A gray image recognition enhancement method and system based on physical knowledge fusion
By fusing physical knowledge through polarization preprocessing, thermodynamic ROI extraction, and wavelet enhancement, the problem of insufficient accuracy in grayscale image recognition under high noise environments is solved, achieving efficient and interpretable image recognition results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INSPUR ENTERPRISE CLOUD TECHNOLOGY (SHANDONG) CO LTD
- Filing Date
- 2026-02-27
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies lack sufficient accuracy in grayscale image recognition under high noise and complex environments. Traditional methods struggle to accurately define regions of interest in high noise environments, have limited feature extraction capabilities, and lack physical prior guidance in data-driven models, making it difficult to improve recognition accuracy and reliability.
An image enhancement method based on physical knowledge fusion is adopted, which combines Stokes parameters, heat conduction equation and Haar wavelet transform through polarization preprocessing, thermodynamic ROI extraction and wavelet enhancement.
It improves the feature extraction accuracy and robustness of grayscale image recognition, reduces computational complexity, and is applicable to fields such as industrial inspection and medical imaging, thereby improving recognition accuracy and interpretability.
Smart Images
Figure CN122391584A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a grayscale image recognition enhancement method and system based on physical knowledge fusion. Background Technology
[0002] Grayscale image recognition technology, as a core branch of computer vision, has been widely applied in key areas such as industrial inspection (e.g., microcrack detection in plate heat exchangers, PCB solder joint quality assessment), medical diagnosis (PET imaging, bone microstructure analysis), remote sensing monitoring, autonomous driving environmental perception, and OCR character recognition. Its low data volume and computational efficiency make it the preferred solution for resource-constrained scenarios. However, as applications expand towards higher precision and more complex environments, existing technologies are revealing multi-dimensional limitations, specifically as follows: (1) Feature extraction failure in high-noise environments In medical imaging (such as PET images), the high noise background caused by the recording of tissue physiological activities makes it difficult for traditional methods to accurately define regions of interest (ROIs), directly affecting the accuracy of lesion localization. In industrial scenarios, reflections or dynamic shadows on metal surfaces can cause grayscale values to fluctuate by ±0.15 (normalized to the [0,1] interval), resulting in a false detection rate as high as 21.85% for traditional threshold segmentation algorithms. A typical example is the missed detection of defects in overexposed areas of automotive parts.
[0003] (2) Insufficient small-scale defect detection capability For core industrial components (such as microcracks on the surface of plate heat exchangers), traditional methods often result in the complete loss of defect features with a width <50μm due to the use of width-adaptive smoothing coefficients to fill grayscale valleys. Even with the introduction of deep learning models, the false negative rate remains above 15% in scenarios with scarce data (e.g., only 200 labeled samples per month) or edge computing scenarios (computing power <10 TOPS). In laser stripe analysis, the traditional grayscale centroid method is affected by noise and threshold selection, resulting in a sub-pixel positioning error of 0.3 pixels, directly limiting the positioning accuracy of precision assembly.
[0004] (3) Limitations of the expressive power of traditional feature algorithms Mainstream feature extraction techniques have inherent defects: grayscale histograms can only reflect the global brightness distribution and are sensitive to noise and illumination changes. In PCB solder joint detection, the standard deviation of the matching score is ±0.15. Local binary mode (LBP) has poor robustness to rotation and scaling transformations. In a 90° rotation scene, the feature matching accuracy drops by 40%. Although scale-invariant feature transform (SIFT) has scale robustness, the feature extraction time for a single image is up to 200ms, which is difficult to meet the real-time detection requirements [5]
[14] . In addition, texture analysis tools such as grayscale co-occurrence matrix (GLCM) cannot distinguish material differences. For example, dark red and dark gray objects are often misjudged as the same type of target because of their similar reflectivity.
