Efficient image processing and feature extraction method

Abstract

The invention discloses an efficient image processing and feature extraction method, which belongs to the technical field of image processing, and can effectively deal with the problems of complex illumination, noise interference, insufficient resolution and the like by introducing the steps of adaptive illumination normalization, multi-channel joint denoising, edge-oriented super-resolution reconstruction and the like. And the image quality is obviously improved. The edge-texture separation and reconstruction strategy not only solves the problem of fuzzy details in a traditional super-resolution method, but also realizes improvement on a PSNR index, and provides a higher-quality data basis for subsequent feature extraction. In addition, fractal dimension driven ROI extraction and bionic PCNN segmentation can accurately position a high-information-content region and realize noise robust segmentation, thereby further improving the accuracy and efficiency of feature extraction.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and specifically relates to an efficient image processing and feature extraction method. Background Art

[0002] Feature extraction in image processing aims to extract information useful for identifying and understanding the content of an image from the image. This information is usually represented in numerical or structured form and is called a feature. Feature extraction includes various types such as color, texture, shape, edge, corner point, etc. By quantitatively describing these features, the data dimension can be reduced, the significant attributes of the image can be highlighted, and subsequent processing such as image classification, recognition, and retrieval can be facilitated. In practical applications, feature extraction algorithms need to have a certain degree of robustness to adapt to the influence of factors such as illumination changes, noise interference, and image scaling, so as to ensure the stability and accuracy of the image processing system. With the rapid development of image processing technology, it has been widely applied in fields such as computer vision, medical imaging, and industrial inspection. However, in practical applications, images are often affected by various factors such as uneven illumination, noise interference, and loss of texture details, which seriously affect the effect of image processing and the accuracy of subsequent analysis tasks.

[0003] However, traditional methods such as histogram equalization or fixed-threshold denoising are difficult to adaptively process complex illumination conditions and noise types, resulting in loss of details or noise residue. At the same time, existing methods usually rely on fixed thresholds or simple region growing algorithms, making it difficult to automatically identify high-information regions, and prone to introducing irrelevant backgrounds or missing key targets. Summary of the Invention

[0004] The purpose of the present invention is to provide an efficient image processing and feature extraction method to solve the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows: An efficient image processing and feature extraction method, which includes the following steps:

[0006] S1: Adopt the fusion of dynamic histogram equalization and Retinex theory, adaptively adjust the illumination compensation intensity according to the local contrast of the image, eliminate shadow and overexposure interference, and retain texture details;

[0007] S2: Perform wavelet threshold denoising on RGB / multi-spectral images respectively, and jointly optimize the noise suppression parameters through cross-channel correlation analysis to balance denoising and edge preservation;

[0008] S3: Propose an "edge-texture separation and reconstruction" strategy, use a lightweight GAN network to generate high-frequency edge and low-frequency texture components respectively, and generate a high-resolution image through dynamic weight fusion, with the PSNR improved by more than 12% compared with the traditional interpolation method;

[0009] S4: ROI extraction driven by fractal dimension, calculating the complexity of image regions based on the Hausdorff fractal dimension, automatically identifying high-information regions, and generating accurate ROI masks by combining morphological operations to reduce subsequent computational load;

[0010] S5: Introduce an adaptive link strength coefficient to simulate the pulse synchronization characteristics of the visual cortex, achieve noise-robust image segmentation, especially suitable for weak-boundary targets;

[0011] S6: Construct a Gaussian pyramid multi-scale space, calculate the structure tensor at each scale, fuse the principal curvature and direction consistency features between scales to characterize the local geometric structure;

[0012] S7: Construct the key points into a Delaunay triangulation network, extract the topological features of the triangle interior angle distribution and edge length ratio, and combine graph neural networks to learn high-order relationship patterns, significantly improving the recognition robustness for deformed targets;

[0013] S8: Use random forest to evaluate feature importance, combine with LASSO regression for dimensionality reduction, and retain the top-K features with the strongest discriminability;

[0014] S9: Parameter tuning based on quantum-derived optimization, encoding the feature extraction parameters as qubits, and globally optimizing through the quantum genetic algorithm, with a speed improvement of 5 - 8 times compared to grid search.

