Night vision full-color imaging control method

Abstract

The present application relates to a kind of night vision full-color imaging control method, adopts polarization light path hardware synchronous acquisition linear polarization modulation micro-light image and wide-spectrum infrared radiation image, and the time series data of scene polarization, atmospheric scattering, sensor quantum noise are synchronously acquired;By constructing non-rigid mutual information topological grid, double-image pixel-level accurate registration is realized;Build space-time variation non-uniform entropy field noise reduction model, rely on pixel dynamic entropy threshold formula to realize pixel-level adaptive differentiation noise reduction;Based on multi-material spectral fingerprint knowledge base combined with physical constraint, pixel-level spectral inversion sample-free real color reconstruction is completed;Finally, construct three-dimensional residual error vector field of spectrum, polarization, radiation, accurately identify distortion type and iteratively correct parameters in reverse, form full-link closed loop optimization.The present application completely gets rid of traditional fusion coloring technology route, imaging detail fidelity is high, color restoration is real, scene adaptability is strong, can effectively improve the full-color imaging quality and stability of complex night vision scene.

Description

Technical Field

[0001] This invention belongs to the technical field of security closed-circuit television imaging systems, and particularly relates to a night vision full-color imaging control method. Background Technology

[0002] Current full-color night vision imaging technologies generally employ the conventional technique of "grayscale fusion of low-light images and infrared images + deep learning color mapping." Relying on traditional feature point registration, globally unified filtering and noise reduction, and color fitting from a sample library to complete night vision imaging, this approach has inherent technical limitations in low-light conditions, stray light interference, and complex nighttime scenes. Specifically, it suffers from two core problems: First: Traditional dual-light image registration has low accuracy and poor denoising mechanism versatility, resulting in loss of detail and severe noise residue in night vision images.

[0003] Existing technologies mostly employ fixed feature point matching methods such as SIFT and ORB to register low-light and infrared images, achieving only coarse-grained pixel alignment. This is unsuitable for nighttime scenes with weak textures and low signal-to-noise ratios, easily leading to problems such as field-of-view deviation, pixel misalignment, and phase shift. Furthermore, traditional noise reduction methods often use uniform noise reduction approaches such as global fixed threshold filtering, Gaussian filtering, and guided filtering, failing to differentiate between structural, textured, and flat regions of the image, nor considering the temporal fluctuations of inherent quantum noise in sensors. During noise reduction, either excessive noise reduction results in the loss of crucial details such as edges and textures, or insufficient noise reduction fails to effectively remove nighttime stray atmospheric light noise, sensor quantum noise, and fixed dark noise points, ultimately causing blurry night vision images, low signal-to-noise ratios, and poor image fidelity.

[0004] Second: Traditional sample mapping AI coloring mechanisms deviate from the laws of physical imaging, resulting in distorted color reproduction at night, poor scene adaptability, and no closed-loop correction capability.

[0005] Current color reconstruction methods for full-color night vision imaging are all based on pixel fitting mapping using deep learning sample libraries. They rely on massive training samples for color transfer, failing to consider physical imaging factors such as nighttime atmospheric scattering, scene polarization characteristics, and the spectral properties of object materials. This type of method is essentially "image-level pseudo-color filling," not true color restoration based on physical spectral laws. It is highly prone to problems such as color homogenization, color shift, and inconsistencies between real and virtual colors for objects of different materials, including vegetation, roads, buildings, and water. Furthermore, existing technologies lack imaging distortion detection and parameter iteration closed-loop mechanisms, making it impossible to reverse-correct imaging defects such as registration deviations, noise reduction mismatches, and spectral distortions. This results in poor imaging stability and low color consistency in complex night vision scenes, making it difficult to achieve high-precision, high-fidelity full-color night vision imaging. Summary of the Invention

[0006] The purpose of this invention is to provide a night vision full-color imaging control method to solve the problems mentioned in the background art.

[0007] In view of this, the present invention provides a night vision full-color imaging control method, the method comprising: based on a polarization-splitting synchronous imaging optical path, simultaneously acquiring linearly polarized modulated low-light images and broadband infrared radiation images of a night vision target scene at the same imaging time and within the same field of view, and simultaneously acquiring scene polarization degree spatial distribution parameters, atmospheric scattered light intensity distribution parameters, and quantum noise time sequence data of the image sensor in real time; Cross-domain feature decomposition is performed on the linearly polarized modulated low-light image and the broadband infrared radiation image. Dual-domain core features are extracted and a non-rigid mutual information topological mesh is constructed to complete the adaptive topological registration of the two images and generate a pixel-level one-to-one corresponding registration coupling image. A spatiotemporal variational non-uniform entropy field denoising model is constructed. The time-series data of quantum noise from the sensor is fused with the spatial entropy field of the registered and coupled image to generate a pixel-level dynamic adaptive entropy threshold. The registered and coupled image is subjected to regional differential denoising suppression to remove quantum noise, stray light noise and fixed dark noise points, resulting in a high-fidelity low-noise night vision base image. A multi-material spectral fingerprint knowledge base adapted to night vision scenarios is pre-built. The spatial distribution parameters of scene polarization degree and atmospheric scattered light intensity distribution parameters are combined as physical constraints. Pixel-level material spectral attribution and physical spectral radiometric inversion are performed on high-fidelity, low-noise night vision base images to complete the real color reconstruction without sample mapping and generate an initial physical full-color image. A three-dimensional residual vector field of spectrum, polarization, and radiation is constructed for the initial physical full-color image. The image distortion type is determined by the residual distribution characteristics. Based on the distortion type, the topological registration parameters, entropy field noise reduction threshold, and spectral color reconstruction coefficient are iteratively corrected in reverse. After closed-loop optimization, the final night vision full-color imaging result is output.

[0008] In a further embodiment of the present invention, the step of performing cross-domain feature decomposition on the linearly polarized modulated low-light image and the broadband infrared radiation image, extracting dual-domain core features and constructing a non-rigid mutual information topological mesh to complete the adaptive topological registration of the two images and generate a pixel-level one-to-one corresponding registration coupled image includes: performing global gradient calculation on the linearly polarized modulated low-light image, extracting the polarization angle parameters at each pixel position of the image, and generating a continuous and stable polarization angle gradient field feature; Radiation extrema detection is performed on the broadband infrared radiation image to extract the radiation intensity topological ridge features at the boundary between light and dark areas of the image and the contour of the object. The two images to be registered are divided into mutual information unit blocks with adaptive size adjustment. The size of the unit block is dynamically scaled according to the local image gradient entropy value. The unit block size is reduced in areas with complex textures and enlarged in areas with flat textures. Calculate the joint mutual information matrix of polarization features and infrared radiation features within each mutual information unit block, and select the pixel points corresponding to the maximum mutual information values ​​in the matrix as cross-domain topological anchor points. With each topological anchor point as the core, radiating connections are made to neighboring related pixels to construct a non-rigid mutual information topological mesh with elastic deformation characteristics, generating many-to-one related topological edges between pixels. The topological mesh is subjected to multiple rounds of elastic deformation iterative optimization to successively eliminate field deviation, pixel distortion and phase shift between the two images, and finally achieve pixel alignment of the two images to generate a registration coupled image.