[0005] (4) Generalization defects of data-driven models Purely data-driven deep learning models face three major challenges: First, labeling dependence. In scenarios such as medical imaging, due to data scarcity (e.g., rare disease samples <100 cases), the convergence speed of models such as ResNet decreases by 60%, and the accuracy of the validation set decreases by 20%-30% compared to RGB input. Second, lack of physical alignment. Shallow CNN networks rely excessively on color features, resulting in "feature hunger" during grayscale training, leading to the loss of structural details in tasks such as bone microstructure anisotropy analysis. Third, insufficient interpretability. In industrial defect detection, the model's decision path is disconnected from the heat conduction law, making it difficult to trace the root cause of misjudgments.
[0006] In summary, existing technologies exhibit a dual deficiency in "data efficiency" and "physical consistency" in complex physical scenarios. Traditional algorithms are limited by manual feature design, and data-driven models lack physical prior guidance, making it difficult to overcome bottlenecks in recognition accuracy and reliability in industrial / medical scenarios with high noise, small sample sizes, and multiple interferences.
[0007] These limitations directly constrain development in key areas: in industrial inspection, the failure to detect microcracks in photovoltaic modules leads to a 35% increase in power plant operation and maintenance costs; in medical diagnosis, low-dose X-ray imaging reduces the early detection rate of emphysema by 28% due to insufficient contrast; and in the field of autonomous driving, accidents caused by grayscale camera failure in tunnels with strong light account for 12%. Therefore, constructing a grayscale image recognition enhancement method that integrates physical knowledge has become an inevitable requirement to overcome existing technological bottlenecks.
[0008] Based on this, the present invention proposes a grayscale image recognition enhancement method and system based on physical knowledge fusion. Summary of the Invention
[0009] To overcome the shortcomings of existing technologies, this invention provides a simple and efficient grayscale image recognition enhancement method based on physical knowledge fusion.
[0010] This invention is achieved through the following technical solution: A grayscale image recognition enhancement method based on physical knowledge fusion includes the following steps: Step S1: Polarization preprocessing Based on the transverse wave characteristics of light, images of objects under different polarization states are acquired, and Stokes parameters, including total intensity I, linear polarization difference Q, and diagonal polarization difference U, are calculated to obtain a polarization degree DOP / polarization angle AOP image to subtract reflective areas. In step S1, the original grayscale image is input, polarization preprocessing is performed on the original grayscale image, the degree of polarization DOP / polarization angle AOP is calculated through the Stokes parameter, and the polarization enhancement image is output after subtracting the reflective area. Among them, the degree of polarization (DOP) is used to quantify the polarization characteristics of the target, while the angle of polarization (AOP) reflects the distribution of polarization direction.
[0011] For highly reflective surfaces such as metals and glass, a pixel-level polarization component subtraction technique is employed: by identifying high-value areas of polarization degree (DOP) (usually corresponding to reflective areas), the polarization component intensity of these areas is subtracted from the original grayscale image, eliminating overexposure information while preserving texture details. In terms of hardware implementation, a single-layer silicon metasurface structure (etched with periodic triangular pores, α=924nm, R=265nm, H=315nm) is used to achieve directional edge detection of polarized light. With a numerical aperture of 0.35, it can effectively separate S-waves and P-waves to enhance edge contrast.
[0012] Step S2: Thermodynamic ROI Extraction The polarization enhancement map is mapped to the pseudo-matter temperature field, and the noise-suppressed ROI feature map is output through iterative optimization of the heat conduction equation. In step S2, a "pseudo-matter-energy diffusion" framework is constructed based on the heat conduction physics model to achieve noise suppression and feature fusion; A grayscale image pixel is defined as a pseudo-substance with a specific specific heat capacity, and pixels with similar grayscale values are assigned similar specific heat capacities. The heating-cooling process is simulated by solving the heat conduction equation, which is as follows: ∂T / ∂t=α²∇²T+Q Where α is the thermal diffusivity, Q is the heat source term, T is the temperature, and t is the time; During the iterative heating phase, pseudo-materials with similar specific heat capacities gradually fuse due to the thermal conduction effect, while pixels with differences exceeding a custom threshold (such as the boundary between the target and the background) form energy barriers due to the temperature gradient. During the cooling stage, the ROI region is converged through thermal balance calculation, which smooths out noisy pixels and highlights target features.
