the distortion in surface fitting of human infrared images under complex photothermal environments caused by reflection interference, local overexposure, or sudden temperature changes, a boundary singularity auditing method based on a unified temporal baseline is proposed. This method identifies and extracts abnormal distributions of fitting residuals caused by anomalous gray levels in a single-frame infrared image, providing a structured energy reference for subsequent pseudo-peak identification and feature extraction. This method completes the entire process of background balancing, residual calculation, energy aggregation mapping, and counterfactual verification using only a single-frame infrared image, without relying on image sequences. The specific implementation steps are as follows: The input human infrared image undergoes grayscale consistency calibration and background drift correction. The input infrared image is converted into a single-channel grayscale matrix, and the infrared radiation intensity value of each pixel is read. The mean and standard deviation of the grayscale value across the entire image are calculated. Using the mean grayscale value as the brightness balance benchmark, linear grayscale normalization is performed, mapping all pixel grayscale values ​​to a standard range between zero and one to enhance the dynamic contrast of the image. Subsequently, a background reference ring is established in the edge region of the image, with a ring width set to 10% of the image width. The average grayscale value of pixels within this region is calculated as the background temperature benchmark. Differential compensation is performed on the entire image using this benchmark value, adjusting the grayscale value of each pixel to maintain a balanced thermal distribution of the overall image under the background temperature reference. This step effectively eliminates global brightness drift caused by uneven ambient temperature or thermal noise from the imaging device, ensuring the stability of the subsequent fitting process.

[0020] Local surface fitting was performed on the calibrated infrared image to extract the fitting residuals. The entire image was divided into several overlapping regions, each a 31-pixel by 31-pixel sliding window with a sliding step of five pixels, ensuring continuous coverage of adjacent regions. For each pixel within the window, a bivariate cubic polynomial surface was fitted using the least squares method, with its coordinate position as the independent variable and its gray value as the dependent variable. The fitting equation included first, second, and third terms for the x and y coordinates, as well as an interaction term, accurately reflecting the local gray-level variation trend. After fitting, the difference between the actual gray value of each pixel within the window and the corresponding gray value on the fitted surface was calculated, and this difference was defined as the fitting residual. For window regions near the image boundary, due to insufficient fitting data, the sliding window was expanded inward to 1.5 times the original window area to ensure that the edge structure was fully included in the calculation. Finally, after traversing the entire image, a complete fitting residual image was generated. The residual value of each pixel represents the magnitude of its gray-level deviation from the local surface trend, reflecting the energy change characteristics of potential anomalous regions.

[0021] Multi-scale energy mapping analysis was performed on the generated residual images to reveal energy concentration characteristics at different spatial scales. Three neighborhood analysis structures were constructed for fine, medium, and wide scales. For the fine scale, a 3x3 neighborhood was used, summing the absolute values ​​of the residuals within each pixel's neighborhood to quantify the concentration of local residual signals. For the medium scale, a 7x7 neighborhood was used, summing the absolute values ​​of the residuals of each pixel within the neighborhood inversely proportional to their Euclidean distance from the center pixel, ensuring the center point's influence is dominant. For the wide scale, an 11x11 neighborhood was used, summing the squared residual values ​​to describe the diffusion of the residual signal over a larger spatial range. The energy response maps generated at each of the three scales were normalized to ensure that the grayscale values ​​of each image were distributed between zero and one. Then, a weighted fusion was performed with a weight ratio of 5:3:2 for the fine, medium, and wide scales to generate a comprehensive energy mapping map. Histogram analysis was performed on this mapping map to calculate its average energy value and standard deviation. Pixel regions with energy values ​​higher than the average plus twice the standard deviation were defined as abnormal energy concentration areas. By using connectivity detection, all high-energy regions are extracted and formed into a binary mask to represent the potential anomalous response distribution. Each anomalous region corresponds to a thermal anomaly region in the image where gray-level fitting errors are concentrated, typically located at boundaries or at points of reflection interference.

[0022] Counterfactual playback is performed on the extracted anomalous energy concentration areas to generate a traction residual distribution map. For each anomalous energy area, its centroid pixel position in the original image is extracted, and a local region is established at this position with a radius set to twice the average radius of the anomalous energy area. Within this region, two sets of counterfactual data are generated according to the sign direction of the residual values: one set amplifies the residual values ​​by a factor of two along the original direction and diffuses them to a five-pixel neighborhood centered on the current pixel; the other set inverts the residual values ​​and diffuses them in the same way. The two sets of data are superimposed on the original grayscale image to form two counterfactual grayscale images under different assumptions. The gradient rate of change distribution of the two counterfactual images in the same region is calculated, and pixels that maintain a significant gradient response under both perturbation conditions are marked as stable residual pixels. These stable residual pixel groups are analyzed by spatial clustering, and the center positions of each cluster are extracted as high traction residual points. Based on the spatial distribution of these high traction residual points, a traction residual distribution map is generated to record the stability of the residual signal of each pixel under counterfactual perturbation, reflecting the ability of the physical structure of the local anomalous response to be preserved. The traction residual distribution map visually reveals the difference between the real structural thermal anomaly region and the pseudo-anomaly caused by random noise or light reflection, providing an accurate energy space reference for subsequent pseudo-peak identification and feature extraction.