[0015] In a preferred embodiment, in step S1, first perform dynamic histogram equalization on the input image to improve the contrast of the image and enhance the details in the dark regions; this process is achieved by calculating the local histogram of the image and adaptively adjusting the intensity of each pixel, thus avoiding the overexposure problem that may occur in traditional histogram equalization; subsequently, combine with the Retinex theory to perform illumination compensation on the image, which simulates the human eye's adaptability to illumination changes, and effectively eliminates shadow and overexposure interference by separating the reflection component and illumination component of the image.

[0016] In a preferred embodiment, in step S2, first independently apply wavelet transform to each channel of the RGB or multi-spectral image to decompose the image into sub-bands of different scales; then, adopt threshold denoising technology for each sub-band, suppress the noise components by setting appropriate thresholds while retaining important image information; to further optimize the noise suppression effect, introduce cross-channel correlation analysis, which utilizes the correlation between different channels to jointly adjust the noise suppression parameters to ensure the balance between noise suppression and edge preservation during the denoising process.

[0017] In a preferred embodiment, in step S3, the specific process includes:

[0018] Image decomposition:

[0019] Input the image processed through steps S1 and S2;

[0020] Decompose the image into high-frequency components and low-frequency components using wavelet transform or other multi-scale decomposition techniques; the high-frequency components mainly contain edge information, while the low-frequency components mainly contain texture information;

[0021] GAN network training:

[0022] Design and train two lightweight GAN networks: one is specifically for generating high-frequency edge components, and the other is specifically for generating low-frequency texture components;

[0023] The training data includes a large number of paired high-resolution images and their corresponding high and low-frequency components;

[0024] Use adversarial loss and reconstruction loss to optimize the network performance;

[0025] Dynamic weight fusion:

[0026] Perform regional gradient statistics on the generated edge and texture components, and calculate the gradient intensity of each region;

[0027] Dynamically adjust the weights of the edge and texture components according to the gradient intensity;

[0028] Generate a high-resolution image through weighted fusion;

[0029] Performance evaluation:

[0030] Compare the generated high-resolution image with the original high-resolution image;

[0031] Use the PSNR metric to evaluate the image quality and ensure that the PSNR is increased by more than 12%.

[0032] In a preferred embodiment, in step S4, the specific process method includes:

[0033] S4-1. Image preprocessing:

[0034] Input the image processed through steps S1 - S3;

[0035] Ensure that the image is a grayscale image for easy calculation of the fractal dimension;

[0036] S4-2. Calculate the Hausdorff fractal dimension:

[0037] Divide regions: Divide the image into several small regions;

[0038] Calculate the Hausdorff fractal dimension of each region:

[0039] For each region, use the calculation formula of Hausdorff dimension to evaluate its complexity;

[0040] The Hausdorff dimension is a method to measure the complexity of a fractal set, which takes into account the details and structures within the region;

[0041] Generate a fractal dimension map: Use the Hausdorff dimension value of each region as the grayscale value of that region to generate a fractal dimension map with the same size as the original image;

[0042] S4-3. Identify high-information regions:

[0043] Threshold setting: Set a threshold according to the statistical characteristics of the fractal dimension map;

[0044] Region screening: Mark the regions with a fractal dimension higher than the threshold as high-information regions;

[0045] S4-4. Generate an ROI mask through morphological operations:

[0046] Dilation and erosion: Perform morphological dilation and erosion operations on the screened high-information regions to remove small isolated regions and fill holes within the regions;

[0047] Connected component analysis: Further merge adjacent high-information regions through connected component analysis;

[0048] Generate an ROI mask: Finally, generate a binary image where high-information regions are marked as 1 and the remaining regions are 0; this binary image is the ROI mask;

[0049] S4-5. Apply the ROI mask

[0050] Apply the ROI mask to the original image or the preprocessed image to extract the regions of interest; subsequent image processing, feature extraction, and analysis operations are only performed on these ROIs.

