Image intelligence processing system and method based on generative images

Abstract

The application belongs to the technical field of image processing, and is specifically an image intelligent processing system and method based on generative images, and specifically includes an image source adaptive analysis module, a generative image generation grading module, a generative image quality quantification module, a source generative image fusion reconstruction module and a fused image intelligent optimization module; the application lays a precise data foundation for subsequent processing through the image source adaptive analysis module, generates according to precision grading through the generative image generation grading module, filters high-quality generated images through multi-dimensional indexes through the generative image quality quantification module, performs feature-level fusion and pixel-level reconstruction through the source generative image fusion reconstruction module, and performs accurate optimization according to optimization coefficients through the fused image intelligent optimization module, and sets verification standards in a targeted manner, guarantees the scene adaptability and reliability of the final output image, significantly improves the processing efficiency and image quality, and adapts to the image processing needs in multiple fields.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an intelligent image processing system and method based on generative images. Background Technology

[0002] With the rapid iteration of generative artificial intelligence technology, generative image processing has been widely applied in many fields such as portrait beautification, industrial inspection and completion, medical image enhancement, and virtual scene construction, becoming a key technology for enhancing the application value of images and expanding their application scenarios.

[0003] For example, Chinese invention patent with publication number CN119379556A discloses "an image denoising method and device based on generative artificial intelligence technology". The technical solution of this invention is to construct an image denoising model based on conditional generative adversarial network and combine it with noise type recognition to achieve accurate denoising of different medical images, so as to solve the problems of poor adaptability and insufficient accuracy of traditional denoising methods, and to a certain extent verify the application value of generative AI in the field of image processing.

[0004] However, the above-mentioned invention focuses on the single scenario of medical image denoising, which is difficult to adapt to the processing needs of multiple types and scenarios of source images. Furthermore, it does not involve the hierarchical generation of generative images, the fusion and reconstruction of source and generated images, and lacks subsequent intelligent optimization and effect verification. It cannot meet the needs of comprehensive image processing in complex scenarios, cannot guarantee the overall quality of the processed images, and cannot take into account the intelligence, precision and efficiency of image processing.

[0005] To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an image intelligent processing system and method based on generative images to address the technical deficiencies mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an image intelligent processing system based on generative images, comprising an image source adaptation and parsing module, a generative image generation and grading module, a generative image quality quantification module, a source image fusion and reconstruction module, and a fused image intelligent optimization module;

[0008] The image source adaptation and parsing module performs multi-dimensional feature parsing on original source images of different formats, resolutions, and scenes, and completes the format normalization, feature extraction, and scene adaptability determination of the source images; the generative image generation and grading module receives the source image adaptation and parsing dataset and divides the generative images into different levels according to the generation accuracy.

[0009] The generative image quality quantization module constructs a multi-dimensional generative image quality quantization system based on a hierarchical generative image dataset, calculates generative image quality scores, and determines whether the generative images are qualified or unqualified based on the quality scores. The original image fusion and reconstruction module realizes feature-level fusion and pixel-level reconstruction of the original source image and the high-quality generative image dataset based on the source image adaptation parsing dataset and the high-quality generative image dataset. The fused image intelligent optimization module is used to calculate optimization coefficients and perform multi-dimensional intelligent optimization of the fused image based on the optimization coefficients.

[0010] Furthermore, the image source adaptation and parsing module receives the original source image from the external input, first performs format recognition and normalization processing on the original source image, and converts all source images of all formats into the system's preset uncompressed high-definition format. At the same time, it performs adaptive scaling according to the resolution of the source image, matching the system's processing resolution while ensuring that image features are not lost.

[0011] The low-level visual features and high-level semantic features of the normalized source image are extracted by a deep learning feature extraction network. The extracted source image features are then used to determine scene adaptability. Based on a pre-set scene feature library, the source image is matched to the corresponding application scene category, generating a source image adaptation parsing dataset.

[0012] Furthermore, if the image source adaptation parsing module receives fusion anomaly data transmitted by the original image fusion reconstruction module, it extracts the feature conflict information in the fusion anomaly data, performs secondary feature parsing in combination with the feature data of the original source image, corrects the feature extraction results, and regenerates the source image adaptation parsing dataset. Finally, the source image adaptation parsing dataset is transmitted to the generative image generation hierarchical module through the system's internal data interface.

[0013] Furthermore, the generative image generation and grading module receives the source image adaptation parsing dataset, parses the scene adaptation categories and feature importance rankings in the dataset, and selects a generative model that matches the scene of the source image from the generative image model library according to the system's preset generative model-scene adaptation mapping table.