[0009] In a further embodiment of the present invention, the step of calculating the joint mutual information matrix of polarization features and infrared radiation features within each mutual information unit block, and selecting the pixel points corresponding to the maximum mutual information values ​​in the matrix as cross-domain topological anchor points, includes: extracting the polarization feature dataset of all low-light polarized pixels and the radiation feature dataset of all infrared pixels within a single mutual information unit block. The probability distribution density of the two datasets is statistically analyzed, and the joint entropy and edge entropy between the two types of features are calculated to obtain the mutual information value of pixels within the unit block. Traverse the mutual information values ​​of all pixel combinations within the unit block to construct a two-dimensional joint mutual information matrix; Set a peak threshold for mutual information between unit blocks, and select pixel coordinates in the matrix that exceed the peak threshold as initial anchor points; The initial anchor points are deduplicated and their neighborhoods are filtered, and the pixels with the most uniform distribution and the strongest feature significance are retained as the final cross-domain topology anchor points.

[0010] In a further embodiment of the present invention, the construction of a spatiotemporal variational non-uniform entropy field noise reduction model, which fuses the temporal sequence data of sensor quantum noise with the spatial entropy field of the registration coupled image to generate a pixel-level dynamic adaptive entropy threshold, performs regional differential noise reduction and suppression on the registration coupled image, removes quantum noise, stray light noise and fixed dark noise points, and obtains a high-fidelity low-noise night vision base image, includes: performing pixel-by-pixel spatial information entropy calculation on the registration coupled image to generate a global continuous spatial variational entropy field, and dividing the image into three types of regions according to the entropy value: high structure entropy region, texture detail entropy region, and low-noise flat entropy region; Extract the temporal sequence of quantum noise from the image sensor for a preset number of consecutive frames, statistically analyze the temporal grayscale fluctuation characteristics pixel by pixel, calculate the temporal quantum noise entropy corresponding to each pixel, and construct the temporal entropy distribution matrix. The spatial variational entropy field and the temporal entropy distribution matrix are fused at the pixel level to generate a global spatiotemporal joint entropy density map of the image. Based on the pixel dynamic entropy threshold calculation formula, an independent adaptive noise reduction threshold is generated for each pixel of the image. Nonlinear noise reduction and suppression operators are matched for high structure entropy region, texture detail entropy region and low noise flat entropy region respectively. Quantum random noise, atmospheric stray light-derived noise and fixed dark noise in the corresponding region are suppressed according to the independent threshold of each pixel. The image contour structure and micro texture details are preserved throughout the process to obtain a high-fidelity low noise night vision base image.

[0011] In a further embodiment of the present invention, the method for generating an independent adaptive noise reduction threshold for each pixel of the image based on the pixel dynamic entropy threshold calculation formula is as follows: Where T(x,y) is the dynamic adaptive noise reduction threshold corresponding to the pixel at coordinates (x,y), T0 is the system-preset baseline noise reduction threshold, and H s (x,y) represents the spatial variational entropy value corresponding to this pixel, H t (x,y) represents the temporal quantum noise entropy value corresponding to this pixel, which is a very small constant approaching 0, used to prevent division by zero errors in calculations. This represents the local variance of pixel grayscale within the neighborhood window of that pixel. The system's preset standard reference variance; Different pixels generate differentiated thresholds based on their own spatial texture features and temporal noise features, achieving adaptive noise reduction without fixed thresholds or globally uniform parameters across the entire domain.

[0012] In a further embodiment of the present invention, the pre-built multi-material spectral fingerprint knowledge base adapted to night vision scenarios, combined with the spatial distribution parameters of scene polarization degree and atmospheric scattered light intensity distribution parameters as physical constraints, performs pixel-level material spectral attribution discrimination and physical spectral radiometric inversion on high-fidelity low-noise night vision base images, completes real color reconstruction without sample mapping, and generates an initial physical full-color image, including: pre-collecting various materials such as vegetation, buildings, roads, sky, metal, and water bodies in night vision scenarios, obtaining the spectral response curves of various materials in the low-light polarization band, infrared radiation band, and visible light color band respectively, extracting the unique spectral feature parameters of each material, and constructing a multi-material spectral fingerprint knowledge base; The spatial distribution of scene polarization degree and atmospheric scattered light intensity are collected in real time as physical constraints to correct the attenuation coefficient and radiation compensation coefficient of spectral fingerprint. The polarization-infrared joint spectral feature vectors of each pixel in the high-fidelity, low-noise night vision substrate image are extracted, and the feature vectors are physically constrained to be matched with the standard fingerprint curves of various materials in the spectral fingerprint knowledge base. The material category corresponding to each pixel is determined based on the optimal matching result, and pixel-level material spectral classification is completed. Based on the atmospheric radiative transfer physical model, the standard spectral fingerprint parameters of the corresponding material are called to retrieve the true color radiance values ​​of the RGB three channels pixel by pixel, complete the physical reconstruction of the whole domain color, and generate the initial physical full-color image.

[0013] In a further embodiment of the present invention, the step of extracting the polarization-infrared joint spectral feature vector of each pixel in the high-fidelity low-noise night vision substrate image and performing physical constraint matching between the feature vector and the standard fingerprint curves of various materials in the spectral fingerprint knowledge base includes: collecting four-dimensional feature parameters of a single pixel, namely polarization response amplitude, polarization phase offset, infrared radiation intensity, and infrared band attenuation coefficient, and splicing them to form a pixel-specific joint spectral feature vector. By introducing polarization bias penalty terms, atmospheric scattering attenuation penalty terms, and infrared radiation bias penalty terms, a multi-dimensional physical matching distance formula is constructed. Calculate the physical matching distance between the current pixel feature vector and the spectral fingerprints of each material in the knowledge base. The smaller the matching distance, the higher the material matching degree. The material fingerprint corresponding to the minimum matching distance is selected as the spectral fingerprint of the current pixel to complete the material identification.

[0014] In a further embodiment of the present invention, the construction of a three-dimensional residual vector field of spectrum, polarization, and radiation for the initial physical full-color image, and the determination of the image distortion type by residual distribution characteristics, includes: performing pixel-by-pixel difference calculations on the pixel color spectrum parameters, polarization response parameters, and infrared radiation parameters of the initial physical full-color image and the original physical parameters of the high-fidelity low-noise night vision substrate image to obtain the global spectral residual matrix, polarization residual matrix, and radiation residual matrix, respectively. The three-dimensional residual matrices are spatially coupled to construct a global three-dimensional residual vector field; Spatial distribution skewness, pixel dispersion, and regional coherence of statistical residual vector fields; If the spectral residual amplitude accounts for the largest proportion and is discretely distributed, it is determined to be color spectral distortion; If the polarization residual amplitude accounts for the largest proportion and is locally clustered, it is determined to be image registration distortion; If the radiation residual amplitude accounts for the largest proportion and is uniformly offset across the entire domain, it is determined to be a distortion due to mismatch of noise reduction parameters.

[0015] In a further embodiment of the present invention, the step of correcting the topology registration parameters, entropy field noise reduction threshold and spectral color reconstruction coefficient based on the distortion type through reverse iteration includes: when it is determined that there is color spectral distortion, fine-tuning the atmospheric scattering attenuation compensation coefficient of the spectral fingerprint knowledge base, re-executing the pixel-level material spectral inversion and color reconstruction operation, and correcting the color shift problem. When image registration distortion is detected, the elastic stiffness coefficient of the non-rigid mutual information topology mesh and the anchor point matching threshold are adjusted, and the dual image topology registration is completed again to eliminate pixel misalignment distortion. When the noise reduction parameter mismatch distortion is identified, the baseline threshold and standard reference variance parameter in the dynamic entropy threshold calculation formula are corrected, and a pixel-level adaptive noise reduction threshold is regenerated to optimize the image noise reduction effect and preserve complete texture details.