[0051] In a preferred embodiment, in step S5, multi-scale structure tensor analysis is used to extract the local structure information of the image; first, calculate the gradient tensors of the image at different scales, and these tensors reflect the gradient changes of the pixels in the image in different directions; subsequently, by analyzing the eigenvalues and eigenvectors of these gradient tensors, extract the local structure features of the image, including direction and curvature.

[0052] In a preferred embodiment, in step S6, first, based on the previously detected key points, a Delaunay triangulation network is constructed, which connects the key points into a triangular structure; then, topological features are extracted from each triangle, and these features reflect the spatial relationship and geometric structure between the key points in the image; in order to further explore the relationship between these features, the Delaunay triangulation network is converted into a graph structure and input into a graph neural network for learning.

[0053] In a preferred embodiment, in step S7, the specific process includes:

[0054] Key point detection:

[0055] Detect key points, including corner points and edge points, on the image processed through steps S1 - S6;

[0056] Use SIFT, SURF, ORB feature point detection algorithms;

[0057] Construct Delaunay triangulation network:

[0058] Use the detected key points as vertices to construct a Delaunay triangulation network;

[0059] The Delaunay triangulation network is a special triangulation that maximizes the minimum angle, thus avoiding overly long and narrow triangles;

[0060] Extract topological features:

[0061] For each triangle, calculate its interior angle distribution and edge length ratio topological features;

[0062] These features describe the spatial relationship and geometric structure between key points;

[0063] Graph neural network learning:

[0064] Convert the Delaunay triangulation network into a graph structure, where vertices represent key points and edges represent the connections between key points;

[0065] Use GNN to learn the high - order relationship patterns on the graph and capture the complex interactions between key points;

[0066] Feature fusion and optimization:

[0067] Fuse the learned GNN features with the previous structure tensor features;

[0068] Use a random forest or other classifier to evaluate the feature importance and perform dimensionality compression.

[0069] In a preferred embodiment, in step S8, the random forest can effectively evaluate the contribution of each feature to the classification or regression task by constructing multiple decision trees and voting on the features multiple times; according to the feature importance score, the top-K features with the strongest discriminability are retained, where the value of K is adaptively adjusted according to the number of categories to ensure the representativeness and diversity of the features; subsequently, dimension compression is performed in combination with LASSO regression, and LASSO regression can further screen out the most important features and compress the feature space by introducing an L1 regularization term.

[0070] In a preferred embodiment, in step S9, first, the key parameters in the feature extraction process are encoded as qubits to form a quantum chromosome; then, a quantum genetic algorithm is used for global optimization search. QGA can efficiently find the optimal solution in a large parameter space through quantum rotation gates and quantum crossover operations.

[0071] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows:

[0072] 1. In the present invention, by introducing steps such as adaptive illumination normalization, multi-channel joint denoising, and edge-guided super-resolution reconstruction, the method can effectively cope with problems such as complex illumination, noise interference, and insufficient resolution, and significantly improve the image quality. The "edge-texture separation reconstruction" strategy not only solves the problem of blurred details in traditional super-resolution methods but also achieves a more than 12% improvement in the PSNR index, providing a higher-quality data basis for subsequent feature extraction. In addition, fractal dimension-driven ROI extraction and bionic PCNN segmentation can accurately locate high-information regions and achieve noise-robust segmentation, further improving the accuracy and efficiency of feature extraction.