[0023] A reflection pseudo-peak discriminator is constructed based on the traction residual distribution map. A physical confidence spectrum is generated by using polarization sensitivity estimation and viewpoint transformation consistency constraints, and the positions of high-risk pseudo-peak nuclei are extracted accordingly. The previous step generated the traction residual distribution map, which is used to identify the concentrated areas of fitted residuals in the infrared image. To further identify non-real heat source responses in the image that may be caused by reflection, scattering, or optical artifacts, a method combining polarization sensitivity estimation and viewpoint consistency constraints is proposed to construct a reflection pseudo-peak discriminator and ultimately extract the locations of high-risk pseudo-peak nuclei. The specific steps are as follows: Based on the generated traction residual distribution map, potential candidate regions for reflectivity are extracted. All pixels in the traction residual distribution map are scanned, and the mean and standard deviation of the residuals across the entire image are calculated. Pixels with residual values ​​greater than the mean plus twice the standard deviation are selected as initial candidate points. For each initial candidate point, an 11×11 pixel local neighborhood window is established centered on that point, and the average residual intensity of all pixels within that neighborhood is calculated. If the average residual intensity of this region is still higher than the average residual level of the entire image, the neighborhood is defined as a high traction residual block. For each high traction residual block, the grayscale distribution of its corresponding location is extracted from the original infrared image, and the grayscale difference between the brightest and darkest pixels within that region is calculated. If this grayscale difference exceeds 30% of the overall dynamic range of the image, the region is considered to have a strong local brightness anomaly. Furthermore, statistical analysis of the gray-scale gradient direction at the edge of the region reveals that if there are significant gradient jumps in more than six directions within a 360-degree range, and the angle differences between these directions are discontinuous, then the region is considered to exhibit an irregular structural abrupt change trend and is a potential source of reflection interference.

[0024] For the selected potential reflection regions, a dual-view image acquisition method is used to perform polarization sensitivity estimation. Two viewpoint images of the current frame are acquired using an infrared camera, with the angle between the two viewpoints set to 7 degrees to ensure a slight difference in the line-of-sight direction without distortion. Conformal feature points in each image are selected using a feature point matching algorithm, and sub-pixel-level registration is performed to ensure pixel-by-pixel alignment of the two images in the same coordinate system. For each potential reflection region, the corresponding grayscale value is extracted from both viewpoint images, and the grayscale difference between the two viewpoints is calculated pixel-by-pixel and recorded as the polarization intensity difference of that pixel. Then, the polarization intensity difference of all pixels in the region is selected, and its standard deviation and maximum value are calculated. If the standard deviation exceeds a preset threshold (e.g., 1.5 times the standard deviation of the original image), or the maximum value exceeds 15% of the difference in grayscale values ​​between the two viewpoints, it indicates that the region has a significant difference in radiation response between the two viewpoints, exhibiting a tendency for strong reflection or specular scattering. Based on this, a polarization sensitivity image is constructed for each region, marking high-sensitivity regions.

[0025] After completing the polarization sensitivity estimation, viewpoint consistency constraint detection is performed on the same region to further eliminate interference from the real heat source region. Using the registered two-view images, the grayscale value sequence of the center point and its five directions (up, down, left, and right) totaling nine pixels is extracted in each potential pseudo-peak region. The image under the second view is mapped to the projection coordinate system under the first view using back projection to maintain pixel position consistency. The grayscale values ​​of the same pixel positions in the two sets of images are compared, and the grayscale difference between each pixel pair is calculated. If more than 50% of the pixel pairs in this region have a grayscale difference exceeding 10% of the overall image dynamic range, and their local gradient direction has a reverse deflection within 180 degrees compared to the first view image, then the region is considered to have significant viewpoint inconsistency. To avoid occasional deviations affecting the overall judgment, the viewpoint consistency ratio, i.e., the ratio of the number of consistent pixels to the total number of pixels, is calculated in this region. When this ratio is less than 0.6, the region is defined as a viewpoint inconsistency region. For each viewpoint inconsistency region, a viewpoint consistency mask image is generated to characterize its consistency weakening distribution.