[0016] In a further embodiment of the present invention, the method of simultaneously acquiring linearly polarized modulated low-light images and broadband infrared radiation images of a night vision target scene at the same imaging time and within the same field of view based on the polarization-splitting synchronous imaging optical path includes: using a coaxial polarization-splitting imaging structure to split the incident light rays of the scene through a polarization-splitting prism, with one light ray connected to a low-light imaging sensor to acquire a linearly polarized modulated low-light image, and the other light ray connected to an infrared imaging sensor to acquire a broadband infrared radiation image; Strictly calibrate the pixel resolution, imaging field of view, optical center coordinates and imaging frame rate of the two sensors to ensure that the spatial dimensions of the two images are completely aligned and there is no inter-frame time difference in the temporal dimension. The entire process involves no mechanical scanning, no post-processing pixel interpolation, and no inter-frame synthesis, achieving physical synchronous acquisition of dual-domain images.

[0017] The beneficial effects of this invention are: Firstly, it achieves pixel-level precise registration and differentiated adaptive noise reduction, completely eliminating imaging noise and preserving the microscopic textures and structural details of night vision scenes to the greatest extent.

[0018] Abandoning traditional fixed feature point registration methods, this approach constructs adaptively sized mutual information unit blocks and non-rigid mutual information topological meshes. It maximizes the mutual information of dual-domain features to filter topological anchor points and combines this with elastic deformation iterative optimization to achieve pixel-level precise alignment between two images, completely resolving issues of field-of-view deviation, pixel distortion, and phase shift in low-texture night vision scenes. Simultaneously, a spatiotemporal variational non-uniform entropy field denoising model and a pixel dynamic entropy threshold calculation formula are used, fusing spatial texture entropy and temporal quantum noise entropy to generate an independent adaptive denoising threshold for each pixel. Differentiated nonlinear denoising operators are matched for image structure regions, texture regions, and flat regions. This accurately distinguishes and eliminates three typical nighttime noise types: quantum random noise, atmospheric stray light noise, and fixed dark noise points. It fundamentally avoids the problems of detail smoothing or incomplete denoising caused by global uniform filtering, significantly improving the signal-to-noise ratio and image fidelity of night vision images.

[0019] Secondly, based on physical spectrum inversion, sample-free true color reconstruction is achieved, coupled with three-dimensional residual closed-loop correction, which greatly improves the realism and scene stability of night vision full-color imaging.

[0020] Abandoning the pseudo-coloring scheme of traditional deep learning sample mapping, this method builds a multi-material spectral fingerprint knowledge base. Using scene polarization and atmospheric scattering intensity as physical constraints, and combining an atmospheric radiative transfer model, it completes pixel-level material spectral attribution and RGB three-channel physical radiometric inversion. Relying on the inherent spectral properties of objects, it restores true colors, completely solving the problems of color homogenization, color shift, and poor scene adaptability of traditional technologies. Simultaneously, it constructs a three-dimensional residual vector field of spectrum, polarization, and radiation to accurately identify three types of imaging defects: color distortion, registration distortion, and noise reduction parameter mismatch. It then iteratively corrects core parameters in a targeted manner, forming a complete imaging closed loop. This effectively solves the pain points of unstable imaging and the inability to self-correct image quality deviations in complex night vision scenarios, achieving highly realistic, consistent, and stable AI night vision full-color imaging effects in all weather conditions and complex interference environments. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0022] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," and "horizontal," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0023] This embodiment provides a night vision full-color imaging control method. First, based on a high-precision self-developed coaxial polarization beam splitting synchronous imaging optical path, the hardware strictly ensures that the two imaging optical paths are coaxial, have the same field of view, are at the same moment, and have the same resolution. Within each frame imaging exposure cycle, under the same imaging moment and within the same complete field of view coverage, the hardware synchronously and in parallel acquires linearly polarized modulated low-light images and broadband infrared radiation images of the night vision target scene.

[0024] Among them, the linearly polarized modulated low-light image is a weak visible light imaging data that has been polarized and filtered by a polarization beam splitter to remove interference from random polarized light and retain the inherent polarization characteristics of the object surface in the scene. It effectively preserves the texture, material, and subtle edge information of objects at night. The broadband infrared radiation image is an imaging data that covers a wide spectrum of infrared waves from 700nm to 1400nm and carries the thermal radiation distribution of all objects in the scene. It is not affected by nighttime lighting, shadows, low light, or haze, and stably provides scene structure contour information. The two images are from the same hardware source, have completely overlapping fields of view, and correspond one-to-one with pixel positions, eliminating inter-frame offset, field of view deviation, and phase misalignment problems from the hardware source.

[0025] At the same time as the dual-channel image hardware synchronous acquisition, the system simultaneously acquires the scene polarization degree spatial distribution parameters, atmospheric scattered light intensity distribution parameters, and quantum noise time series data of the image sensor in real time at a high-frequency sampling frequency of 100Hz. Among them, the scene polarization degree spatial distribution parameters are used to characterize the polarization reflection, transmission, and attenuation characteristics of the air medium and object surface material at night on a pixel-by-pixel basis, which is the core physical parameter for distinguishing different object materials. The atmospheric scattered light intensity distribution parameters are used to characterize the light scattering, attenuation, and color shift interference characteristics caused by aerosols, fog, and suspended particles at night on a pixel-by-pixel basis. The sensor quantum noise time series data is used to continuously record the temporal random fluctuation law of the inherent electronic noise in the photoelectric conversion process of the image sensor in low light environment, accurately characterize the original noise characteristics under low illumination at night, and provide comprehensive and high-precision pre-physical constraint data for subsequent accurate registration, adaptive dynamic noise reduction, physical color inversion, and closed-loop error correction. This fundamentally solves the problems of inaccurate imaging, color distortion, and inaccurate noise reduction caused by traditional algorithms relying only on image grayscale information and lacking physical parameter constraints.

[0026] Furthermore, the acquired linearly polarized low-light images and broadband infrared radiation images are subjected to pixel-by-pixel, cross-dimensional, heterogeneous dual-domain feature decomposition processing. The unique polarization angle and polarization intensity gradient features of the linearly polarized low-light images and the unique thermal radiation intensity and topological contour ridge features of the broadband infrared images are extracted respectively, and the irreplaceable dual-domain core effective features of the two types of heterogeneous images are completely extracted. Based on the image local gradient entropy adaptive block strategy, a non-rigid mutual information topological mesh with elastic deformation capability is dynamically constructed. This overcomes the inherent defects of traditional fixed feature point registration methods such as SIFT, ORB, and Harris, which rely on gray-level saliency, fail in weak texture scenes, have sparse registration points, and cannot adapt to non-rigid deformation. The adaptive topological accurate registration and correction of the two images is completed through global topological association + local elastic fine-tuning, realizing pixel-level one-to-one correspondence and accurate alignment of the two heterogeneous modal images, generating a high-precision registered and coupled image without misalignment, geometric distortion, phase deviation, or local stretching.

[0027] This step completely solves the technical pain points of traditional dual-light registration in weak texture scenes such as nighttime roads, sky, and solid-color walls, including registration failure, large-area pixel misalignment, detail shift, and ghosting. It adopts a mechanism of dual-domain physical feature deep decomposition and non-rigid topological mesh adaptive matching. This significantly improves the accuracy of heterogeneous image fusion and provides a high-fidelity, unbiased pre-image input foundation for subsequent high-precision noise reduction and real physical color reconstruction.