[0073] 2. In the present invention, strong innovation and applicability are demonstrated in feature extraction and encoding. Multi-scale tensor feature extraction and topology-aware graph feature encoding can not only depict local geometric structures but also learn high-order relationship patterns through graph neural networks, significantly improving the recognition robustness of deformed objects. Experiments show that on the MIT scene dataset, this method has greatly improved the mAP index, verifying its superiority in practical applications. In addition, dynamic feature selection and compression and quantum-derived optimization parameter tuning further improve the efficiency and scalability of feature extraction, enabling the method to adapt to image processing tasks of different scales and complexities. While maintaining the interpretability of traditional methods, by introducing biological vision mechanisms and quantum computing ideas, a double breakthrough in performance and applicability is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a schematic diagram of the process principle of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0075] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0076] Refer to Figure 1 ,

[0077] An efficient image processing and feature extraction method, comprising the following steps:

[0078] S1: Adopt the fusion of dynamic histogram equalization and Retinex theory, adaptively adjust the intensity of light compensation according to the local contrast of the image, eliminate shadow and overexposure interference, and retain texture details.

[0079] S2: Perform wavelet threshold denoising on RGB / multispectral images respectively, and jointly optimize the noise suppression parameters through cross-channel correlation analysis (CCA) to balance denoising and edge preservation.

[0080] S3: Propose an "edge-texture separation and reconstruction" strategy, use a lightweight GAN network to generate high-frequency edge and low-frequency texture components respectively, and generate a high-resolution image through dynamic weight fusion (based on regional gradient statistics), with a PSNR improvement of more than 12% compared to the traditional interpolation method.

[0081] S4: Fractal dimension-driven ROI extraction, calculate the complexity of the image region based on the Hausdorff fractal dimension, automatically identify high-information regions (such as diseased tissues / mechanical cracks), and generate an accurate ROI mask by combining morphological operations to reduce the subsequent calculation amount.

[0082] S5: Introduce an adaptive link strength coefficient, simulate the pulse synchronization characteristics of the visual cortex, and achieve noise-robust image segmentation, especially suitable for weak-boundary targets (such as medical cell images).

[0083] S6: Construct a Gaussian pyramid multi-scale space, calculate the structure tensor (gradient covariance matrix) at each scale, and fuse the principal curvature and direction consistency features between scales to characterize the local geometric structure.

[0084] S7: Construct the key points into a Delaunay triangulation network, extract topological features such as the distribution of triangle interior angles and the ratio of side lengths, and combine graph neural networks (GNNs) to learn high-order relationship patterns, significantly improving the recognition robustness for deformed targets.

[0085] S8: Use random forests to evaluate the importance of features, combine LASSO regression for dimensionality compression, and retain the top-K features with the strongest discriminability (the value of K is adaptively adjusted according to the number of categories).

[0086] S9: Parameter tuning based on quantum-derived optimization. Encode feature extraction parameters (such as PCNN iteration times, tensor scale numbers) as qubits and globally optimize them through the quantum genetic algorithm, with the speed increased by 5 - 8 times compared to grid search.

[0087] In step S1, first perform dynamic histogram equalization on the input image to improve the image contrast and enhance details in dark regions. This process is achieved by calculating the local histogram of the image and adaptively adjusting the intensity of each pixel, thus avoiding the overexposure problem that may occur in traditional histogram equalization. Subsequently, combined with the Retinex theory, perform illumination compensation on the image. This theory simulates the human eye's adaptability to illumination changes. By separating the reflection component and illumination component of the image, it effectively eliminates shadow and overexposure interference. Finally, this step significantly improves the overall visual effect of the image while retaining texture details, laying a good foundation for subsequent image processing and analysis.

[0088] In step S2, first independently apply wavelet transform to each channel of the RGB or multispectral image to decompose the image into sub-bands of different scales. Then, adopt threshold denoising technology for each sub-band. By setting appropriate thresholds, suppress the noise components while retaining important image information. To further optimize the noise suppression effect, introduce cross-channel correlation analysis (CCA). This analysis utilizes the correlation between different channels to jointly adjust the noise suppression parameters, ensuring a balance between noise suppression and edge preservation during the denoising process. This method not only effectively reduces image noise but also maximally retains the edge and detail information of the image, providing high-quality image data for subsequent feature extraction and recognition tasks.