[0013] In step S2, the temperature profile analysis is performed using MATLAB's improfile function. The thermal gradient and abnormal regions are extracted by combining binarization and boundary detection techniques, providing robust features for subsequent identification.
[0014] Step S3, Wavelet Enhancement A signal or image multi-resolution analysis method based on Haar wavelet basis functions is used to perform multi-scale wavelet decomposition and threshold denoising on the ROI feature map. After reconstruction, an enhanced image with clear edges is obtained, which is finally input into a classification model (such as CNN) to complete the recognition. In step S3, Haar wavelet decomposition is used, high-frequency coefficients are processed by dynamic thresholding, and weighted reconstruction is used to enhance the edges. By combining the multi-scale properties of wave optics with the time-frequency analysis capabilities of wavelet transform, image denoising and edge enhancement are achieved. Two-dimensional decomposition is performed using a Haar wavelet basis. First, perform multi-layer wavelet transform on the image to obtain approximation coefficients (cA, reflecting low-frequency smoothness information) and detail coefficients (horizontal cH, vertical cV, diagonal cD, reflecting high-frequency edges and noise). Then, a soft thresholding function is applied to the detail coefficients. λ (w) ｛j,k｝ ) = sign(w ｛j,k｝ )·max(0,│w ｛j,k｝ The process involves processing │−λ to suppress noise while preserving edge energy; Among them, w ｛j,k｝ Here, λ represents the wavelet detail coefficients, and λ is a preset threshold. Finally, the image is reconstructed and enhanced using inverse wavelet transform.
[0015] In step S3, multiple decomposition-reconstruction strategies are supported, including wavedec2-waverec2 and dwt2-idwt2. The optimal wavelet method can be dynamically selected based on the image noise characteristics to extract 12 noise statistical features (mean, variance, slope, peak value, etc.) for subsequent classification tasks.
[0016] Step S4, Dynamic Fusion A dynamic weight adjustment strategy is introduced, which assigns feature weights according to the local entropy value of the image that reflects the texture complexity. In high-entropy regions (such as dense textures), the weight of wavelet detail coefficients is increased, and in low-entropy regions (such as smooth backgrounds), the thermodynamic fusion intensity is enhanced, so as to achieve adaptive complementarity of physical features.
[0017] A grayscale image recognition enhancement system based on physical knowledge fusion is provided to implement the aforementioned method. It adopts a modular and integrated architecture, achieving grayscale image recognition enhancement through the synergistic effect of three major physical modules: polarized light imaging, thermodynamic feature extraction, and wave optics-wavelet transform. Each module constructs a mathematical model based on physical principles, forming an end-to-end processing link through the organic connection of data flow, including: The polarization imaging module is responsible for acquiring images of objects under different polarization states based on the transverse wave characteristics of light, calculating Stokes parameters, including total intensity I, linear polarization difference Q and diagonal polarization difference U, obtaining polarization degree DOP / polarization angle AOP image, and outputting a polarization enhancement map after subtracting the reflective area. The thermodynamic feature extraction module is responsible for defining grayscale image pixels as pseudo-substances with specific specific heat capacities, assigning similar grayscale value pixels with similar specific heat capacities, and simulating the heating-cooling process by solving the heat conduction equation, which is as follows: ∂T / ∂t=α²∇²T+Q Where α is the thermal diffusivity, Q is the heat source term, T is the temperature, and t is the time; During the iterative heating phase, pseudo-materials with similar specific heat capacities gradually fuse due to the thermal conduction effect, while pixels with differences exceeding a custom threshold (such as the boundary between the target and the background) form energy barriers due to the temperature gradient. During the cooling stage, the ROI region is converged through thermal balance calculation, which smooths out noisy pixels and highlights target features. The Wave Optics and Wavelet Transform module is responsible for using