[0014] Based on the feature importance ranking of the source image, the generation parameters of the generation model are adaptively adjusted. The high-importance features of the source image are used as the core generation constraints of the generative image, and the low-importance features are used as auxiliary generation constraints. At the same time, according to the system's preset generation accuracy level classification standard, the generative image is divided into three levels: primary generation level, intermediate generation level, and advanced generation level.

[0015] Based on the adjusted generation parameters and the selected generation accuracy level, the matching generation model is driven to generate generative images. During the generation process, the feature data of the generative images are extracted in real time and initially matched with the feature data of the source images. Generative images with obvious feature conflicts are eliminated, and generative images that pass the initial matching are retained. Generation level labels and generation feature datasets are added to generative images of each level to form a hierarchical generative image dataset.

[0016] Furthermore, if the generative image generation grading module receives substandard generative image data transmitted by the generative image quality quantization module, it analyzes the reasons for the substandard data, readjusts the generation parameters of the generation model or replaces the generation model with a suitable one, and improves the generation accuracy level to re-grade and generate the data. Finally, the graded generative image dataset is transmitted to the generative image quality quantization module.

[0017] Furthermore, the generative image quality quantification module receives a hierarchical generative image dataset, parses the generative images, generation level labels, and generation feature datasets in the hierarchical generative image dataset, and simultaneously retrieves the source image feature data from the source image adaptation parsing dataset to construct a generative image quality quantification index system. It sets a comprehensive scoring formula for generative image quality to calculate the Q value; compares the calculated Q value with the pass threshold of the corresponding generation level to determine whether the generative image is passable or failable.

[0018] The qualified generative images are screened, and quality score labels are added to form a high-quality generative image dataset. The high-quality generative image dataset is then transmitted to the original image fusion and reconstruction module. Meanwhile, the unqualified generative image data is labeled with the reason for unqualification and fed back to the generative image generation and grading module for regeneration.

[0019] Furthermore, the original image fusion and reconstruction module performs in-depth analysis of the low-level visual features and high-level semantic features of the source image and the high-quality generated image to determine the fusion region and non-fusion region of the source image and the generated image.

[0020] First, feature-level fusion is performed, which weights and fuses the fusion region features of the generated image with the fusion region features of the source image. The fusion weights are adjusted according to the scene adaptation category of the source image. After feature-level fusion is completed, pixel-level reconstruction is performed, and pixel-level interpolation, smoothing and sharpening are performed on the fused feature data to optimize the edge transition effect between the fused and non-fused regions.

[0021] After pixel-level reconstruction is completed, the original fused reconstructed image is generated. At the same time, the data in the fusion process is extracted. If fusion anomalies occur, the abnormal data is extracted and a fusion anomaly dataset is fed back to the image source adaptation and parsing module for secondary parsing. If there are no anomalies in the fusion process, the original fused reconstructed image and the corresponding fusion feature dataset are transmitted to the fused image intelligent optimization module.

[0022] Furthermore, the intelligent image optimization module constructs a multi-dimensional image optimization system that includes color optimization, contrast optimization, sharpness optimization, and edge transition optimization. It also sets a formula for intelligent image optimization coefficients to calculate the K value. Based on the magnitude of the K value, the optimization level is divided into three levels: light optimization, medium optimization, and heavy optimization. Different optimization levels correspond to different adjustment ranges of optimization parameters. Based on the determined optimization level and the scene adaptation category of the source image, the fused image is intelligently optimized in multiple dimensions.

[0023] Furthermore, the fusion image intelligent optimization module communicates with the processed image effect verification module. The processed image effect verification module constructs a processing image effect verification system to determine whether the processed image is qualified. Qualified processed images are used as the final output result of the system, while unqualified image data is fed back to the fusion image intelligent optimization module. The specific operation process is as follows:

[0024] The system receives the fused and optimized image and optimization parameter dataset transmitted by the intelligent image fusion optimization module, and simultaneously retrieves the source image adaptation parsing dataset. Based on the scene adaptation category of the source image, it determines the effect verification standard of the processed image. After calculating the P-value, it compares the calculated P-value with the corresponding qualified threshold. If the P-value is not less than the corresponding qualified threshold, the fused and optimized image is judged to be qualified; if the P-value is less than the corresponding qualified threshold, the fused and optimized image is judged to be unqualified.

[0025] The present invention also proposes an intelligent image processing method based on generative images, comprising the following steps:

[0026] Step 1: Source Image Parsing and Transmission:

[0027] The image source adaptation and parsing module is activated to perform format normalization, feature extraction, and scene adaptation determination on the input original source image, and generate an adaptation parsing dataset which is then transmitted to the generative image generation and hierarchical module; if abnormal fusion data is received, it is parsed and corrected a second time before being retransmitted.