[0028] Furthermore, a spatiotemporal variational non-uniform entropy field denoising model adapted to complex night vision scenarios in all weather conditions is constructed, overcoming the inherent technical shortcomings of traditional global unified denoising and fixed threshold denoising methods such as Gaussian filtering, guided filtering, mean filtering, and BM3D filtering. The sensor quantum noise temporal sequence data obtained from multiple consecutively sampled frames is deeply fused pixel-by-pixel with the spatial texture entropy field of the registered coupled image. Simultaneously, considering the complexity of the local spatial texture of the image and the long-term temporal noise fluctuation pattern of pixels, a dynamic adaptive entropy threshold is independently generated for each pixel of the image. Based on the texture characteristics of the image region, fine-grained partitioning is performed, and the registered coupled image is divided... This precise, regional, differentiated, and non-linear noise reduction method accurately distinguishes and targets four typical types of nighttime noise unique to nighttime scenes: sensor quantum random noise, atmospheric stray light-derived noise, lens-fixed dark noise, and pixel defects. During the noise reduction process, the noise reduction intensity is weakened and details are preserved in edge structure and fine texture areas, while the noise reduction intensity is strengthened and noise is completely removed in flat background areas. While fully preserving image edge contours, micro-texture details, and target structural information, it achieves full-domain, deep, clean, and residue-free noise reduction, ultimately resulting in a low-noise night vision base image with high signal-to-noise ratio, high fidelity, ultra-clear clarity, and no residual noise. This step completely solves the dual technical shortcomings of traditional noise reduction algorithms: either excessive noise reduction leading to blurred images, smoothed details, and blurred object edges, or insufficient noise reduction leading to residual noise and numerous image blemishes. It significantly improves the clarity, detail restoration, and image purity of nighttime images.

[0029] Furthermore, a multi-material spectral fingerprint knowledge base adapted to complex night vision scenarios in all weather conditions is pre-built. This knowledge base is entirely based on real physical spectral experimental data collection, without relying on any deep learning training samples or involving sample fitting and AI pseudo-coloring. Real-time collected scene polarization degree spatial distribution parameters and atmospheric scattered light intensity distribution parameters are used as dual dynamic physical constraints to correct the nighttime attenuation deviation of spectral parameters in real time. Pixel-by-pixel material spectral feature precise matching and material attribution discrimination are performed on the denoised high-fidelity, low-noise night vision base image to accurately distinguish the real scene material attributes corresponding to each pixel. Based on the standard atmospheric radiative transfer physics equation and combined with real-time nighttime environmental parameters, pixel-by-pixel physical spectral radiometric inversion is completed, completely eliminating the sample mapping pseudo-color defects of traditional AI coloring. This achieves real physical color reconstruction without training samples, fitting bias, or color homogenization, generating an initial physical full-color image with colors that closely match the real daytime scene, rich layers, clear material distinction, and no distortion. This step thoroughly solves the core problems of traditional night vision coloring—color distortion, material color homogenization, night scene color cast, poor scene adaptability, and severe pseudo-coloring—from the perspective of physical imaging mechanisms.

[0030] Furthermore, for the generated initial physical full-color image, precise imaging parameters in the spectral, polarization, and thermal radiation dimensions are extracted pixel by pixel across the entire domain simultaneously. A three-dimensional residual vector field covering the entire image is constructed to achieve full-domain quantification and structured characterization of multi-dimensional imaging errors. Through multi-dimensional statistical characteristics such as residual amplitude ratio, spatial distribution morphology, dispersion, and regional coherence, the specific distortion type of the current image is accurately determined, and three core imaging defects—color spectral distortion, registration misalignment distortion, and noise reduction parameter mismatch distortion—are precisely distinguished. Based on different distortion types, the topology registration grid elastic parameters, entropy field noise reduction dynamic threshold parameters, and spectral color reconstruction radiation compensation coefficients are iteratively corrected in a targeted and differentiated manner. This forms a full-link autonomous optimization mechanism of hardware synchronous acquisition → dual-domain topology registration → spatiotemporal entropy field noise reduction → physical spectral colorization → three-dimensional residual verification → parameter closed-loop correction, continuously iterating to converge image quality errors, and finally outputting a high-fidelity, high-consistency, high-stability, and high-detail final night vision full-color imaging result, achieving adaptive, self-detection, self-correction, and self-optimization imaging effects in complex night scenes.

[0031] The process involves cross-domain feature decomposition of the linearly polarized modulated low-light image and the broadband infrared radiation image, extraction of dual-domain core features, and construction of a non-rigid mutual information topological mesh to achieve adaptive topological registration of the two images and generate pixel-level one-to-one corresponding registration coupled images. This includes: performing full-domain pixel-by-pixel and four-neighbor gradient calculations on the linearly polarized modulated low-light image, traversing the polarization vector angle changes and polarization intensity changes at all coordinate positions of the image, and accurately extracting the real-time polarization angle parameters at each pixel position of the image; eliminating polarization angle jump errors caused by single-point noise through a 3×3 neighbor pixel weighted smoothing fitting method, and generating continuous, unbroken, highly stable, and detail-rich polarization angle gradient field features. These polarization gradient field features can accurately reflect the subtle differences in surface roughness, contour edges, texture undulations, and surface changes of objects at night, effectively compensating for the defects of single-level and weak detail representation in low-light grayscale images, and providing unique feature support for registration of weak texture scenes.

[0032] Simultaneously, the broadband infrared radiation image is subjected to a full-domain pixel-by-pixel radiation extreme value traversal detection. The radiation intensity difference between the image and its eight neighboring pixels is compared pixel by pixel to accurately extract the radiation intensity topological ridge features at the image's light-dark boundary, object contour boundary, target abrupt change location, and thermal radiation abrupt change location. These ridge features are generated entirely based on the object's own thermal radiation differences and are not affected by nighttime illumination, shadows, low light, fog, or occlusion. The features are extremely stable and can stably characterize the structural boundaries and spatial positions of all objects at night, providing highly robust structural prior information for accurate registration of dual-modal images.

[0033] Subsequently, the two heterogeneous images to be registered are simultaneously subjected to adaptive dynamic block processing, dividing them into mutually information unit blocks whose size can be dynamically adjusted. The size of the unit block is adaptively scaled and adjusted in real time according to the local image gradient entropy value. Specifically, for complex areas with dense texture, complex edges, rich details, and high gradient entropy values, the system automatically reduces the size of the unit block, which can be adjusted to a minimum of 4×4 pixels to improve the accuracy of local fine registration. For background areas with flat images, uniform grayscale, no obvious texture, and low gradient entropy values, the system automatically enlarges the size of the unit block, which can be adjusted to a maximum of 32×32 pixels to significantly improve the registration operation efficiency, achieve a dynamic optimal balance between imaging accuracy and operation speed, and adapt to night vision scenes of different complexity.

[0034] For each adaptive mutual information unit block, the joint correlation between the distribution of low-light polarization features and the distribution of infrared radiation features within the block is calculated. After normalization, a joint mutual information matrix specific to the unit block is generated. The matching correlation strength of the dual-domain features in the local area is accurately quantified, and the optimal pixel corresponding to the maximum mutual information value within the matrix is ​​selected as the cross-domain topological anchor point. This type of topological anchor point is different from traditional grayscale feature points. It does not depend on the significance of pixel grayscale brightness and darkness, but is determined entirely by the intrinsic correlation of dual-domain heterogeneous features. Even in low-texture, low-difference scenes such as nighttime roads, skies, and walls, it can still stably, densely, and uniformly extract matching anchor points, completely solving the problem of no effective feature points in weak-texture scenes in traditional registration methods.

[0035] Using each selected high-reliability topological anchor point as the core node, the eight-neighborhood connectivity criterion is adopted to radiate and connect all surrounding related pixels, constructing a non-rigid mutual information topological mesh with elastic deformation, local fine-tuning, and global convergence characteristics, generating many-to-one accurate topological edges between pixels; through the global constraint capability of the topological mesh, stable association matching of pixels in weak feature and weak texture regions is achieved, effectively making up for the shortcomings of traditional feature registration that relies only on sparse key points and cannot cover all pixels, and completely solving the problems of large-area misalignment, blurring, and ghosting in weak texture regions.