Haar wavelet decomposition, combining the multi-scale characteristics of wave optics with the time-frequency analysis capabilities of wavelet transform, to achieve image denoising and edge enhancement. Two-dimensional decomposition is performed using the Haar wavelet basis: First, perform multi-layer wavelet transform on the image to obtain approximation coefficients (cA, reflecting low-frequency smoothness information) and detail coefficients (horizontal cH, vertical cV, diagonal cD, reflecting high-frequency edges and noise). Then, a soft thresholding function is applied to the detail coefficients. λ (w) ｛j,k｝ ) = sign(w ｛j,k｝ )·max(0,│w ｛j,k｝ The process involves processing │−λ to suppress noise while preserving edge energy; Among them, w ｛j,k｝ Here, λ represents the wavelet detail coefficients, and λ is a preset threshold. Finally, the image is reconstructed and enhanced using inverse wavelet transform; The dynamic fusion module is responsible for introducing a dynamic weight adjustment strategy. It assigns feature weights according to the local entropy value of the image, which reflects the complexity of the texture. In high-entropy regions (such as dense textures), the weight of wavelet detail coefficients is increased, and in low-entropy regions (such as smooth backgrounds), the thermodynamic fusion intensity is enhanced to achieve adaptive complementarity of physical features.
[0018] A grayscale image recognition enhancement device based on physical knowledge fusion includes a memory and a processor; the memory is used to store a computer program, and the processor is used to execute the computer program to implement the above-described method steps.
[0019] A readable storage medium storing a computer program that, when executed by a processor, implements the above-described method steps.
[0020] The beneficial effects of this invention are: the grayscale image recognition enhancement method and system based on physical knowledge fusion breaks through the generalization bottleneck of traditional pure data-driven methods, avoids the decision ambiguity of black box models, reduces computational complexity, and improves feature extraction accuracy, providing a high-performance solution for fields such as industrial inspection and medical imaging. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Appendix Figure 1 This is a schematic diagram of the grayscale image recognition enhancement method based on physical knowledge fusion according to the present invention. Detailed Implementation
[0023] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions in the embodiments of this invention will be clearly and completely described below in conjunction with the embodiments of this invention. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.
[0024] This grayscale image recognition enhancement method based on physical knowledge fusion includes the following steps: Step S1: Polarization preprocessing Based on the transverse wave characteristics of light, images of objects under different polarization states are acquired, and Stokes parameters, including total intensity I, linear polarization difference Q, and diagonal polarization difference U, are calculated to obtain a polarization degree DOP / polarization angle AOP image to subtract reflective areas. In step S1, the original grayscale image is input, polarization preprocessing is performed on the original grayscale image, the degree of polarization DOP / polarization angle AOP is calculated through the Stokes parameter, and the polarization enhancement image is output after subtracting the reflective area. Among them, the degree of polarization (DOP) is used to quantify the polarization characteristics of the target, while the angle of polarization (AOP) reflects the distribution of polarization direction.
[0025] For highly reflective surfaces such as metals and glass, a pixel-level polarization component subtraction technique is employed: by identifying high-value areas of polarization degree (DOP) (usually corresponding to reflective areas), the polarization component intensity of these areas is subtracted from the original grayscale image, eliminating overexposure information while preserving texture details. In terms of hardware implementation, a single-layer silicon metasurface structure (etched with periodic triangular pores, α=924nm, R=265nm, H=315nm) is used to achieve directional edge detection of polarized light. With a numerical aperture of 0.35, it can effectively separate S-waves and P-waves to enhance edge contrast.