[0026] The obtained polarization sensitivity image and viewpoint consistency mask image are fused to generate a physical confidence spectrum, and the locations of high-risk pseudo-peak kernels are extracted accordingly. First, the polarization sensitivity image and viewpoint consistency image are normalized to ensure their grayscale values ​​are uniformly between 0 and 1. For each pixel, the polarization sensitivity value is multiplied by 0.4, and the viewpoint consistency value is multiplied by 0.6. These two values ​​are then added together to obtain the final physical confidence value for that pixel, generating a confidence image. To remove noise, the confidence image is Gaussian smoothed with a kernel size of 5×5 and a standard deviation of 1.2. Subsequently, the mean and standard deviation of the confidence image are calculated, and pixel locations with values ​​less than the mean minus one standard deviation are defined as physically unconfidential points. Eight-neighborhood connectivity analysis is used to cluster all physically unconfidential points. For regions with more than 25 pixels in a cluster, a minimum value localization operation is performed, finding the pixel with the lowest confidence value in each connected component, extracting its spatial coordinates, and marking it as a high-risk pseudo-peak kernel. The location of the pseudo-peak kernel will be stored for use in the subsequent gradient orthogonal registration step, to guide the elimination of abnormal gradient directions and the construction of stable edges.

[0027] Based on the location of high-risk pseudo-peak kernels, a gradient orthogonal registration chain is established. The horizontal and vertical gradients of the surface-fitted image are phase-conjugate registered to output a stable edge evidence map. After identifying and locating the spatial position of the reflection pseudo-peak kernels, to avoid their interference with the gradient direction of the infrared image, a gradient direction registration method guided by high-risk pseudo-peak kernels is proposed. Phase conjugation processing corrects directional anomaly regions in the image, thereby generating a stable and continuous edge evidence map, laying the foundation for subsequent texture feature extraction. The specific implementation steps are as follows: Using the identified high-risk pseudo-peak locations as references, initial gradient direction extraction and response suppression processing are performed on the surface-fitted infrared image. After smoothing the infrared image using a pre-sequence polynomial surface fitting, an image with suppressed background noise and a clear heat source structure is obtained. First-order difference calculations are performed on this image along the horizontal and vertical directions to obtain two gradient maps, representing the horizontal and vertical gray-level change rate images, respectively. To prevent isolated noise from misleading gradient calculations, a 3x3 median filter is performed on the image beforehand to smooth local outliers. After extracting the gradient maps, a circular neighborhood with a diameter of 11 pixels is established in the gradient map, centered on the coordinates of each high-risk pseudo-peak. Within this neighborhood, the gradient magnitudes in the horizontal and vertical directions are statistically analyzed, and their orientation angles are calculated as the dominant direction within the entire neighborhood. Subsequently, the gradient orientation angle of each pixel within this neighborhood is corrected so that its direction deviates from the dominant direction by no more than 15 degrees, while its gradient magnitude is reduced to 85% of the neighborhood mean. This step does not erase the original image structure, but only performs centralized correction and amplitude convergence on the gradient direction in the pseudo-peak perturbation region, thereby effectively weakening the local sudden gradient abruptness caused by reflective interference.

[0028] After completing the regional orientation-focused processing, an orthogonal registration chain is constructed based on the full-image gradient map, and pixel-level phase conjugate registration is performed to improve the continuity and physical consistency of edge orientations. Each pixel in the image is used as a processing unit, and its gradient magnitude and gradient orientation angle at corresponding positions in the horizontal and vertical gradient maps are extracted simultaneously. Using this pixel as the center, a 7x7 pixel local neighborhood is constructed, and the gradient orientation distribution of all pixels within this neighborhood is calculated. Orientation projection is performed on the gradient orientation values ​​of all neighboring pixels, transforming the horizontal and vertical gradient orientations into a unified polar coordinate representation. If, within this neighborhood, multiple pixels have orientation angles deviating from the principal orientation angle by more than 30 degrees, or if the orientation angle distribution exhibits a bimodal characteristic (i.e., the orientation distribution is concentrated in two opposing directions), then this region is determined to be an orientation-inconsistent region. For each region with inconsistent orientation, a phase conjugate registration operation is performed: First, the principal orientation angle of the neighborhood's orientation distribution is calculated. Then, the cosine of the angle between the orientation angle of each pixel and the principal orientation angle is compared. If the angle is less than 30 degrees, the original orientation is maintained; if the angle is greater than 30 degrees, the pixel orientation is adjusted so that its orientation vector is conjugate with the principal orientation vector, with the orientation angle difference controlled within 15 degrees. Simultaneously, the original amplitude is preserved, and only the orientation is adjusted. In this way, without sacrificing image gradient energy, the gradient orientation of local regions is unified, achieving directional continuity of the edge structure and ensuring that the contour of the same heat source does not exhibit edge breakage or orientation distortion due to false peak interference.