[0028] Step 2: Generate images in stages:

[0029] The generative image generation grading module matches the generation model corresponding to the scene, adjusts the parameters, and generates images according to accuracy grading. After preliminary screening, a graded generation dataset is formed and transmitted to the generative image quality quantification module. If unqualified data is received, the parameters / model are adjusted, the image is regenerated, and then retransmitted.

[0030] Step 3: Quality Quantitative Judgment

[0031] The generative image quality quantification module constructs a four-dimensional index system, calculates the comprehensive quality score Q, sets a threshold according to the generation level to determine the passability, and the qualified images form a high-quality dataset that is transmitted to the original image fusion and reconstruction module, while the unqualified data is fed back to the generative image generation and grading module.

[0032] Step 4: Original Fusion and Reconstruction

[0033] The original image fusion and reconstruction module performs feature-level fusion and pixel-level reconstruction on the source image and the high-quality generated image, and transmits the fused image to the fused image intelligent optimization module; if a fusion anomaly occurs, it is fed back to the image source adaptation and parsing module for secondary parsing.

[0034] Step 5: Image fusion optimization:

[0035] The image intelligent optimization module calculates the optimization coefficient K, performs multi-dimensional optimization at different levels, generates an optimized image and transmits it to the image processing effect verification module; if unqualified data is received, the coefficient is recalculated, the level is increased, the optimization is repeated and the image is retransmitted.

[0036] Step Six: Output the Result Verification:

[0037] The image processing effect verification module calculates the scene adaptation matching degree P, judges the passability according to the threshold, and outputs the passable image as the final result. The unqualified data is fed back to the fusion image intelligent optimization module for re-optimization.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] 1. In this invention, the original source images of different formats, resolutions, and scenes are normalized, multi-dimensional features are extracted, and scene adaptation is determined to avoid subsequent adaptation problems. Furthermore, images are generated in grades according to generation accuracy to meet the differentiated needs of different application scenarios and improve the adaptability between generated images and source images. In addition, by constructing a quality quantification system, the qualification of generated images is determined based on accurate scoring, and high-quality generated images are selected for subsequent fusion to avoid low-quality generated images affecting the fusion effect.

[0040] 2. In this invention, feature-level fusion and pixel-level reconstruction of source images and high-quality generated images are used to avoid distortion caused by simple superposition or splicing. Furthermore, a multi-dimensional optimization system is constructed, which divides optimization levels based on optimization coefficients, performs precise optimization in combination with scene adaptation categories, balances visual effects and image realism, and sets targeted verification standards based on the scene of the source image to calculate the scene adaptation matching degree to determine the qualification of the processed image, thus ensuring the reliability and scene adaptability of the final output image. Attached Figure Description

[0041] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0042] Figure 1 This is a system block diagram of Embodiment 1 of the present invention;

[0043] Figure 2 This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1: As Figure 1 As shown, the image intelligent processing system based on generative images proposed in this invention includes an image source adaptation and parsing module, a generative image generation and grading module, a generative image quality quantification module, a source image fusion and reconstruction module, a fused image intelligent optimization module, and a processed image effect verification module.

[0046] The image source adaptation and parsing module is the initial data processing unit of the system. It realizes multi-dimensional feature parsing of original source images of different formats, resolutions, and scenes, and completes the format normalization, feature extraction, and scene adaptability determination of the source images. It provides accurate source image feature data and scene adaptation basis for the hierarchical generation of subsequent generative images. It realizes the normalization processing and refined feature parsing of source images of multiple formats, resolutions, and scenes, avoiding the problem of poor adaptability between subsequent generative images and source images caused by inaccurate feature extraction of source images. At the same time, it can receive fusion abnormal data from the source image fusion and reconstruction module and perform secondary parsing to optimize the feature extraction results and improve the accuracy of source image feature data, providing reliable basic data for the hierarchical generation of generative images.

[0047] Specifically, the image source adaptation and parsing module first receives the original source image from the external input. The image formats include common and professional image formats such as JPG, PNG, TIFF, and RAW, and the resolutions cover low-resolution images from 320 to 240 to ultra-high-resolution images from 10K to 8K. The application scenarios include portrait photography, landscape photography, industrial inspection, medical imaging, virtual simulation, etc.

[0048] After receiving the image, the original source image is first processed for format recognition and normalization. All source images of different formats are converted into the system's preset uncompressed high-definition format. At the same time, adaptive scaling is performed according to the resolution of the source image to match the system's processing resolution while ensuring that image features are not lost.

[0049] Subsequently, a deep learning feature extraction network is used to extract low-level visual features and high-level semantic features from the normalized source image. The low-level visual features include pixel grayscale values, color channel distribution, edge contour features, and texture features, while the high-level semantic features include the main image target, scene environment information, and spatial relationships between targets.