[0026] Step S2: Thermodynamic ROI Extraction The polarization enhancement map is mapped to the pseudo-matter temperature field, and the noise-suppressed ROI feature map is output through iterative optimization of the heat conduction equation. In step S2, a "pseudo-matter-energy diffusion" framework is constructed based on the heat conduction physics model to achieve noise suppression and feature fusion; A grayscale image pixel is defined as a pseudo-substance with a specific specific heat capacity, and pixels with similar grayscale values are assigned similar specific heat capacities. The heating-cooling process is simulated by solving the heat conduction equation, which is as follows: ∂T / ∂t=α²∇²T+Q Where α is the thermal diffusivity, Q is the heat source term, T is the temperature, and t is the time; During the iterative heating phase, pseudo-materials with similar specific heat capacities gradually fuse due to the thermal conduction effect, while pixels with differences exceeding a custom threshold (such as the boundary between the target and the background) form energy barriers due to the temperature gradient. During the cooling stage, the ROI region is converged through thermal balance calculation, which smooths out noisy pixels and highlights target features.
[0027] In step S2, the temperature profile analysis is performed using MATLAB's improfile function. The thermal gradient and abnormal regions are extracted by combining binarization and boundary detection techniques, providing robust features for subsequent identification.
[0028] Step S3, Wavelet Enhancement A signal or image multi-resolution analysis method based on Haar wavelet basis functions is used to perform multi-scale wavelet decomposition and threshold denoising on the ROI feature map. After reconstruction, an enhanced image with clear edges is obtained, which is finally input into a classification model (such as CNN) to complete the recognition. In step S3, Haar wavelet decomposition is used, high-frequency coefficients are processed by dynamic thresholding, and weighted reconstruction is used to enhance the edges. By combining the multi-scale properties of wave optics with the time-frequency analysis capabilities of wavelet transform, image denoising and edge enhancement are achieved. Two-dimensional decomposition is performed using a Haar wavelet basis. First, perform multi-layer wavelet transform on the image to obtain approximation coefficients (cA, reflecting low-frequency smoothness information) and detail coefficients (horizontal cH, vertical cV, diagonal cD, reflecting high-frequency edges and noise). Then, a soft thresholding function is applied to the detail coefficients. λ (w) ｛j,k｝ ) = sign(w ｛j,k｝ )·max(0,│w ｛j,k｝ The process involves processing │−λ to suppress noise while preserving edge energy; Among them, w ｛j,k｝ Here, λ represents the wavelet detail coefficients, and λ is a preset threshold. Finally, the image is reconstructed and enhanced using inverse wavelet transform.
[0029] In step S3, multiple decomposition-reconstruction strategies are supported, including wavedec2-waverec2 and dwt2-idwt2. The optimal wavelet method can be dynamically selected based on the image noise characteristics to extract 12 noise statistical features (mean, variance, slope, peak value, etc.) for subsequent classification tasks.
[0030] Step S4, Dynamic Fusion A dynamic weight adjustment strategy is introduced, which assigns feature weights according to the local entropy value of the image that reflects the texture complexity. In high-entropy regions (such as dense textures), the weight of wavelet detail coefficients is increased, and in low-entropy regions (such as smooth backgrounds), the thermodynamic fusion intensity is enhanced, so as to achieve adaptive complementarity of physical features.