[0029] After completing the phase conjugate registration of the entire image, an edge magnitude map is constructed by integrating the gradient information in the horizontal and vertical directions, and a stable edge evidence map is further generated. Specifically, for each pixel in the image, the square root of the sum of the squares of its horizontal and vertical gradients is taken to obtain the gradient magnitude of that pixel. The gradient magnitudes of all pixels form an edge magnitude map. To extract the linear structure of the edges, a non-maximum suppression operation is performed on the edge magnitude map. This operation compares the magnitude of the current pixel with the magnitude of its two adjacent pixels in the principal gradient direction of each pixel. If the magnitude of the current pixel is greater than that of the two adjacent pixels, its value is retained; otherwise, it is set to zero, thereby highlighting local maxima to form edge lines. To remove low-intensity false edges in the background, a double-threshold filtering process is applied to the edge map after non-maximum suppression. The high threshold is set to the mean of the edge magnitudes of the entire image plus one standard deviation, and the low threshold is set to the mean minus one standard deviation. Only pixels with a threshold higher than the low threshold are retained, and weak edge pixels connected to pixels with a high threshold are tracked to form a continuous edge chain. Finally, linear normalization is performed on the edge chain strength, and the output edge response map is used as a stable edge evidence map. After being guided by high-risk pseudo-peak kernels, directional registration, and amplitude enhancement, this edge evidence map has the characteristics of clear structure, unified direction, and strong ability to suppress false responses, and can be directly used as the input basis for subsequent texture feature extraction and target localization algorithms.

[0030] An enhanced gray-level co-occurrence matrix feature structure is constructed based on a stable edge evidence map. The local contrast response is improved by a joint mechanism of extreme value triplet ranking statistics and dual neighborhood differential entropy, which suppresses brightness distortion caused by high-risk pseudo-peak kernels and generates a confidence contrast spectrum. After constructing the stable edge evidence map, to further highlight the texture features of weak target regions in infrared images against a complex background, an enhanced gray-level co-occurrence matrix feature construction method is proposed. This method combines extreme value triplet ranking statistics and a dual neighborhood difference entropy joint mechanism to generate a contrast spectrum with confidence indication. The specific steps are as follows: Based on salient edge regions in the stable edge evidence map, a direction-aware gray-level co-occurrence matrix statistical structure is established in the original grayscale image. Specifically, each pixel in the edge evidence map whose edge intensity is greater than the mean plus standard deviation of the image edge intensity is traversed and used as the center pixel. A square analysis window of size 21 pixels by 21 pixels is constructed in the original grayscale image. This window expands the gray-level co-occurrence statistics in four fixed directions, using the center pixel as the base point: horizontal (0 degrees), upper left to lower right (45 degrees), vertical (90 degrees), and upper right to lower left (135 degrees). In each direction, the gray-level combination frequency of adjacent pixel pairs is counted, and the gray-level is divided into 16 levels, obtaining a 16×16 gray-level co-occurrence matrix for each direction. Next, the gray-level co-occurrence matrix for each direction is normalized so that each matrix element represents the probability of occurrence of that gray-level pair. Furthermore, directional weights are assigned based on the gradient directional intensity of the current center pixel in the stable edge evidence map. The weights are calculated as follows: the proportion of the local gradient magnitude in each of the four directions to the total gradient energy at that point is used as the weighting factor of the gray-level co-occurrence matrix in that direction, thus enhancing the gray-level statistical contribution of the main direction. In the co-occurrence matrix after directional weighting, the probability value of each gray-level pair reflects the local texture generation characteristics in its corresponding direction. Through the above operations, a gray-level co-occurrence matrix feature representation structure that simultaneously possesses directional responsiveness and edge structure sensitivity is constructed, laying the foundation for recognizing the texture stability of real thermal targets.

[0031] Based on the constructed directional weighted gray-level co-occurrence matrix, local extreme gray-level pairs are extracted from each matrix, and an extreme triplet ranking statistical operation is performed. The specific execution process is as follows: In each co-occurrence matrix, the top 3 gray-level pairs by frequency value are selected as the extreme triplets in that direction. The gray-level combinations and probability values ​​of these three gray-level pairs are recorded respectively, constructing a triplet sequence. In each of the four directions, these three gray-level pairs are sorted from highest to lowest probability of occurrence. Next, the consistency of the sorting structure in the four directions is analyzed. If the gray-level combinations and their sorting positions of the triples are consistent in all four directions, it is considered a consistent sort; if the sorting order is different or the gray-level pair combinations change in at least two directions, it is considered a non-consistent sort. Based on this, a triplet structure consistency score is assigned to the current pixel: 1 for consistent sorting and 0 for non-consistent sorting. Then, the sum of the consistency scores is calculated within a 5x5 neighborhood centered on the current pixel and normalized to a structure consistency index between 0 and 1. A higher index indicates a stable grayscale texture dominance in the region across multiple directions; a lower index indicates disordered grayscale transitions, potentially dominated by spurious peaks or background noise. This triplet structure consistency index serves as a crucial basis for the reliability of the texture structure and is used as a reference for weight allocation in subsequent entropy analysis.