[0050] Next, scene adaptability is determined for the extracted source image features. Based on the preset scene feature library, the source image is matched to the corresponding application scene category, and a source image adaptation parsing dataset containing source image format information, resolution information, low-level visual feature data, high-level semantic feature data, scene adaptation category, and feature importance ranking is generated.

[0051] Furthermore, if the image source adaptation parsing module receives fusion anomaly data transmitted by the original image fusion reconstruction module, it extracts the feature conflict information in the fusion anomaly data, performs secondary feature parsing in combination with the feature data of the original source image, corrects the feature extraction results, and regenerates the source image adaptation parsing dataset. Finally, the source image adaptation parsing dataset is transmitted to the generative image generation hierarchical module through the system's internal data interface.

[0052] The generative image generation and grading module is the core generation unit of generative images. Based on the source image adaptation and parsing dataset transmitted by the image source adaptation and parsing module, and combined with the preset generative image model library, it realizes the graded generation and preliminary screening of generative images. It selects the appropriate generation model according to the scene adaptation category and feature importance of the source image, and divides the generative images into different levels according to the generation accuracy, providing a grading basis for subsequent quality quantification. It can also receive unqualified generative image data from the generative image quality quantification module and re-grade and generate them to meet the differentiated needs of different application scenarios and improve the adaptability of generative images to source images.

[0053] Specifically, the generative image generation and hierarchical module first receives the source image adaptation parsing dataset transmitted by the image source adaptation parsing module, parses the scene adaptation categories and feature importance ranking in the dataset, and selects the generative model that matches the scene of the source image from the generative image model library according to the system's preset generative model-scene adaptation mapping table. The generative image model library includes diffusion models, GAN models, VAE models, etc. optimized for different scenes, and each model has completed adaptation training with the corresponding scene source image.

[0054] Subsequently, based on the feature importance ranking of the source images, the generation parameters of the generation model are adaptively adjusted. The high-importance features of the source images are used as the core generation constraints of the generated images, while the low-importance features are used as auxiliary generation constraints. At the same time, according to the system's preset generation accuracy level classification standard, the generated images are divided into three levels: primary generation level, intermediate generation level, and advanced generation level. The primary generation level meets the basic visual effect requirements, the intermediate generation level meets the requirements for accurate feature matching, and the advanced generation level meets the requirements for professional applications with ultra-high definition and high fit.

[0055] Based on the adjusted generation parameters and the selected generation accuracy level, the matching generation model is driven to generate generative images. During the generation process, the feature data of the generative images are extracted in real time and initially matched with the feature data of the source images. Generative images with obvious feature conflicts are eliminated, and generative images that pass the initial matching are retained. Generation level labels and generation feature datasets are added to generative images of each level to form a hierarchical generative image dataset.

[0056] Furthermore, if the generative image generation grading module receives substandard generative image data transmitted by the generative image quality quantization module, it analyzes the reasons for the substandard data, readjusts the generation parameters of the generation model or replaces the generation model with a suitable one, and improves the generation accuracy level to re-grade and generate the data. Finally, it transmits the graded generative image dataset to the generative image quality quantization module.

[0057] The generative image quality quantification module is a quality assessment unit for generative images. Based on the hierarchical generative image dataset transmitted by the generative image generation and grading module, it constructs a multi-dimensional generative image quality quantification system. By setting a reasonable quality quantification formula, it achieves accurate calculation of generative image quality scores. Based on the quality scores, it determines whether the generative images are qualified or unqualified. At the same time, it feeds back unqualified data to the generative image generation and grading module, which filters out high-quality generative images for subsequent fusion and reconstruction of original images, effectively avoiding the problem of poor fusion effect caused by poor quality of generative images.

[0058] Specifically, the generative image quality quantification module first receives the hierarchical generative image dataset transmitted by the generative image generation and grading module. It then parses the generative images, generation level labels, and generation feature datasets within the dataset. Simultaneously, it retrieves source image feature data from the source image adaptation and parsing dataset transmitted by the image source adaptation and parsing module. This process constructs a generative image quality quantification index system encompassing four dimensions: feature matching degree, visual smoothness, color consistency, and edge distortion. To achieve accurate quality quantification, a comprehensive generative image quality scoring formula is defined, as follows:

[0059] Q = α × F + β × S + γ × C - δ × E;

[0060] Q: Overall quality score of the generative image, ranging from 0 to 100 points. A higher score indicates better quality of the generative image.

[0061] F: Feature matching score, ranging from 0 to 100, represents the degree of matching between the features of the generated image and the features of the source image; F is obtained by calculating the cosine similarity between the feature data of the generated image and the feature data of the source image. The closer the cosine similarity is to 1, the higher the F score.