[0031] This grayscale image recognition enhancement system, based on physics knowledge fusion, implements the aforementioned method. It employs a modular and integrated architecture, leveraging the synergistic effects of three major physical modules—polarized light imaging, thermodynamic feature extraction, and wave optics-wavelet transform—to achieve grayscale image recognition enhancement. Each module constructs a mathematical model based on physical principles, forming an end-to-end processing link through the organic connection of data flows, including: The polarization imaging module is responsible for acquiring images of objects under different polarization states based on the transverse wave characteristics of light, calculating Stokes parameters, including total intensity I, linear polarization difference Q and diagonal polarization difference U, obtaining polarization degree DOP / polarization angle AOP image, and outputting a polarization enhancement map after subtracting the reflective area. The thermodynamic feature extraction module is responsible for defining grayscale image pixels as pseudo-substances with specific specific heat capacities, assigning similar grayscale value pixels with similar specific heat capacities, and simulating the heating-cooling process by solving the heat conduction equation, which is as follows: ∂T / ∂t=α²∇²T+Q Where α is the thermal diffusivity, Q is the heat source term, T is the temperature, and t is the time; During the iterative heating phase, pseudo-materials with similar specific heat capacities gradually fuse due to the thermal conduction effect, while pixels with differences exceeding a custom threshold (such as the boundary between the target and the background) form energy barriers due to the temperature gradient. During the cooling stage, the ROI region is converged through thermal balance calculation, which smooths out noisy pixels and highlights target features. The Wave Optics and Wavelet Transform module is responsible for using Haar wavelet decomposition, combining the multi-scale characteristics of wave optics with the time-frequency analysis capabilities of wavelet transform, to achieve image denoising and edge enhancement. Two-dimensional decomposition is performed using the Haar wavelet basis: First, perform multi-layer wavelet transform on the image to obtain approximation coefficients (cA, reflecting low-frequency smoothness information) and detail coefficients (horizontal cH, vertical cV, diagonal cD, reflecting high-frequency edges and noise). Then, a soft thresholding function is applied to the detail coefficients. λ (w) ｛j,k｝ ) = sign(w ｛j,k｝ )·max(0,│w ｛j,k｝ │−λ) processing suppresses noise while preserving edge energy; ȵ Among them, w ｛j,k｝ Here, λ represents the wavelet detail coefficients, and λ is a preset threshold. Finally, the image is reconstructed and enhanced using inverse wavelet transform; The dynamic fusion module is responsible for introducing a dynamic weight adjustment strategy. It assigns feature weights according to the local entropy value of the image, which reflects the complexity of the texture. In high-entropy regions (such as dense textures), the weight of wavelet detail coefficients is increased, and in low-entropy regions (such as smooth backgrounds), the thermodynamic fusion intensity is enhanced to achieve adaptive complementarity of physical features.
[0032] The grayscale image recognition enhancement device based on physical knowledge fusion includes a memory and a processor; the memory is used to store a computer program, and the processor is used to execute the computer program to implement the above-described method steps.
[0033] The readable storage medium stores a computer program that, when executed by a processor, implements the above-described method steps.
[0034] Compared with existing technologies, this grayscale image recognition enhancement method based on physical knowledge fusion has the following characteristics: 1) Deep coupling of physical models: The mathematical properties of Stokes polarization theory, heat conduction equation and wavelet transform are organically integrated, breaking through the generalization bottleneck of traditional pure data-driven methods.
[0035] 2) End-to-end interpretability: The outputs of each module have clear physical meanings (such as DOP representing polarization characteristics and temperature gradient reflecting regional differences), avoiding the decision ambiguity of black box models.
[0036] 3) Lightweight implementation: The learnable parameter fusion network (X1) and the task network (X2) are jointly trained, and the polarization parameter analysis and feature extraction are integrated into the end-to-end framework, which reduces the computational complexity.
[0037] The above technical solution improves the feature extraction accuracy by 15%-20% compared with traditional methods in the task of recognizing grayscale images with strong noise and low contrast, providing a high-performance solution for fields such as industrial inspection and medical imaging.
[0038] This grayscale image recognition enhancement system based on physical knowledge fusion achieves performance breakthroughs in each core component through collaborative optimization of multiple physical modules, as detailed below: 1) Polarized light imaging The polarization sensitivity has been improved by 30%, effectively removing strong reflective interference from overexposed areas in industrial settings and in-vehicle imaging in intelligent transportation, significantly improving image clarity. The F1-score for defect detection based on polarization degree calculation has been improved by 0.28, and the average grayscale difference in the material polarization degree map can achieve accurate identification of different materials within a distance of 100 meters underwater.