[0036] Finally, the constructed topological mesh is subjected to at least five rounds of iterative elastic deformation optimization. In each round of iteration, the coordinates of the mesh nodes and the local deformation offset are finely adjusted point by point to gradually correct small field deviations, pixel geometric distortions, gray-level phase shifts, and local stretching distortions. After multiple rounds of convergence, high-precision pixel alignment of the two heterogeneous images is achieved, with the pixel correspondence error controlled within 0.1 pixels. Finally, a high-precision registered and coupled image with one-to-one pixel position correspondence, perfect structural fit, no edge offset, and no misalignment distortion is generated, providing zero-deviation high-quality input for subsequent accurate noise reduction and physical color reconstruction.

[0037] The step of calculating the joint mutual information matrix of polarization features and infrared radiation features within each mutual information unit block, and selecting the pixel points corresponding to the maximum mutual information values ​​in the matrix as cross-domain topological anchor points, includes: accurately extracting multi-dimensional polarization feature datasets such as polarization angle, polarization intensity, and polarization gradient corresponding to all micro-polarized pixels within a single mutual information unit block, as well as multi-dimensional radiation feature datasets such as radiation intensity, radiation gradient, and neighborhood radiation difference corresponding to all infrared pixels, to ensure that the feature data of a single block is complete, without omissions or missing data.

[0038] The Gaussian kernel density estimation method is used to statistically calculate the continuous probability distribution density of the two sets of feature datasets to avoid the loss of precision caused by discrete statistics. Based on the principle of information entropy, the edge entropy of polarization feature and infrared radiation feature are calculated respectively. The joint entropy between the two types of heterogeneous features is further solved. The mutual information value of any pixel combination within the unit block is accurately calculated by the difference between the joint entropy and the edge entropy, and the intrinsic correlation matching degree of dual-domain pixels is accurately quantified.

[0039] Iterate through the mutual information values ​​of all pairs of pixels within a unit block, construct a complete two-dimensional joint mutual information matrix according to the correspondence of pixel two-dimensional spatial coordinates. The horizontal and vertical coordinates of the matrix strictly correspond to the spatial position of the pixels, and the matrix values ​​correspond to the matching strength of the dual-domain feature association, thus completely, accurately and comprehensively representing the matching distribution characteristics of the local region.

[0040] The peak threshold of mutual information of unit blocks is dynamically set adaptively based on the real-time illumination conditions of the current night vision scene. The lower the illumination and the more complex the scene, the threshold is adaptively lowered to ensure that a sufficient number of effective anchor points can still be extracted in low-light scenes. All pixel coordinates in the filter matrix that exceed the peak threshold are used as the initial candidate anchor point set to initially remove low-association, invalid matching, and pseudo-matching pixels.

[0041] The initial candidate anchor point set is optimized through a dual screening process. First, coordinate deduplication is performed to remove duplicate anchor points. Then, a neighborhood non-maximum suppression algorithm is used for neighborhood screening to remove locally clustered redundant anchor points and weakly correlated pseudo-anchor points. Pixels with uniform spatial distribution, the highest correlation between dual-domain features, the strongest feature significance, and the best matching stability are retained as the final cross-domain topological anchor points. This ensures that the anchor points are uniformly distributed across the entire domain, fully cover high and low texture regions, have high matching accuracy, and strong robustness, greatly improving the accuracy and global stability of subsequent topological registration.

[0042] The proposed spatiotemporal variational non-uniform entropy field denoising model fuses the temporal sequence data of sensor quantum noise with the spatial entropy field of the registered and coupled image to generate a pixel-level dynamic adaptive entropy threshold. This model performs differentiated denoising suppression on the registered and coupled image across different regions, eliminating quantum noise, stray light noise, and fixed dark noise points to obtain a high-fidelity, low-noise night vision base image. This includes: accurately calculating the spatial information entropy of the registered and coupled image pixel by pixel and across the entire image; generating a continuous and smooth spatial variational entropy field based on the disorder of the grayscale distribution of 8×8 neighboring pixels; and refining the image region according to a preset entropy value range threshold. The image is automatically classified into three types of feature regions: high-structure entropy regions (object edges, contours, abrupt change regions), texture detail entropy regions (vegetation, road texture, building detail regions), and low-noise flat entropy regions (sky, solid-color walls, flat road surfaces). This achieves refined and differentiated scene region discrimination, providing accurate regional basis for regional denoising.

[0043] The system continuously extracts the temporal sequence of quantum noise from the image sensor for a preset number of consecutive frames (configurable to 5–15 frames), statistically analyzes the small fluctuations, random jumps, and amplitude perturbations of grayscale over a long time sequence pixel by pixel, calculates the temporal quantum noise entropy corresponding to each pixel, accurately quantifies the temporal noise activity and intensity of each pixel, constructs a complete temporal entropy distribution matrix, and accurately characterizes the spatiotemporal distribution law of the sensor's inherent quantum noise under low light conditions.

[0044] By performing strict pixel-level one-to-one weighted fusion of the global spatial variational entropy field and the temporal entropy distribution matrix, and simultaneously taking into account the complexity of the image spatial texture structure and the long-term temporal noise fluctuation characteristics of pixels, a high-precision spatiotemporal joint entropy density map of the entire image is generated, which enables precise localization and quantitative differentiation of noise location, noise intensity, and noise type.

[0045] Based on the pixel dynamic entropy threshold calculation formula of this invention, it is independent of any fixed global parameters and generates an exclusive adaptive noise reduction threshold for each pixel of the image, truly realizing fine noise reduction control with one threshold per pixel, scene adaptation, and noise adaptation.

[0046] For the three regions that have been divided, nonlinear noise reduction and suppression operators with different response characteristics are matched: the high structure entropy region adopts the edge-weighted noise reduction operator to weaken the noise reduction intensity and prioritize the integrity of the contour structure; the texture detail entropy region adopts the fine texture preservation noise reduction operator to slightly smooth noise and completely preserve micro-texture; the low noise flat entropy region adopts the global depth noise reduction operator to completely remove residual noise in flat areas; based on the independent threshold of each pixel, the quantum random noise, atmospheric stray light-derived noise and lens fixed dark noise in the corresponding region are precisely suppressed, and the image contour structure and micro-texture details are completely preserved throughout the process, eliminating the problems of detail smoothing, image blurring and edge blurring, and finally obtaining a high-fidelity, high-definition, noise-free and distortion-free low-noise night vision base image.

[0047] The pixel dynamic entropy threshold calculation formula generates an independent adaptive noise reduction threshold for each pixel of the image. The pixel dynamic entropy threshold calculation formula is as follows: Where T(x,y) is the dynamic adaptive noise reduction threshold corresponding to the pixel at coordinates (x,y), T0 is the system-preset baseline noise reduction threshold used to adapt to the basic night vision noise reduction intensity, which can be finely adjusted according to the night illumination level; Hs(x,y) is the spatial variational entropy value corresponding to the pixel, representing the local texture complexity, the richer the texture, the higher the entropy value; Ht(x,y) is the temporal quantum noise entropy value corresponding to the pixel, representing the temporal noise activity intensity of the pixel, the more intense the noise, the higher the entropy value; ε is a very small constant approaching 0 (value 10). -6 ), used to prevent division by zero errors and ensure stable and error-free operation throughout the entire process; σ 2 (x,y) represents the local variance of pixel grayscale within the 8×8 neighborhood window of this pixel, characterizing the degree of drastic local grayscale fluctuation; σ0 2 The system's preset standard reference variance is used to normalize local fluctuation characteristics.