[0062] S: Visual fluency, ranging from 0 to 100, represents the overall visual effect and texture smoothness of the generated image; S is obtained by scoring the overall visual effect of the generated image through a deep learning visual evaluation model, which is trained on a large number of image visual effect samples from different scenes.

[0063] C: Color consistency coefficient, ranging from 0 to 100 points, characterizes the degree of consistency between the color channel distribution of the generated image and the color channel distribution of the source image; C is obtained by calculating the mean deviation rate of the three RGB color channels of the generated image and the source image. The smaller the mean deviation rate, the higher the C score.

[0064] E: Edge distortion, ranging from 0 to 100 points, characterizes the degree of distortion between the edge contours of the generated image and the edge contours of the source image; E is obtained by calculating the Euclidean distance between the edge contour features of the generated image and the source image. The larger the Euclidean distance, the higher the E score.

[0065] α, β, γ, δ: Weight coefficients of each indicator, and satisfy α+β+γ+δ=1. The weight coefficients are adaptively adjusted according to the scene adaptation category of the source image and retrieved from the system's preset scene-weight coefficient mapping table. This mapping table has been verified by a large number of image processing experiments in different scenes and can be updated according to actual application scenarios. For example, α takes a larger value in professional scenarios such as industrial inspection and medical imaging, while β and γ take larger values ​​in visual scenarios such as portrait photography and landscape shooting. δ is a fixed negative weight in all scenarios.

[0066] The numerical values ​​and weight coefficients of each indicator are obtained according to the above parameter acquisition method. They are then substituted into the comprehensive quality scoring formula to calculate the Q value of each generated image. Subsequently, a quality qualification threshold is set, where the quality qualification threshold for the primary generation level is 60 points, the quality qualification threshold for the intermediate generation level is 75 points, and the quality qualification threshold for the advanced generation level is 90 points. The calculated Q value is compared with the qualification threshold of the corresponding generation level to determine whether the generated image is qualified or unqualified.

[0067] The qualified generative images are screened, and quality score labels are added to form a high-quality generative image dataset. The high-quality generative image dataset is then transmitted to the original image fusion and reconstruction module. Meanwhile, the unqualified generative image data is labeled with the reason for unqualification and fed back to the generative image generation and grading module for regeneration.

[0068] The original image fusion and reconstruction module is the core fusion unit of the source image and the generated image. Based on the source image adaptation and parsing dataset of the image source adaptation and parsing module and the high-quality generated image dataset of the generated image quality quantization module, it realizes feature-level fusion and pixel-level reconstruction of the original source image and the high-quality generated image. It avoids the fusion distortion problem caused by simple pixel superposition or feature splicing, and realizes the fine fusion of the source image and the generated image, effectively improving the overall fit and visual effect of the fused image.

[0069] Simultaneously, abnormal data during the fusion process is extracted and fed back to the image source adaptation and parsing module for secondary parsing, optimizing the source image feature data, further improving the effect of subsequent fusion processing, and avoiding the repeated occurrence of fusion distortion problems;

[0070] Specifically, the original image fusion and reconstruction module first receives a high-quality generative image dataset, and at the same time calls the source image adaptation and parsing dataset transmitted by the image source adaptation and parsing module. It performs in-depth analysis on the low-level visual features and high-level semantic features of the source image and the high-quality generative image to determine the fusion region and non-fusion region of the source image and the generative image. The fusion region is the area of ​​the source image that needs to be repaired, enhanced and completed, and the non-fusion region is the core feature region of the source image, which needs to keep the original features unchanged.

[0071] First, feature-level fusion is performed, which weights the fusion region features of the generative image and the fusion region features of the source image. The fusion weights are adjusted according to the scene adaptation category of the source image. In professional scenes, the feature weights of the source image are higher, while in visual scenes, the feature weights of the generative image are higher. The consistency of feature fusion is detected in real time during the feature-level fusion process to avoid semantic conflicts.

[0072] After feature-level fusion is completed, pixel-level reconstruction is performed. The fused feature data is then subjected to pixel-level interpolation, smoothing, and sharpening to optimize the edge transition between the fused and non-fused regions, making the overall visual effect of the fused image natural and without stitching marks.

[0073] After pixel-level reconstruction is completed, the original fused reconstructed image is generated. At the same time, the data in the fusion process is extracted. If fusion anomalies such as feature conflicts or edge transition distortion occur, the abnormal data is extracted and a fusion anomaly dataset is formed. The fusion anomaly dataset is fed back to the image source adaptation and parsing module for secondary parsing. If there are no anomalies in the fusion process, the original fused reconstructed image and the corresponding fusion feature dataset are transmitted to the fused image intelligent optimization module.