[0039] 2) Thermodynamic Feature Extraction Thermodynamic ROI extraction accuracy reaches 99%, automated processing is quick, and the average difference from manual annotation by physicians is less than 1%, making it suitable for clinical diagnostic scenarios such as liver fibrosis assessment. The thermodynamic convergence model improves the signal-to-noise ratio (SNR) of PET images by 15dB, providing stable feature support for high-noise medical image analysis.
[0040] 3) Wave optics and wavelet transform It improves noise resistance by 25% compared to the Sobel operator, separates noise from signal through multi-scale analysis, achieves an edge preservation of 0.92 (SSIM index), and improves PSNR by 1.8dB compared to traditional hard thresholding denoising. It enhances image details while eliminating block artifacts, adjusts image size without affecting detection accuracy, and exhibits excellent adaptability to different shooting devices and lighting conditions.
[0041] The embodiments described above are merely one specific implementation of the present invention. Ordinary changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included within the protection scope of the present invention.
Claims
1. A grayscale image recognition enhancement method based on physical knowledge fusion, characterized in that: Includes the following steps: Step S1: Polarization preprocessing Based on the transverse wave characteristics of light, images of objects under different polarization states are acquired, and Stokes parameters, including total intensity I, linear polarization difference Q and diagonal polarization difference U, are calculated to obtain the degree of polarization DOP / polarization angle AOP image to subtract reflective areas. Step S2: Thermodynamic ROI Extraction The polarization enhancement map is mapped to the pseudo-matter temperature field, and the noise-suppressed ROI feature map is output through iterative optimization of the heat conduction equation. Step S3, Wavelet Enhancement A signal or image multi-resolution analysis method based on Haar wavelet basis functions is used to perform multi-scale wavelet decomposition and threshold denoising on the ROI feature map. After reconstruction, an enhanced image with clear edges is obtained, which is finally input into a classification model to complete the recognition. Step S4, Dynamic Fusion A dynamic weight adjustment strategy is introduced, which assigns feature weights according to the local entropy value of the image that reflects the texture complexity. In high-entropy regions, the weight of wavelet detail coefficients is increased, and in low-entropy regions, the thermodynamic fusion intensity is enhanced to achieve adaptive complementarity of physical features.
2. The grayscale image recognition enhancement method based on physical knowledge fusion according to claim 1, characterized in that: In step S1, the original grayscale image is input, polarization preprocessing is performed on the original grayscale image, the degree of polarization DOP / polarization angle AOP is calculated through the Stokes parameter, and the polarization enhancement image is output after subtracting the reflective area. Among them, the degree of polarization (DOP) is used to quantify the polarization characteristics of the target, while the angle of polarization (AOP) reflects the distribution of polarization direction.
3. The grayscale image recognition enhancement method based on physical knowledge fusion according to claim 1, characterized in that: In step S2, a "pseudo-matter-energy diffusion" framework is constructed based on the heat conduction physics model to achieve noise suppression and feature fusion; A grayscale image pixel is defined as a pseudo-substance with a specific specific heat capacity, and pixels with similar grayscale values are assigned similar specific heat capacities. The heating-cooling process is simulated by solving the heat conduction equation, which is as follows: ∂T / ∂t=α²∇²T+Q Where α is the thermal diffusivity, Q is the heat source term, T is the temperature, and t is the time; During the iterative heating phase, pseudo-materials with similar specific heat capacities gradually fuse due to the thermal conduction effect, while pixels with differences exceeding a custom threshold form energy barriers due to temperature gradients. During the cooling stage, the ROI region is converged through thermal balance calculation, which smooths out noisy pixels and highlights target features.
4. The grayscale image recognition enhancement method based on physical knowledge fusion according to claim 3, characterized in that: In step S2, the MATLAB improfile function is used to perform temperature profile analysis. Combined with binarization imbinarize and boundary detection bwboundaries technology, thermal gradients and abnormal regions are extracted to provide robust features for subsequent identification.