[0032] After obtaining the triplet structure consistency index, to further enhance the discrimination ability of low-contrast regions, a joint calculation method of dual neighborhood differential entropy is proposed, and a confidence contrast spectrum is constructed based on this method. The specific implementation is as follows: Based on each center pixel, two rectangular regions are established around it: one is a horizontal neighborhood with a size of 15 pixels wide and 5 pixels high; the other is a vertical neighborhood with a size of 5 pixels wide and 15 pixels high. Gray-level distributions are extracted in both neighborhoods, and gray-level histograms are constructed with 16 gray-level levels. Gray-level differential entropy is calculated in each direction, and the entropy value reflects the complexity of the local gray-level distribution in that direction. If the differential entropy in one direction is much higher than that in another direction, for example, a difference exceeding 0.3 (the unit interval of normalized entropy values), it indicates that the gray-level change in that region has strong directionality; such regions usually correspond to target edges or texture abrupt change points. If the entropy values ​​in both directions are low, it indicates that the region has uniform gray-level and weak texture structure, possibly corresponding to background regions or pseudo-peak cores. To integrate the directional judgment result with the aforementioned triplet structure consistency index, the difference between the entropy values ​​of the two directions for each pixel is used as a directional sensitivity score, which is then multiplied by the structure consistency index to obtain the confidence contrast score for that pixel. After calculating these scores across the entire image, a confidence contrast spectrum is obtained. To enhance visual discrimination and structural continuity, this spectrum is smoothed using a Gaussian filter with a 5x5 kernel size and a standard deviation of 1.0. Finally, the spectrum is linearly normalized to limit its values ​​to between 0 and 1, serving as input for subsequent thresholding steps.

[0033] A cross-frame adaptive threshold engine is built based on confidence contrast spectrum. A residual sensitivity feedback loop is constructed using a time-robust threshold mechanism based on mean and standard deviation to generate a dynamic equilibrium threshold and suppress detection ghosting. In the previous step, a confidence contrast spectrum was obtained to reflect the reliability of each pixel region in the infrared image in terms of texture structure and brightness features. To stably identify real thermal anomaly regions in the image sequence and effectively suppress detection tailing caused by background residue, response hysteresis, or high-brightness reflection, this step proposes a dynamic thresholding mechanism combining temporal statistical analysis and residual adjustment to construct a temporally stable thresholding engine. The specific steps are as follows: A cross-frame feature statistical structure is constructed based on the confidence contrast ratio spectrum, and the temporally stable feature quantity of each pixel is extracted. Specifically, ten consecutive frames of images from the current infrared image sequence are selected, ensuring consistent acquisition intervals, constant exposure parameters, and the same image resolution. Image preprocessing, stable edge extraction, co-occurrence matrix construction, and contrast ratio spectrum generation are performed on each of these ten frames to ensure that each frame has a confidence contrast ratio spectrum with the same dimension and semantics. Using the geometric center of the image as a reference, the ten frames are spatially aligned using a sub-pixel precision feature point registration algorithm to eliminate positional inconsistencies caused by slight viewpoint shifts. After alignment, the confidence contrast ratio value of each pixel in the ten frames is extracted, forming a time series of length ten. Statistical processing is performed on this time series to calculate the mean and standard deviation of the pixel in the time dimension, forming a time-series mean image and a time-series standard deviation image, respectively. The time-series mean image is used to characterize the long-term response trend of the pixel, while the time-series standard deviation image reflects the degree of fluctuation of the pixel in the time dimension. This method compares the confidence contrast value of each pixel in the current frame with its corresponding historical mean and standard deviation. When the value exceeds the mean plus twice the standard deviation, the pixel is considered to have undergone a significant contrast change in the current frame and possesses the potential for thermal target response. Unlike traditional global histogram distributed thresholding methods, this method customizes the judgment threshold for the temporal behavior pattern of each pixel, exhibiting stronger local adaptability and dynamic response capabilities.

[0034] After establishing the basic threshold judgment mechanism, a sensitivity feedback structure associated with temporal residual fluctuations is constructed to adaptively fine-tune the threshold to cope with potential thermal structural abrupt changes, rapid background drift, or drastic changes in the appearance and disappearance of targets in the image. Specifically, in the current frame image, all candidate target pixels that meet the basic threshold conditions are extracted. For these pixels, the difference between their confidence contrast values ​​in the current frame and the previous frame is calculated and recorded as cross-frame residual values. All residual values ​​are statistically analyzed, and the mean and range of residuals for the entire image are calculated, defined as the residual fluctuation index. This index reflects whether the potential target region in the current frame has undergone significant intensity fluctuations. When the residual fluctuation index exceeds 20% of the index of the previous frame, it indicates a strong grayscale structure change in the image, and the system may mistakenly identify noise as a target. In this case, the threshold benchmark of the current frame is appropriately increased, raising the detection threshold by 10% to suppress false target triggering. Conversely, when the residual fluctuation index is less than 20% of the previous frame's index, it indicates that the image response is stable and the target boundary is stable. In this case, the threshold benchmark of the current frame is reduced by 10% to enhance the sensitivity to targets with edge decay. This residual feedback mechanism automatically adjusts the threshold according to the actual response behavior of the image. In various dynamic processes such as the appearance, change, stabilization, or disappearance of the target, it can maintain the convergence of detection conditions and real-time response capability, effectively improving the detection accuracy and environmental adaptability.