[0074] The intelligent optimization module for fused images is a precise optimization unit for fused images. Based on the original fused and reconstructed images and fused feature datasets transmitted by the original image fusion and reconstruction module, and combined with the quality score of the generative image, it sets the optimization coefficient formula for fused images to achieve accurate calculation of the optimization coefficient. Based on the optimization coefficient, it performs multi-dimensional intelligent optimization of the fused images, which not only ensures the visual effect of the fused images, but also avoids image distortion caused by over-optimization. It effectively improves the scene adaptability and overall quality of the fused and optimized images and the original images. It can also receive and process unqualified image data from the image effect verification module and re-optimize them.

[0075] Specifically, the fused image intelligent optimization module first receives the original fused and reconstructed image and fused feature dataset transmitted by the original image fusion and reconstruction module. Simultaneously, it retrieves the quality score Q of the corresponding generative image transmitted by the generative image quality quantification module. Combining this with the scene adaptation category of the source image, it constructs a multi-dimensional fused image optimization system that includes color optimization, contrast optimization, sharpness optimization, and edge transition optimization. To achieve precise optimization, the module sets a formula for the fused image intelligent optimization coefficients, as follows:

[0076] K = λ × Q1 + μ × M;

[0077] Wherein, K: intelligent optimization coefficient of fused image, with a value range of 0 to 1. The larger the value of K, the higher the degree of optimization required for the fused image;

[0078] Q1: Normalized generative image quality score, Q1=Q / 100, with a value range of 0 to 1;

[0079] M: Fusion feature matching deviation rate, with a value range of 0 to 1, characterizing the degree of matching deviation between the feature data of the fused image and the feature data of the source image. It is obtained by calculating the feature deviation value between the fused feature dataset of the fused image and the source image adaptation parsing dataset of the source image, and then performing normalization processing. The larger the feature deviation value, the higher the M value.

[0080] λ and μ: weight coefficients, and satisfy λ+μ=1, where λ is the quality score weight and μ is the feature matching deviation rate weight. Preferably, the fixed values ​​are λ=0.4 and μ=0.6, that is, the degree of optimization is determined based on the degree of feature matching deviation.

[0081] The values ​​of Q1 and M are obtained according to the above parameter acquisition method. The values ​​are then substituted into the optimization coefficient formula to calculate the K value. Based on the magnitude of the K value, the optimization degree is divided into three levels: mild optimization (K≤0.3), moderate optimization (0.3<K≤0.7), and severe optimization (K>0.7). Different optimization levels correspond to different adjustment ranges of optimization parameters.

[0082] Then, based on the determined optimization level and the scene adaptation category of the source image, the fused image is intelligently optimized in multiple dimensions. Light optimization only makes minor adjustments to the color and contrast of the fused image. Medium optimization adds moderate adjustments to sharpness and edge transition on the basis of color and contrast adjustments. Heavy optimization makes significant and refined adjustments to all optimization dimensions. During the optimization process, the optimization effect is displayed in real time and feature matching detection is performed to ensure that there are no feature conflicts in the optimized image. After optimization is completed, the fused optimized image and optimization parameter dataset are generated and transmitted to the image processing effect verification module.

[0083] Furthermore, if the fused image intelligent optimization module receives unqualified fused optimization image data transmitted by the processed image effect verification module, it will analyze the reasons for the unqualification, recalculate the optimization coefficient K and increase the optimization level for re-optimization. After the second optimization is completed, the fused optimization image and optimization parameter dataset are regenerated and transmitted to the processed image effect verification module.

[0084] The image processing effect verification module is the final effect verification unit of the system. Based on the fused and optimized image and optimization parameter dataset transmitted by the fused image intelligent optimization module, it constructs an image processing effect verification system. By setting an effect matching degree formula, it calculates the matching degree between the processed image and the source image to meet the scene adaptation requirements. Based on the matching degree, it determines whether the processed image is qualified. The qualified processed image is used as the final output result of the system, and the unqualified image data is fed back to the fused image intelligent optimization module. The specific operation process is as follows:

[0085] First, the system receives the fused and optimized image and optimization parameter dataset from the fused image intelligent optimization module. Simultaneously, it retrieves the source image adaptation and analysis dataset from the image source adaptation and analysis module. Based on the scene adaptation category of the source image, it determines the effect verification standard for the processed image. For professional scenes such as industrial inspection and medical imaging, the verification standard focuses on feature accuracy and data authenticity; for visual scenes such as portrait photography and landscape photography, the verification standard focuses on visual effect and color aesthetics. To achieve accurate effect verification, a formula for the scene adaptation matching degree of the processed image is set, as follows:

[0086] ;

[0087] Wherein, P is the scene adaptation matching degree of the processed image, and the value ranges from 0% to 100%. The higher the P value, the more the processed image matches the scene adaptation requirements of the source image.