[0048] This formula jointly regulates the threshold through three dimensions: positive regulation of spatial texture entropy, inverse constraint of temporal noise entropy, and correction of local variance exponent. This allows the threshold to be adaptively lowered in complex texture areas to protect details from being blurred; the threshold to be adaptively raised in areas with drastic noise fluctuations to enhance noise reduction capabilities; and the threshold to be adaptively adapted in flat and stable areas to completely eliminate noise. Different pixels automatically generate differentiated optimal thresholds based on their own spatial texture features and temporal noise features, achieving adaptive and precise noise reduction across the entire domain without fixed thresholds or globally unified parameters, while simultaneously ensuring both extreme noise reduction effects and ultra-high detail fidelity.

[0049] The pre-built multi-material spectral fingerprint knowledge base adapted to night vision scenarios, combined with the spatial distribution parameters of scene polarization degree and atmospheric scattered light intensity distribution parameters as physical constraints, performs pixel-level material spectral attribution discrimination and physical spectral radiometric inversion on high-fidelity, low-noise night vision base images, completes real color reconstruction without sample mapping, and generates an initial physical full-color image. This includes: pre-collecting standard samples of six typical materials that frequently appear in night vision scenarios—vegetation, buildings, roads, sky, metals, and water bodies—in a standard darkroom environment and in multi-time outdoor night vision environments; accurately testing the complete spectral response curves of each material in the low-light polarization band, infrared radiation band, and visible light color band; accurately extracting unique spectral feature parameters such as spectral absorption peaks, radiation gain coefficients, band attenuation laws, and polarization response characteristics of each material; and storing these parameters in a structured manner according to material categories to construct a complete, accurate, and dynamically calibrated multi-material spectral fingerprint knowledge base. This knowledge base is entirely based on real physical spectral experimental collection and does not rely on any training samples, possessing extremely strong physical realism, scene generalization, and anti-interference capabilities.

[0050] By using the real-time collected spatial distribution of scene polarization degree and atmospheric scattered light intensity distribution as dual dynamic physical constraints, the nighttime attenuation coefficient and atmospheric radiation compensation coefficient of various materials in the knowledge base are dynamically corrected according to real-time environmental parameters. This offsets the spectral distortion deviation caused by nighttime air scattering, polarization attenuation, and light absorption, allowing the standard parameters of the knowledge base to adapt to the current real nighttime environment in real time and eliminating color deviation caused by fixed parameters.

[0051] The polarization response features and infrared radiation features of each pixel in a high-fidelity, low-noise night vision substrate image are extracted pixel by pixel and fused to generate a polarization-infrared joint spectral feature vector, which comprehensively and accurately characterizes the optical properties of the object material corresponding to the current pixel. The real-time pixel feature vector is matched with the standard fingerprint curves of various materials in the spectral fingerprint knowledge base using physical constraints. Unlike traditional pixel grayscale matching, this solution is based on underlying physical optical feature matching, which is not affected by illumination, brightness, or noise, and has extremely high matching accuracy.

[0052] Based on the optimal matching result of the minimum physical matching distance across the entire area, the true material category corresponding to each pixel is accurately determined, and pixel-level fine-grained material spectral classification is completed, realizing pixel-by-pixel accurate differentiation of materials across the entire image, providing accurate material basis for subsequent color physical inversion.

[0053] Based on the standard atmospheric radiative transfer physical model, combined with real-time atmospheric scattering and polarization attenuation environmental parameters, and calling the standard spectral fingerprint parameters of the corresponding pixel material, the true physical color radiation values ​​of the RGB three channels are retrieved pixel by pixel. Through band energy compensation, spectral attenuation correction, and polarization color difference correction, the physical reconstruction of the entire domain color is completed, completely getting rid of the sample fitting deviation, color homogenization, and pseudo-color problems of traditional AI coloring, and generating an initial physical full-color image with true color, rich layers, clear material distinction, and close to the true visual perception of the human eye.

[0054] The process involves extracting the polarization-infrared joint spectral feature vectors of each pixel in a high-fidelity, low-noise night vision substrate image, and then performing a physical constraint-based matching between these feature vectors and the standard fingerprint curves of various materials in the spectral fingerprint knowledge base. This includes: precisely collecting four core physical feature parameters of a single pixel on a pixel-by-pixel basis, namely polarization response amplitude, polarization phase shift, infrared radiation intensity, and infrared band attenuation coefficient; and then sequentially splicing and normalizing these four heterogeneous physical parameters to form a pixel-specific four-dimensional joint spectral feature vector, which comprehensively and multidimensionally characterizes the optical properties of the pixel material.

[0055] Three types of physical penalty factors are introduced: polarization deviation penalty, atmospheric scattering attenuation penalty, and infrared radiation deviation penalty. A multi-dimensional physical matching distance formula is constructed to constrain and offset the feature deviation caused by nighttime environmental interference in real time during the matching calculation, thereby avoiding material mismatch and incorrect matching caused by nighttime stray light, polarization attenuation, and fog scattering.

[0056] The physical matching distance between the current pixel's four-dimensional feature vector and the spectral fingerprint of each material in the knowledge base is calculated pixel by pixel. The smaller the matching distance value, the higher the consistency between the current pixel's optical features and the corresponding material's standard spectral features, and the higher the material matching degree.

[0057] By traversing all material categories, the material fingerprint corresponding to the minimum matching distance is selected as the spectral fingerprint of the current pixel, thus completing high-precision material identification for a single pixel. This ensures that the material basis for the full-domain color inversion is real and reliable, and completely eliminates color confusion, material homogenization, and night scene color cast problems from the root.

[0058] The step of constructing a three-dimensional residual vector field of spectrum, polarization, and radiation for the initial physical full-color image and determining the image distortion type through residual distribution characteristics includes: performing pixel-by-pixel precise difference calculations on the pixel-by-pixel color spectral parameters, polarization response parameters, and infrared radiation parameters of the initial physical full-color image and the original physical parameters of the denoised high-fidelity low-noise night vision substrate image, respectively, to obtain the global spectral residual matrix, polarization residual matrix, and radiation residual matrix, and to accurately quantify the magnitude and direction of the three-dimensional imaging error of each pixel.

[0059] By precisely coupling spatial coordinates and fusing data of the three-dimensional residual matrix, a three-dimensional residual vector field covering the entire image is constructed, enabling full-domain visualization, quantification, and structured characterization of imaging errors, and accurately locating the error distribution area and error magnitude.

[0060] By statistically analyzing the spatial distribution skewness, pixel dispersion, and regional coherence of the residual vector field region by region, the distribution pattern, source, and distortion type of error can be accurately determined.

[0061] The specific distortion determination logic is as follows: If the spectral residual amplitude accounts for the largest proportion and the residual distribution is scattered and irregularly discrete, it is determined to be color spectral distortion, corresponding to the problem of insufficient adaptation of spectral inversion parameters and color shift caused by inaccurate atmospheric compensation; if the polarization residual amplitude accounts for the largest proportion and the residual is locally clustered and concentrated at the edge of the object, exhibiting local clustering characteristics, it is determined to be image registration distortion, corresponding to the problem of minor deviation in topological mesh matching and slight pixel misalignment; if the radiation residual amplitude accounts for the largest proportion and the residual is uniformly offset across the entire domain with consistent overall deviation, it is determined to be distortion due to denoising parameter mismatch, corresponding to the problem of mismatch between the global denoising threshold and scene noise characteristics.