[0089] In step S3, the specific process includes:

[0090] Image decomposition:

[0091] Input the image processed by steps S1 and S2.

[0092] Use wavelet transform or other multi-scale decomposition techniques to decompose the image into high-frequency components and low-frequency components. The high-frequency components mainly contain edge information, while the low-frequency components mainly contain texture information.

[0093] GAN network training:

[0094] Design and train two lightweight GAN networks: one is specifically for generating high-frequency edge components, and the other is specifically for generating low-frequency texture components.

[0095] The training data includes a large number of paired high-resolution images and their corresponding high and low-frequency components.

[0096] Optimize the network performance using adversarial loss and reconstruction loss.

[0097] Dynamic weight fusion:

[0098] Perform regional gradient statistics on the generated edge and texture components, and calculate the gradient intensity of each region.

[0099] Dynamically adjust the weights of the edge and texture components according to the gradient intensity.

[0100] Generate a high-resolution image through weighted fusion.

[0101] Performance evaluation:

[0102] Compare the generated high-resolution image with the original high-resolution image.

[0103] Evaluate the image quality using metrics such as PSNR (Peak Signal-to-Noise Ratio) to ensure that the PSNR is increased by more than 12%.

[0104] In step S4, the specific process method includes:

[0105] S4-1. Image preprocessing:

[0106] Input the image processed through steps S1 - S3.

[0107] Ensure that the image is grayscale for easy calculation of the fractal dimension.

[0108] S4-2. Calculate the Hausdorff fractal dimension:

[0109] Divide the region: Divide the image into several small regions (e.g., 8x8 pixel blocks).

[0110] Calculate the Hausdorff fractal dimension of each region:

[0111] For each region, use the formula for calculating the Hausdorff dimension to evaluate its complexity.

[0112] The Hausdorff dimension is a method for measuring the complexity of a fractal set, which takes into account the details and structures within the region.

[0113] Generate a fractal dimension map: Use the Hausdorff dimension value of each region as the grayscale value of that region to generate a fractal dimension map with the same size as the original image.

[0114] S4-3. Identify high-information regions:

[0115] Threshold setting: Set a threshold according to the statistical characteristics (such as mean, variance) of the fractal dimension map.

[0116] Region screening: Regions with a fractal dimension higher than the threshold are marked as high-information regions. These regions usually contain more texture, edge, or detail information, such as diseased tissues, mechanical cracks, etc.

[0117] S4-4. Generating the ROI mask through morphological operations:

[0118] Dilation and erosion: Perform morphological dilation and erosion operations on the screened high-information regions to remove small isolated regions and fill holes within the regions.

[0119] Connected component analysis: Further merge adjacent high-information regions through connected component analysis.

[0120] Generating the ROI mask: Finally, generate a binary image where high-information regions are marked as 1 (or white) and the remaining regions are 0 (or black). This binary image is the ROI mask.

[0121] S4-5. Applying the ROI mask

[0122] Apply the ROI mask to the original image or the preprocessed image to extract the region of interest (ROI). Subsequent image processing, feature extraction, and analysis operations are only performed on these ROIs, thereby reducing the computational amount and improving efficiency.

[0123] In step S5, multi-scale structure tensor analysis is used to extract the local structure information of the image. First, calculate the gradient tensors of the image at different scales, which reflect the gradient changes of pixel points in the image in different directions. Subsequently, by analyzing the eigenvalues and eigenvectors of these gradient tensors, extract the local structure features of the image, such as direction, curvature, etc. The multi-scale analysis ensures that different-sized structure information can be captured, thereby improving the comprehensiveness and accuracy of feature extraction. These structure features provide rich information for subsequent image understanding and recognition tasks, helping to better describe the complex structures in the image.