[0088] If represents the quantized value of the core features of the fused and optimized image, and is a quantization index that characterizes the core features of the fused and optimized image. If is obtained by extracting the corresponding core features from the optimization parameter dataset of the fused and optimized image and quantizing them. The quantization method is completely consistent with Is, ensuring the comparability of the parameters.

[0089] Is is the core feature quantization value of the source image, representing the quantification index of the core features of the source image; Is is obtained by extracting core features from the source image adaptation parsing dataset and quantizing them according to the scene adaptation category of the source image. The core features are the key features corresponding to the scene adaptation category. For example, the core feature of the industrial inspection scene is the pixel coordinates of the product defect contour, the core feature of the medical imaging scene is the gray value distribution of the lesion area, and the core feature of the portrait photography scene is the contour features and skin color value of the portrait face.

[0090] That is, |If-Is| is the absolute difference between the quantized values ​​of the core features of the fused and optimized image and the source image;

[0091] The values ​​of If and Is are obtained according to the above parameter acquisition method. The P value is calculated by substituting them into the scene adaptation matching degree formula. The general acceptable threshold for the system is set to P≥85%, and a higher acceptable threshold of P≥95% is set for the processed images of advanced generation level.

[0092] The calculated P-value is compared with the corresponding qualified threshold. If the P-value reaches or exceeds the qualified threshold, the fused and optimized image is determined to be qualified. The module adds scene adaptation tags and processing quality reports to the qualified images, forming the final processed image dataset of the system and transmitting it externally as the system's output. If the P-value does not reach the qualified threshold, the fused and optimized image is determined to be unqualified. The module adds verification reason tags to the unqualified images and feeds them back to the fused image intelligent optimization module for re-optimization.

[0093] Example 2: Figure 2 As shown, the difference between this embodiment and Embodiment 1 is that the image intelligent processing method based on generative images proposed in this invention includes the following steps:

[0094] Step 1: Source Image Parsing and Transmission:

[0095] The image source adaptation and parsing module is activated to perform format normalization, feature extraction, and scene adaptation determination on the input original source image, and generate an adaptation parsing dataset which is then transmitted to the generative image generation and hierarchical module; if abnormal fusion data is received, it is parsed and corrected a second time before being retransmitted.

[0096] Step 2: Generate images in stages:

[0097] The generative image generation grading module matches the generation model corresponding to the scene, adjusts the parameters, and generates images according to accuracy grading. After preliminary screening, a graded generation dataset is formed and transmitted to the generative image quality quantification module. If unqualified data is received, the parameters / model are adjusted, the image is regenerated, and then retransmitted.

[0098] Step 3: Quality Quantitative Judgment

[0099] The generative image quality quantification module constructs a four-dimensional index system, calculates the comprehensive quality score Q, sets a threshold according to the generation level to determine the passability, and the qualified images form a high-quality dataset that is transmitted to the original image fusion and reconstruction module, while the unqualified data is fed back to the generative image generation and grading module.

[0100] Step 4: Original Fusion and Reconstruction

[0101] The original image fusion and reconstruction module performs feature-level fusion and pixel-level reconstruction on the source image and the high-quality generated image, and transmits the fused image to the fused image intelligent optimization module; if a fusion anomaly occurs, it is fed back to the image source adaptation and parsing module for secondary parsing.

[0102] Step 5: Image fusion optimization:

[0103] The image intelligent optimization module calculates the optimization coefficient K, performs multi-dimensional optimization at different levels, generates an optimized image and transmits it to the image processing effect verification module; if unqualified data is received, the coefficient is recalculated, the level is increased, the optimization is repeated and the image is retransmitted.

[0104] Step Six: Output the Result Verification:

[0105] The image processing effect verification module calculates the scene adaptation matching degree P, judges the passability according to the threshold, and outputs the passable image as the final result. The unqualified data is fed back to the fusion image intelligent optimization module for re-optimization.

[0106] The working principle of this invention is as follows: In use, the image source adaptation and parsing module realizes the normalization and refined feature extraction of source images of multiple formats, resolutions, and scenes, laying a precise data foundation for subsequent processing. The generative image generation and grading module matches the scene adaptation model and adjusts the parameters as needed, generating images according to accuracy to meet the differentiated needs of different application scenarios. The generative image quality quantification module selects high-quality generated images through multi-dimensional indicators to avoid low-quality data affecting the fusion effect. The original image fusion and reconstruction module realizes feature-level fusion and pixel-level reconstruction, optimizes edge transitions, and reduces fusion distortion. The fused image intelligent optimization module performs precise optimization according to the optimization coefficient, taking into account both visual effect and data authenticity. The processed image effect verification module sets targeted verification standards to ensure the scene adaptability and reliability of the final output image.