[0035] Combining cross-frame statistical results and residual feedback output, a dynamic equilibrium threshold map for the current frame is generated, and ghosting suppression and target mask output are completed. Based on the temporal mean image, twice the standard deviation of each pixel is added, and then the dynamic adjustment value output by the residual feedback mechanism is superimposed to form the dynamic detection threshold image of that pixel. Subsequently, the confidence contrast spectrum of the current frame is compared pixel by pixel with the dynamic threshold image. If the confidence contrast value of a pixel is greater than its corresponding detection threshold value, the pixel is marked as a candidate target pixel. The initially obtained candidate target images may have false flashes caused by short-term noise in terms of temporal consistency. To solve this problem, this implementation introduces a temporal consistency verification mechanism. Specifically, in the candidate target image, for each pixel marked as a target, it is reviewed whether it also meets the target condition in the previous two frames. If the pixel is a target in at least two of the three frames, the pixel is retained as the final target point; otherwise, it is regarded as an isolated response without temporal stability and is discarded. This method ensures that the target response is continuous over time, effectively eliminating trailing or sporadic jumps. After completing the temporal consistency check, eight-neighbor connectivity analysis is performed on the retained pixels to filter out scattered regions with an area smaller than 30 pixels, retaining only target response regions with complete boundaries. The final generated target mask image possesses high-confidence, structurally intact, and clearly defined target contour information, which can be used for subsequent contour extraction, localization, and discrimination tasks.

[0036] Based on the dynamic equilibrium threshold, multi-field coupling dynamic control is performed. Through the same frequency driving method of optical phase injection and thermal radiation perturbation, the surface fitting weight and threshold gain are synchronously adjusted to complete the extinguishing of false bright spots and the closed-loop self-healing control of the detection algorithm. After generating the dynamic equilibrium threshold map, it is necessary to further adjust the surface fitting weights and detection threshold gain in the image synchronously to achieve closed-loop control of the suppression and detection process for false bright spots in complex photothermal environments. This includes the following steps: Based on the dynamic balance threshold map and confidence contrast spectrum of the current frame, all pixel regions marked as bright responses in the target detection mask but lacking temporal continuity in the previous two frames are identified, and a candidate map of false responses is constructed. On this basis, neighborhood structure analysis is performed on each candidate bright response. A circular region with a radius of 5 pixels centered at that point is selected, and its grayscale value, edge gradient direction, and thermal texture distribution features are extracted. The differences between this region and its corresponding data at the same location in the previous and next frames are calculated. If the grayscale value of this region is greater than twice the overall average of the current frame image, and the edge structure is discontinuous and broken, while lacking stable response support in the temporal dimension, then this region is identified as a high-risk false bright spot and requires subsequent dynamic adjustment operations.

[0037] After identifying the false bright areas, the polarization angle response of these areas in the physical illumination path is calculated by combining the current optical incident angle of the imaging device, the background illuminance level, and the receiving surface direction of the thermal imaging equipment. An optical phase field compensation map based on the true incident path is then constructed. Eight equidistant sampling rings are constructed around the false response areas. Each ring collects edge phase change values ​​in eight directions, and Fourier analysis is performed on the phase shifts in each direction to determine the main perturbation directions. When the phase shift amplitude in one direction exceeds twice that in other directions, and the phase change exhibits a fast peak response characteristic, it indicates that the area is affected by a non-uniform transmitted light field, requiring the introduction of a phase modulation factor into the fitting process. This modulation factor is injected as a position-specific weight into the fitting algorithm, reducing the response curvature in the fitting process of this area by 20% to 30%, thereby suppressing the generation of abnormal fitting surfaces and preventing fitting peaks caused by reflection from interfering with subsequent detection and judgment.

[0038] In addition to optical feature manipulation, thermal radiation perturbation injection is also required for the aforementioned region to construct a radiation fluctuation correction map consistent with the actual heat source response. Within the same region, the thermal intensity change trajectory of this region is extracted from five frames of images, and the degree of deviation between this trajectory and the average thermal drift curve of the entire frame image is compared. If the thermal intensity of this region increases or decreases drastically in a short period of time without showing a stable trend, it is considered a local thermal perturbation rather than a real heat source. During the surface fitting process, a high-frequency, low-amplitude perturbation term is injected into this region, making it identified as a high-variability, low-confidence region in the fitting algorithm. Correspondingly, its curvature response weight is reduced to 60% of the default value of the fitting model to ensure that this region is not overestimated as the target region in the overall background modeling.