5. The grayscale image recognition enhancement method based on physical knowledge fusion according to claim 1, characterized in that: In step S3, Haar wavelet decomposition is used, high-frequency coefficients are processed by dynamic thresholding, and weighted reconstruction is used to enhance the edges. By combining the multi-scale properties of wave optics with the time-frequency analysis capabilities of wavelet transform, image denoising and edge enhancement are achieved. Two-dimensional decomposition is performed using a Haar wavelet basis. First, perform multi-level wavelet transform on the image to obtain approximation coefficients and detail coefficients; Then, a soft thresholding function is applied to the detail coefficients. λ (w) ｛j,k｝ ) = sign(w ｛j,k｝ )·max(0,│w ｛j,k｝ The process involves processing │−λ to suppress noise while preserving edge energy; Among them, w ｛j,k｝ Here, λ represents the wavelet detail coefficients, and λ is a preset threshold. Finally, the image is reconstructed and enhanced using inverse wavelet transform.
6. The grayscale image recognition enhancement method based on physical knowledge fusion according to claim 5, characterized in that: In step S3, multiple decomposition-reconstruction strategies are supported, including wavedec2-waverec2 and dwt2-idwt2. The optimal wavelet method can be dynamically selected based on the image noise characteristics to extract noise statistical features for subsequent classification tasks.
7. A grayscale image recognition enhancement system based on physical knowledge fusion, characterized in that: To implement the method according to any one of claims 1 to 6, comprising: The polarization imaging module is responsible for acquiring images of objects under different polarization states based on the transverse wave characteristics of light, calculating Stokes parameters, including total intensity I, linear polarization difference Q and diagonal polarization difference U, obtaining polarization degree DOP / polarization angle AOP image, and outputting a polarization enhancement map after subtracting the reflective area. The thermodynamic feature extraction module is responsible for defining grayscale image pixels as pseudo-substances with specific specific heat capacities, assigning similar grayscale value pixels with similar specific heat capacities, and simulating the heating-cooling process by solving the heat conduction equation, which is as follows: ∂T / ∂t=α²∇²T+Q Where α is the thermal diffusivity, Q is the heat source term, T is the temperature, and t is the time; During the iterative heating phase, pseudo-materials with similar specific heat capacities gradually fuse due to the thermal conduction effect, while pixels with differences exceeding a custom threshold form energy barriers due to temperature gradients. During the cooling stage, the ROI region is converged through thermal balance calculation, which smooths out noisy pixels and highlights target features. The Wave Optics and Wavelet Transform module is responsible for using Haar wavelet decomposition, combining the multi-scale characteristics of wave optics with the time-frequency analysis capabilities of wavelet transform, to achieve image denoising and edge enhancement. Two-dimensional decomposition is performed using the Haar wavelet basis: First, perform multi-level wavelet transform on the image to obtain approximation coefficients and detail coefficients; Then, a soft thresholding function is applied to the detail coefficients. λ (w) ｛j,k｝ ) = sign(w ｛j,k｝ )·max(0,│w ｛j,k｝ The process involves processing │−λ to suppress noise while preserving edge energy; Among them, w ｛j,k｝ Here, λ represents the wavelet detail coefficients, and λ is a preset threshold. Finally, the image is reconstructed and enhanced using inverse wavelet transform; The dynamic fusion module is responsible for introducing a dynamic weight adjustment strategy. It assigns feature weights according to the local entropy value of the image, which reflects the texture complexity. It increases the weight of wavelet detail coefficients in high-entropy regions and enhances the thermodynamic fusion intensity in low-entropy regions to achieve adaptive complementarity of physical features.
8. A grayscale image recognition enhancement device based on physical knowledge fusion, characterized in that: It includes a memory and a processor; the memory is used to store a computer program, and the processor is used to execute the computer program to implement the method according to any one of claims 1 to 6.
9. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 6.