[0062] The method of correcting the topology registration parameters, entropy field noise reduction threshold, and spectral color reconstruction coefficients based on the distortion type includes: when a color spectral distortion is determined, the system automatically fine-tunes the atmospheric scattering attenuation compensation coefficient and polarization spectral correction coefficient corresponding to the spectral fingerprint knowledge base, and re-executes the pixel-level material spectral inversion and global color reconstruction operation to accurately correct color shift, color cast, and color dimming caused by nighttime environmental interference, and restore the true original color of various materials.

[0063] When an image registration distortion is detected, the system adaptively adjusts the elastic stiffness coefficient of the non-rigid mutual information topology mesh and the anchor point matching threshold, optimizes the mesh deformation constraint, improves the anchor point matching accuracy, and re-completes the iterative optimization of dual-image topology registration, completely eliminating registration distortion problems such as subtle pixel misalignment, edge offset, and local ghosting.

[0064] When the noise reduction parameters are determined to be mismatched or distorted, the system adaptively corrects the baseline threshold and standard reference variance parameters in the dynamic entropy threshold calculation formula, accurately adapts to the real noise characteristics of the current scene, regenerates the global pixel-level adaptive noise reduction threshold, and optimizes the image noise reduction effect a second time. While thoroughly filtering out various types of nighttime noise, it fully preserves the scene texture details, ensuring the ultimate clarity and richness of the image.

[0065] The polarization-splitting synchronous imaging optical path simultaneously acquires linearly polarized modulated low-light images and broadband infrared radiation images of the night vision target scene at the same imaging time and within the same field of view. This includes: employing a high-precision industrial-grade coaxial polarization-splitting imaging structure to ensure that the incident light path is strictly coaxial, without eccentricity or offset; physically splitting the nighttime incident light from the scene using a high-precision polarization-splitting prism, resulting in stable beam splitting ratios, balanced optical path loss, and minimal spectral deviation; one path of weak visible light, after polarization filtering, is connected to a high-sensitivity low-light imaging sensor to acquire a linearly polarized modulated low-light image with complete polarization characteristics of the scene; the other path of infrared light is connected to a broadband infrared imaging sensor to acquire a broadband infrared radiation image covering complete thermal radiation information of the scene.

[0066] During the factory calibration and self-calibration phases of the equipment, the pixel resolution, imaging field of view, optical center coordinates, imaging focal length, and imaging frame rate of the two sensors are strictly and uniformly calibrated to ensure that the spatial dimensions of the two images are completely aligned and the pixels correspond one-to-one. In the temporal dimension, there is no inter-frame time difference or timing misalignment, achieving true hardware-level, physical-level, and high-precision synchronous acquisition.

[0067] The entire process abandons the traditional night vision imaging methods that are prone to errors, such as mechanical scanning and image acquisition, post-processing software pixel interpolation, multi-frame temporal synthesis, and image stretching and correction. It fully realizes hardware physical synchronous acquisition of dual-domain images, completely eliminating parallax, misalignment, pseudo-pixels, temporal deviation, and artificial distortion from the source of acquisition. This provides real, accurate, reliable, and distortion-free raw imaging data support for backend topology registration, precise noise reduction, and physical color reconstruction.

[0068] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A night vision full-color imaging control method, characterized in that, The method includes: based on a polarization-splitting synchronous imaging optical path, simultaneously acquiring linearly polarized modulated low-light images and broadband infrared radiation images of a night vision target scene at the same imaging time and within the same field of view, and simultaneously acquiring scene polarization degree spatial distribution parameters, atmospheric scattered light intensity distribution parameters, and quantum noise time sequence data of the image sensor in real time. Cross-domain feature decomposition is performed on the linearly polarized modulated low-light image and the broadband infrared radiation image. Dual-domain core features are extracted and a non-rigid mutual information topological mesh is constructed to complete the adaptive topological registration of the two images and generate a pixel-level one-to-one corresponding registration coupling image. A spatiotemporal variational non-uniform entropy field denoising model is constructed. The time-series data of quantum noise from the sensor is fused with the spatial entropy field of the registered and coupled image to generate a pixel-level dynamic adaptive entropy threshold. The registered and coupled image is subjected to regional differential denoising suppression to remove quantum noise, stray light noise and fixed dark noise points, resulting in a high-fidelity low-noise night vision base image. A multi-material spectral fingerprint knowledge base adapted to night vision scenarios is pre-built. The spatial distribution parameters of scene polarization degree and atmospheric scattered light intensity distribution parameters are combined as physical constraints. Pixel-level material spectral attribution and physical spectral radiometric inversion are performed on high-fidelity, low-noise night vision base images to complete the real color reconstruction without sample mapping and generate an initial physical full-color image. A three-dimensional residual vector field of spectrum, polarization, and radiation is constructed for the initial physical full-color image. The image distortion type is determined by the residual distribution characteristics. Based on the distortion type, the topological registration parameters, entropy field noise reduction threshold, and spectral color reconstruction coefficient are iteratively corrected in reverse. After closed-loop optimization, the final night vision full-color imaging result is output.

2. The night vision full-color imaging control method according to claim 1, characterized in that, The step of performing cross-domain feature decomposition on the linearly polarized modulated low-light image and the broadband infrared radiation image, extracting dual-domain core features and constructing a non-rigid mutual information topological mesh to complete the adaptive topological registration of the two images and generate a pixel-level one-to-one corresponding registration coupled image includes: performing global gradient calculation on the linearly polarized modulated low-light image, extracting the polarization angle parameters at each pixel position of the image, and generating a continuous and stable polarization angle gradient field feature. Radiation extrema detection is performed on the broadband infrared radiation image to extract the radiation intensity topological ridge features at the boundary between light and dark areas of the image and the contour of the object. The two images to be registered are divided into mutual information unit blocks with adaptive size adjustment. The size of the unit block is dynamically scaled according to the local image gradient entropy value. The unit block size is reduced in areas with complex textures and enlarged in areas with flat textures. Calculate the joint mutual information matrix of polarization features and infrared radiation features within each mutual information unit block, and select the pixel points corresponding to the maximum mutual information values ​​in the matrix as cross-domain topological anchor points. With each topological anchor point as the core, radiating connections are made to neighboring related pixels to construct a non-rigid mutual information topological mesh with elastic deformation characteristics, generating many-to-one related topological edges between pixels. The topological mesh is subjected to multiple rounds of elastic deformation iterative optimization to successively eliminate field deviation, pixel distortion and phase shift between the two images, and finally achieve pixel alignment of the two images to generate a registration coupled image.

3. The night vision full-color imaging control method according to claim 2, characterized in that, The step of calculating the joint mutual information matrix of polarization features and infrared radiation features within each mutual information unit block, and selecting the pixel points corresponding to the maximum mutual information values ​​in the matrix as cross-domain topological anchor points, includes: extracting the polarization feature dataset of all low-light polarized pixels and the radiation feature dataset of all infrared pixels within a single mutual information unit block. The probability distribution density of the two datasets is statistically analyzed, and the joint entropy and edge entropy between the two types of features are calculated to obtain the mutual information value of pixels within the unit block. Traverse the mutual information values ​​of all pixel combinations within the unit block to construct a two-dimensional joint mutual information matrix; Set a peak threshold for mutual information between unit blocks, and select pixel coordinates in the matrix that exceed the peak threshold as initial anchor points; The initial anchor points are deduplicated and their neighborhoods are filtered, and the pixels with the most uniform distribution and the strongest feature significance are retained as the final cross-domain topology anchor points.