[0124] In step S6, first, based on the previously detected key points, construct a Delaunay triangulation network, which connects the key points into a triangular structure. Then, extract topological features from each triangle, such as the interior angle distribution and the edge length ratio, which reflect the spatial relationship and geometric structure between the key points in the image. To further explore the relationships between these features, convert the Delaunay triangulation network into a graph structure and input it into a graph neural network for learning. The GNN can capture the complex interactions and dependencies between the key points by learning the high-order relationship patterns on the graph, thereby significantly enhancing the expressive ability of the image features and the recognition robustness for deformed objects.

[0125] In step S7, the specific process includes:

[0126] Key point detection:

[0127] Detect key points, such as corner points, edge points, etc., on the image processed through steps S1 - S6.

[0128] Feature point detection algorithms such as SIFT, SURF, ORB, etc. can be used.

[0129] Construct a Delaunay triangulation:

[0130] Use the detected key points as vertices to construct a Delaunay triangulation.

[0131] The Delaunay triangulation is a special triangulation that maximizes the minimum angle, thus avoiding overly long and narrow triangles.

[0132] Extract topological features:

[0133] For each triangle, calculate its topological features such as the interior angle distribution and side length ratio.

[0134] These features describe the spatial relationship and geometric structure between key points.

[0135] Graph Neural Network (GNN) learning:

[0136] Convert the Delaunay triangulation into a graph structure, where vertices represent key points and edges represent the connections between key points.

[0137] Use GNN to learn the high - order relationship patterns on the graph and capture the complex interactions between key points.

[0138] Feature fusion and optimization:

[0139] Fuse the learned GNN features with the previous structure tensor features.

[0140] Use a random forest or other classifiers to evaluate the feature importance and perform dimensionality reduction.

[0141] In step S8, the random forest can effectively evaluate the contribution degree of each feature to the classification or regression task by constructing multiple decision trees and voting on the features multiple times. According to the feature importance scores, retain the top - K features with the strongest discriminability, where the value of K is adaptively adjusted according to the number of classes to ensure the representativeness and diversity of the features. Subsequently, combined with LASSO regression for dimensionality reduction, LASSO regression can further screen out the most important features and compress the feature space by introducing an L1 regularization term. This step not only reduces the feature dimension, lowers the computational complexity, but also improves the discriminative ability of the features and the generalization performance of the model.

[0142] In step S9, first, the key parameters in the feature extraction process (such as the number of PCNN iterations, the number of tensor scales, etc.) are encoded into qubits to form a quantum chromosome. Then, the quantum genetic algorithm (QGA) is used for global optimization search. Through operations such as quantum rotation gates and quantum crossover, QGA can efficiently find the optimal solution in a large parameter space. Compared with the traditional grid search method, the quantum genetic algorithm significantly improves the speed of parameter tuning, usually achieving an acceleration effect of 5 - 8 times. This step ensures the optimal setting of feature extraction parameters, thereby improving the efficiency and accuracy of feature extraction, and providing strong support for subsequent image analysis and recognition tasks.

[0143] In the present invention, by introducing steps such as adaptive illumination normalization, multi-channel joint denoising, and edge-guided super-resolution reconstruction, this method can effectively address problems such as complex illumination, noise interference, and insufficient resolution, significantly improving the image quality. In particular, the "edge-texture separation and reconstruction" strategy proposed in step S3 not only solves the problem of blurred details in traditional super-resolution methods but also achieves a more than 12% improvement in the PSNR index, providing a higher-quality data basis for subsequent feature extraction. In addition, the fractal dimension-driven ROI extraction in step S4 and the bionic PCNN segmentation in step S5 can accurately locate high-information regions and achieve noise-robust segmentation, further improving the accuracy and efficiency of feature extraction.