[0039] To synchronize the surface fitting process with the detection threshold judgment, a three-variable response coupling matrix is ​​constructed, with thermal response residual, optical phase perturbation amplitude, and threshold gain amplitude as inputs. By statistically analyzing the coupling strength of each false highlight region in the current frame image, a fitting-detection dual-path adjustment spectrum is formed. When a detection region exhibits strong fluctuations in both the optical perturbation and thermal residual channels, and its dynamic threshold gain reaches or exceeds 1.2 times the default setting, the system simultaneously tightens the curvature control of the fitted surface and increases the threshold for that region to raise the recognition threshold and avoid false detections. When a region shows only slight anomalies in both channels and has a low threshold gain, the fitting accuracy control is relaxed to make it closer to the real texture and enhance the detectability of low-contrast targets. This joint control strategy fuses and models information from various signal sources, avoiding misjudgments, response delays, or false target proliferation caused by the separation of processing stages in traditional methods.

[0040] After completing the above adjustment operations, a detection closed-loop self-healing process is executed. This involves simultaneously updating the surface fitting adjustment map and the dynamic threshold adjustment map to the input of the current detection process, and re-executing the gradient generation, co-occurrence matrix feature extraction, and candidate target discrimination processes for the entire image. Consistency analysis is then performed again on the identified target regions, removing pixels identified as perturbation responses and updating the target mask map. Simultaneously, the adjustment parameters are stored in the current frame's historical adjustment parameter set, serving as the basis for the next frame's photothermal interference prediction and response adjustment. Through this continuous update, information reinjection, and structure re-identification mechanism, a complete detection self-healing closed loop is constructed, ensuring stable detection performance even in scenarios with multi-source perturbations and multi-target interference. This suppresses false detections and enhances the detection capability of weak targets, forming a highly robust and adaptable intelligent detection process.

[0041] This invention achieves accurate identification and structural mapping of fitting anomaly regions by introducing a traction residual distribution map; it establishes a false peak identification mechanism by combining polarization sensitivity and viewpoint consistency constraints, improving the resolution of false bright spots; it ensures the consistency of edge features and the robustness of texture expression through gradient orthogonal registration and differential entropy enhancement strategies; it introduces a cross-frame temporal threshold mechanism and a residual sensitive feedback loop to achieve temporal dynamic stability of detection conditions; and it constructs a dynamic coupling control strategy through optical phase injection and thermal radiation perturbation driving, completing the synchronous adaptive adjustment of surface fitting weights and detection threshold gain, enabling the detection process to have automatic identification, dynamic response, self-repair, and closed-loop control capabilities, effectively suppressing false targets and improving detection accuracy and robustness, especially suitable for infrared image environments with significant photothermal interference, complex structures, and weak targets.

[0042] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A human infrared image small target detection method based on improved FGLCM features, characterized in that, The method comprises the following steps: A boundary singularity audit baseline is established under a unified time baseline, multi-scale energy mapping is performed on the residual error of surface fitting of the human infrared image, and a residual error distribution map is generated; A reflection false peak discriminator is constructed based on the residual error distribution map, combined with polarization sensitivity estimation and view angle transformation consistency constraint, and a high-risk false peak core position is extracted; A gradient orthogonal registration chain is established based on the false peak core position, the horizontal direction gradient and the vertical direction gradient of the image are phase conjugate registered, and an edge evidence map is generated; A gray level co-occurrence matrix feature structure is constructed based on the edge evidence map, an extreme three-tuple ranking statistics and a dual neighborhood difference entropy joint mechanism are used to generate a confidence contrast spectrum; A cross-frame adaptive threshold engine is constructed based on the confidence contrast spectrum, and a dynamic balance threshold is generated by using the time sequence threshold mechanism of mean value and standard deviation; Based on the dynamic balance threshold, multi-field coupling dynamic regulation is performed, the surface fitting weight and threshold gain are adjusted by the same frequency driving mode of optical phase injection and thermal radiation perturbation, and the detection process is controlled.

2. The human infrared image small target detection method based on improved FGLCM features according to claim 1, characterized in that, The residual error distribution map generation step is as follows: Perform gray consistency calibration and background drift correction, normalize the infrared image gray scale, and take the average gray scale of the image edge area as the background temperature reference to complete the whole image gray difference compensation; Perform local surface fitting and extract fitting residual error, use a fixed size sliding window to perform polynomial fitting, calculate the gray difference between each pixel and the fitting surface, and generate a fitting residual error image; Based on the fitting residual error image, a multi-scale energy mapping is constructed, and fine-scale, medium-scale and wide-scale residual error energy maps are extracted and weighted fused to identify energy abnormal areas to form a binary mask; Perform counterfactual playback operation on the energy abnormal area to generate two groups of perturbation images, extract stable residual error pixels and complete clustering, and construct a residual error distribution map.