4. The night vision full-color imaging control method according to claim 1, characterized in that, The construction of the spatiotemporal variational non-uniform entropy field denoising model fuses the temporal sequence data of sensor quantum noise with the spatial entropy field of the registration coupled image to generate a pixel-level dynamic adaptive entropy threshold. It then performs regional differential denoising suppression on the registration coupled image, removing quantum noise, stray light noise, and fixed dark noise points to obtain a high-fidelity, low-noise night vision base image. This includes: calculating the spatial information entropy of the registration coupled image pixel by pixel to generate a global continuous spatial variational entropy field, and dividing the image into three regions based on the entropy value: a high-structure entropy region, a texture detail entropy region, and a low-noise flat entropy region. Extract the temporal sequence of quantum noise from the image sensor for a preset number of consecutive frames, statistically analyze the temporal grayscale fluctuation characteristics pixel by pixel, calculate the temporal quantum noise entropy corresponding to each pixel, and construct the temporal entropy distribution matrix. The spatial variational entropy field and the temporal entropy distribution matrix are fused at the pixel level to generate a global spatiotemporal joint entropy density map of the image. Based on the pixel dynamic entropy threshold calculation formula, an independent adaptive noise reduction threshold is generated for each pixel of the image. Nonlinear noise reduction and suppression operators are matched for high structure entropy region, texture detail entropy region and low noise flat entropy region respectively. Quantum random noise, atmospheric stray light-derived noise and fixed dark noise in the corresponding region are suppressed according to the independent threshold of each pixel. The image contour structure and micro texture details are preserved throughout the process to obtain a high-fidelity low noise night vision base image.

5. The night vision full-color imaging control method according to claim 4, characterized in that, The pixel dynamic entropy threshold calculation formula generates an independent adaptive noise reduction threshold for each pixel of the image. The pixel dynamic entropy threshold calculation formula is as follows: Where T(x,y) is the dynamic adaptive noise reduction threshold corresponding to the pixel at coordinates (x,y), T0 is the system-preset baseline noise reduction threshold, and H s (x,y) represents the spatial variational entropy value corresponding to this pixel, H t (x,y) represents the temporal quantum noise entropy value corresponding to this pixel, which is a very small constant approaching 0, used to prevent division by zero errors in calculations. This represents the local variance of pixel grayscale within the neighborhood window of that pixel. The system's preset standard reference variance; Different pixels generate differentiated thresholds based on their own spatial texture features and temporal noise features, achieving adaptive noise reduction without fixed thresholds or globally uniform parameters across the entire domain.

6. The night vision full-color imaging control method according to claim 1, characterized in that, The pre-built multi-material spectral fingerprint knowledge base adapted to night vision scenarios, combined with the spatial distribution parameters of scene polarization degree and atmospheric scattered light intensity distribution parameters as physical constraints, performs pixel-level material spectral attribution discrimination and physical spectral radiometric inversion on high-fidelity low-noise night vision base images, completes real color reconstruction without sample mapping, and generates an initial physical full-color image. This includes: pre-collecting various materials such as vegetation, buildings, roads, sky, metals, and water bodies in night vision scenarios, obtaining the spectral response curves of various materials in the low-light polarization band, infrared radiation band, and visible light color band, extracting the unique spectral feature parameters of each material, and constructing a multi-material spectral fingerprint knowledge base; The spatial distribution of scene polarization degree and atmospheric scattered light intensity are collected in real time as physical constraints to correct the attenuation coefficient and radiation compensation coefficient of spectral fingerprint. The polarization-infrared joint spectral feature vectors of each pixel in the high-fidelity, low-noise night vision substrate image are extracted, and the feature vectors are physically constrained to be matched with the standard fingerprint curves of various materials in the spectral fingerprint knowledge base. The material category corresponding to each pixel is determined based on the optimal matching result, and pixel-level material spectral classification is completed. Based on the atmospheric radiative transfer physical model, the standard spectral fingerprint parameters of the corresponding material are called to retrieve the true color radiance values ​​of the RGB three channels pixel by pixel, complete the physical reconstruction of the whole domain color, and generate the initial physical full-color image.

7. The night vision full-color imaging control method according to claim 6, characterized in that, The process involves extracting the polarization-infrared joint spectral feature vectors of each pixel in a high-fidelity, low-noise night vision substrate image, and performing physical constraint matching between the feature vectors and the standard fingerprint curves of various materials in the spectral fingerprint knowledge base. This includes collecting four-dimensional feature parameters of a single pixel, such as polarization response amplitude, polarization phase offset, infrared radiation intensity, and infrared band attenuation coefficient, and then splicing them together to form a pixel-specific joint spectral feature vector. By introducing polarization bias penalty terms, atmospheric scattering attenuation penalty terms, and infrared radiation bias penalty terms, a multi-dimensional physical matching distance formula is constructed. Calculate the physical matching distance between the current pixel feature vector and the spectral fingerprints of each material in the knowledge base. The smaller the matching distance, the higher the material matching degree. The material fingerprint corresponding to the minimum matching distance is selected as the spectral fingerprint of the current pixel to complete the material identification.

8. The night vision full-color imaging control method according to claim 1, characterized in that, The step of constructing a three-dimensional residual vector field of spectrum, polarization, and radiation for the initial physical full-color image and determining the image distortion type through residual distribution characteristics includes: calculating the pixel color spectrum parameters, polarization response parameters, and infrared radiation parameters of the initial physical full-color image with the original physical parameters of the high-fidelity low-noise night vision substrate image pixel by pixel difference, to obtain the global spectral residual matrix, polarization residual matrix, and radiation residual matrix, respectively. The three-dimensional residual matrices are spatially coupled to construct a global three-dimensional residual vector field; Spatial distribution skewness, pixel dispersion, and regional coherence of statistical residual vector fields; If the spectral residual amplitude accounts for the largest proportion and is discretely distributed, it is determined to be color spectral distortion; If the polarization residual amplitude accounts for the largest proportion and is locally clustered, it is determined to be image registration distortion; If the radiation residual amplitude accounts for the largest proportion and is uniformly offset across the entire domain, it is determined to be a distortion due to mismatch of noise reduction parameters.

9. The night vision full-color imaging control method according to claim 8, characterized in that, The method of correcting the topology registration parameters, entropy field noise reduction threshold and spectral color reconstruction coefficient based on the distortion type includes: when it is determined to be a color spectral distortion, fine-tuning the atmospheric scattering attenuation compensation coefficient of the spectral fingerprint knowledge base, re-executing the pixel-level material spectral inversion and color reconstruction operation, and correcting the color shift problem. When image registration distortion is detected, the elastic stiffness coefficient of the non-rigid mutual information topology mesh and the anchor point matching threshold are adjusted, and the dual image topology registration is completed again to eliminate pixel misalignment distortion. When the noise reduction parameter mismatch distortion is identified, the baseline threshold and standard reference variance parameter in the dynamic entropy threshold calculation formula are corrected, and a pixel-level adaptive noise reduction threshold is regenerated to optimize the image noise reduction effect and preserve complete texture details.

10. The night vision full-color imaging control method according to claim 1, characterized in that, The polarization-splitting synchronous imaging optical path, which simultaneously acquires linearly polarized modulated low-light images and broadband infrared radiation images of night vision target scenes at the same imaging time and within the same field of view, includes: using a coaxial polarization-splitting imaging structure, splitting the incident light rays of the scene through a polarization beam splitter, with one light ray connected to a low-light imaging sensor to acquire a linearly polarized modulated low-light image, and the other light ray connected to an infrared imaging sensor to acquire a broadband infrared radiation image; Strictly calibrate the pixel resolution, imaging field of view, optical center coordinates and imaging frame rate of the two sensors to ensure that the spatial dimensions of the two images are completely aligned and there is no inter-frame time difference in the temporal dimension. The entire process involves no mechanical scanning, no post-processing pixel interpolation, and no inter-frame synthesis, achieving physical synchronous acquisition of dual-domain images.

Images