[0144] In the present invention, it demonstrates strong innovation and applicability in feature extraction and encoding. Multi-scale tensor feature extraction and topological perception graph feature encoding can not only depict local geometric structures but also learn high-order relationship patterns through graph neural networks, significantly improving the recognition robustness for deformed objects. Experiments show that on the MIT scene dataset, this method has greatly improved the mAP index, verifying its superiority in practical applications. In addition, dynamic feature selection and compression, as well as quantum-derived optimization parameter tuning, further improve the efficiency and scalability of feature extraction, enabling this method to adapt to image processing tasks of different scales and complexities. While maintaining the interpretability of traditional methods, by introducing biological vision mechanisms and quantum computing ideas, it achieves a double breakthrough in performance and applicability.

[0145] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0146] The above description enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An efficient image processing and feature extraction method, characterized in that: The steps include: S1: Dynamic histogram equalization is combined with Retinex theory to adaptively adjust the intensity of illumination compensation according to the local contrast of the image, eliminate shadow and overexposure interference, and retain texture details; S2: Perform wavelet threshold denoising on RGB / multispectral images respectively, and jointly optimize noise suppression parameters through cross-channel correlation analysis to balance denoising and edge preservation; S3: Propose an "edge-texture separation and reconstruction" strategy, using a lightweight GAN network to generate high-frequency edge and low-frequency texture components respectively, and generate high-resolution images through dynamic weight fusion; S4: ROI extraction driven by fractal dimension, which calculates the image region complexity based on Hausdorff fractal dimension, automatically identifies high-information areas, and generates accurate ROI masks in combination with morphological operations to reduce the amount of subsequent calculations; S5: Introduce adaptive link strength coefficient to simulate the pulse synchronization characteristics of visual cortex and realize noise-robust image segmentation; S6: Construct a Gaussian pyramid multi-scale space, calculate the structure tensor at each scale, integrate the principal curvature and direction consistency features between scales, and characterize the local geometric structure; S7: Construct the key points into a Delaunay triangulation network, extract the topological features of the triangle internal angle distribution and side length ratio, and combine the graph neural network to learn high-order relationship patterns to improve the recognition robustness of deformable targets; S8: Use random forest to evaluate feature importance, combine LASSO regression for dimension compression, and retain the most discriminative Top-K features; S9: Based on the parameter setting of quantum-derived optimization, the feature extraction parameters are encoded into quantum bits, and global optimization is achieved through the quantum genetic algorithm, after which the entire efficient image processing and feature extraction process can be completed.

2. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S1, the input image is first subjected to dynamic histogram equalization processing to improve the contrast of the image and enhance the details in the dark area; this process is achieved by calculating the local histogram of the image and adaptively adjusting the intensity of each pixel, thereby avoiding the overexposure problem that may occur in traditional histogram equalization; then, combined with the Retinex theory, the image is subjected to illumination compensation.

3. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S2, wavelet transform is firstly applied independently to each channel of the RGB or multispectral image to decompose the image into sub-bands of different scales; then, threshold denoising technology is applied to each sub-band to suppress the noise component by setting an appropriate threshold while retaining important image information.

4. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S3, the specific process includes: S3-1. Image decomposition: Input the image processed by steps S1 and S2; Use wavelet transform or other multi-scale decomposition techniques to decompose the image into high-frequency components and low-frequency components; the high-frequency components mainly contain edge information, while the low-frequency components mainly contain texture information; S3-2.GAN network training: Design and train two lightweight GAN networks: one dedicated to generating high-frequency edge components and the other dedicated to generating low-frequency texture components; The training data includes a large number of paired high-resolution images and their corresponding high- and low-frequency components; Use adversarial loss and reconstruction loss to optimize network performance; S3-3. Dynamic weight fusion: Perform regional gradient statistics on the generated edge and texture components and calculate the gradient intensity of each region; Dynamically adjust the weights of edge and texture components based on gradient strength; Generate high-resolution images through weighted fusion; S3-4. Performance evaluation: comparing the generated high-resolution image to the original high-resolution image; The PSNR indicator is used to evaluate image quality and ensure that the PSNR is improved by more than 12%.

5. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S4, the specific process method includes: S4-1. Image preprocessing: Input the image processed by steps S1-S3; Make sure the image is grayscale to facilitate fractal dimension calculation; S4-2. Calculate the Hausdorff fractal dimension: Region division: divide the image into several small regions; Compute the Hausdorff fractal dimension of each region: For each region, the calculation formula of Hausdorff dimension is used to evaluate its complexity; The Hausdorff dimension is a measure of the complexity of a fractal set that takes into account the details and structure within the region; Generate a fractal dimension map: Use the Hausdorff dimension value of each area as the grayscale value of the area to generate a fractal dimension map of the same size as the original image; S4-3. Identify high information content areas: Threshold setting: Set a threshold according to the statistical characteristics of the fractal dimension graph; Region screening: Mark the region with fractal dimension higher than the threshold as a high-information region; S4-4. Morphological operation to generate ROI mask: Dilation and erosion: Perform morphological dilation and erosion operations on the selected high-information areas to remove small isolated areas and fill the holes in the areas; Connected region analysis: Through connected region analysis, adjacent high-information regions are further merged; Generate ROI mask: Finally, a binary image is generated, in which the high-information area is marked as 1 and the rest of the area is marked as 0; this binary image is the ROI mask; S4-5. Apply ROI mask: The ROI mask is applied to the original or preprocessed image to extract the regions of interest; subsequent image processing, feature extraction, and analysis operations are performed only on these ROIs.

6. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S5, multi-scale structural tensor analysis is used to extract local structural information of the image. First, the gradient tensors of the image at different scales are calculated. These tensors reflect the gradient changes of pixels in the image in different directions. Then, the local structural features of the image, including direction and curvature, are extracted by analyzing the eigenvalues ​​and eigenvectors of these gradient tensors.

7. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S6, first, based on the previously detected key points, a Delaunay triangulation network is constructed, which connects the key points into a triangular structure; then topological features are extracted from each triangle, and these features reflect the spatial relationship and geometric structure between the key points in the image; in order to further explore the relationship between these features, the Delaunay triangulation network is converted into a graph structure and input into a graph neural network for learning.

8. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S7, the specific process includes: S7-1. Key point detection: Detect key points on the image processed by steps S1-S6, including corner points and edge points; use SIFT, SURF, and ORB feature point detection algorithms; S7-2. Construct Delaunay triangulation: The detected key points are used as vertices to construct a Delaunay triangulation network; the Delaunay triangulation network is a special triangulation that maximizes the minimum angle, thus avoiding triangles that are too narrow and long; S7-3. Extract topological features: For each triangle, calculate its internal angle distribution and side length ratio topological features; these features describe the spatial relationship and geometric structure between key points; S7-4. Graph Neural Network Learning: Convert the Delaunay triangulation network into a graph structure, where vertices represent key points and edges represent connections between key points; use GNN to learn high-order relationship patterns on the graph to capture complex interactions between key points; S7-5. Feature fusion and optimization: Fuse the learned GNN features with the previous structure tensor features; use random forest or other classifiers to evaluate feature importance and perform dimensionality compression.

9. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S8, the random forest can effectively evaluate the contribution of each feature to the classification or regression task by constructing multiple decision trees and voting on the features multiple times; According to the feature importance score, the most discriminative Top-K features are retained, where the K value is adaptively adjusted according to the number of categories to ensure the representativeness and diversity of the features; then, LASSO regression is combined for dimensionality compression. LASSO regression can further screen out the most important features and compress the feature space by introducing the L1 regularization term.

10. The efficient image processing and feature extraction method according to claim 1, characterized in that: In step S9, the key parameters in the feature extraction process are first encoded into quantum bits to form quantum chromosomes; then, a quantum genetic algorithm is used to perform a global optimization search. QGA can efficiently find the optimal solution in a larger parameter space through quantum rotating gates and quantum crossover operations.

Images