3. The human infrared image small target detection method based on improved FGLCM features according to claim 2, characterized in that, The high-risk false peak core position extraction process is as follows: According to the residual error distribution map, high residual error blocks are extracted and regions with local brightness abnormalities and multi-directional gradient mutations are identified; Perform polarization sensitivity estimation on the above-mentioned regions using dual-view images, calculate the gray difference between the two views, and extract the polarization strong response region; Perform view consistency constraint detection on the same region to identify regions with inconsistent gray difference and gradient direction; Weighted fusion of polarization sensitivity image and view consistency mask image generates a physical credibility spectrum which is then Gaussian smoothed and connected domain clustered, and the lowest credibility pixel is extracted as the high-risk false peak core position.

4. The human infrared image small target detection method based on improved FGLCM features according to claim 3, characterized in that, The edge evidence map generation step is as follows: Based on the high-risk false peak core position, the horizontal direction gradient and the vertical direction gradient of the infrared image after surface fitting are extracted and directionally concentrated, and the direction angle and amplitude in the neighborhood are corrected; Based on the image global, an orthogonal registration chain is constructed, and the phase conjugate registration operation is performed on the regions with inconsistent directions to unify the local gradient direction distribution; Fuse the corrected horizontal direction and vertical direction gradient information, perform non-maximum suppression and double-threshold edge extraction, and generate a stable edge evidence map for subsequent texture feature extraction and target positioning.

5. The human infrared image small target detection method based on improved FGLCM features according to claim 4, characterized in that, When performing the phase conjugate registration operation, the gradient direction adjustment is based on the cosine value of the included angle between the main direction angle and the original direction angle, and when the included angle exceeds a set threshold, the original direction is adjusted to be in a conjugate state with the main direction, and the difference between the direction angles does not exceed a preset range, so as to ensure the continuity and consistency of the local gradient direction.

6. The human infrared image small target detection method based on improved FGLCM features according to claim 4, characterized in that, The confidence contrast spectrum generation step is as follows: Based on the high edge intensity pixel points in the stable edge evidence map, a direction perception gray level co-occurrence matrix structure is established in the original gray level image, the gray level co-occurrence relationship is counted in multiple fixed directions, and the direction weight is allocated according to the gradient intensity; Local extreme gray pairs are extracted from the direction weighted gray level co-occurrence matrix, the extreme value triplets are ranked and counted, the direction sorting consistency is analyzed, and the structure consistency index is calculated, which is used to represent the region texture stability; Based on the structure consistency index, a dual neighborhood difference entropy joint mechanism is established, the horizontal and vertical direction gray difference entropies are calculated respectively, and the direction sensitivity and consistency index are fused to generate the confidence contrast spectrum, which is output after smoothing and normalization as the input basis for subsequent threshold discrimination.

7. The human infrared image small target detection method based on improved FGLCM features according to claim 6, characterized in that, The dynamic balance threshold generation step is as follows: Based on the confidence contrast spectrum, the confidence contrast values in the current frame and multiple historical images are extracted, the time sequence mean and standard deviation of each pixel are calculated, and the time sequence mean image and standard deviation image are formed; A sensitivity feedback structure related to the cross-frame residual fluctuation is constructed, the residual fluctuation index is calculated based on the confidence contrast difference between the current frame and the previous frame of the candidate target region, and the detection threshold of the current frame is adjusted according to the fluctuation change; The time sequence statistical result and the residual feedback result are fused to generate the dynamic balance threshold image of the current frame, the confidence contrast spectrum image and the dynamic threshold image are compared pixel by pixel, the time sequence consistency verification and connected region analysis are combined, and the high confidence target mask image is output.

8. The human infrared image small target detection method based on improved FGLCM features according to claim 7, characterized in that, Based on the dynamic balance threshold, multi-field coupling dynamic regulation is performed, the fitting weight and threshold gain are adjusted through the same frequency driving mode of optical phase injection and thermal radiation perturbation, and the control detection process is as follows: Based on the dynamic balance threshold image and the confidence contrast spectrum image, false highlight regions without time sequence continuity in the current frame are identified, and high-risk false highlight spots are determined through neighborhood structure analysis; Based on the optical incidence angle and phase disturbance, a phase compensation map is established, position-specific weights are injected into the fitting surface to suppress abnormal curvature response; The thermal intensity trajectory is extracted and the thermal disturbance residual is calculated, and disturbance items are injected into abnormal regions to reduce their fitting response reliability; A three-variable coupling matrix composed of optical phase disturbance amplitude, thermal response residual and threshold gain is constructed, and the fitting curvature and threshold gain strategies are adjusted synchronously; The fitting adjustment map and the threshold adjustment map are updated together to input the detection process, the detection result is updated, the adjustment parameters are stored, and a closed-loop self-healing control structure is constructed.

Images