[0107] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize it. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An image intelligent processing system based on generative images, characterized in that, It includes an image source adaptation and parsing module, a generative image generation and grading module, a generative image quality quantification module, a source image fusion and reconstruction module, and a fused image intelligent optimization module; The image source adaptation and parsing module performs format normalization, feature extraction and scene adaptability determination on the source image, and the generative image generation and grading module divides the generative image into different levels according to the generation accuracy. The generative image quality quantization module constructs a multi-dimensional generative image quality quantization system, and judges the generative image as qualified or unqualified based on the quality score. The original image fusion and reconstruction module is used to perform feature-level fusion and pixel-level reconstruction of the original source image and the high-quality generated image. The fused image intelligent optimization module performs multi-dimensional intelligent optimization of the fused image based on the optimization coefficient.

2. The image intelligent processing system based on generative images according to claim 1, characterized in that, The image source adaptation and parsing module performs format recognition and normalization on the original source image, and adaptively scales it according to the resolution of the source image. The low-level visual features and high-level semantic features of the normalized source image are extracted by a deep learning feature extraction network. The extracted source image features are then used to determine scene adaptability. Based on a pre-set scene feature library, the source image is matched to the corresponding application scene category, generating a source image adaptation parsing dataset.

3. The image intelligent processing system based on generative images according to claim 2, characterized in that, If the image source adaptation parsing module receives fusion anomaly data transmitted by the original image fusion reconstruction module, it extracts the feature conflict information in the fusion anomaly data, performs secondary feature parsing in combination with the feature data of the original source image, corrects the feature extraction results, and regenerates the source image adaptation parsing dataset. Finally, the source image adaptation parsing dataset is transmitted to the generative image generation hierarchical module through the system's internal data interface.

4. The image intelligent processing system based on generative images according to claim 1, characterized in that, The generative image generation and classification module receives the source image adaptation parsing dataset, parses the scene adaptation categories and feature importance ranking in the dataset, and selects the generative model that matches the scene of the source image from the generative image model library according to the system's preset generative model-scene adaptation mapping table. Based on the feature importance ranking of the source images, the generation parameters of the generation model are adaptively adjusted. At the same time, according to the system's preset generation accuracy level classification standard, the generated images are divided into three levels: primary generation level, intermediate generation level, and advanced generation level. Based on the adjusted generation parameters and the selected generation accuracy level, the matching generation model is driven to generate generative images. During the generation process, the feature data of the generative images are extracted in real time and initially matched with the feature data of the source images. Generation level labels and generation feature datasets are added to each level of generative images to form a hierarchical generative image dataset.

5. The image intelligent processing system based on generative images according to claim 4, characterized in that, If the generative image generation grading module receives substandard generative image data transmitted by the generative image quality quantization module, it analyzes the reasons for the substandard data, readjusts the generation parameters of the generation model or replaces the generation model with a suitable one, and improves the generation accuracy level to re-grade and generate the data. Finally, it transmits the graded generative image dataset to the generative image quality quantization module.

6. The image intelligent processing system based on generative images according to claim 5, characterized in that, The generative image quality quantification module constructs a generative image quality quantification index system, sets a comprehensive scoring formula for generative image quality to calculate the Q value, compares the calculated Q value with the corresponding qualified threshold of the generation level, and determines whether the generative image is qualified or unqualified. The qualified generative images are then screened, quality score labels are added to form a high-quality generative image dataset, and the high-quality generative image dataset is transferred to the original image fusion and reconstruction module.

7. The image intelligent processing system based on generative images according to claim 6, characterized in that, The original image fusion and reconstruction module performs in-depth analysis of the low-level visual features and high-level semantic features of the source image and the high-quality generative image to determine the fusion region and non-fusion region of the source image and the generative image. First, feature-level fusion is performed. After feature-level fusion is completed, pixel-level reconstruction is performed. The fused feature data is then subjected to pixel-level interpolation, smoothing, and sharpening to optimize the edge transition effect between the fused and non-fused regions. After pixel-level reconstruction is completed, the original fused and reconstructed image is generated.

8. The image intelligent processing system based on generative images according to claim 1, characterized in that, The intelligent image optimization module constructs a multi-dimensional image optimization system. Based on the calculated K value, the optimization level is divided into three levels: light optimization, medium optimization, and heavy optimization. Different optimization levels correspond to different adjustment ranges of optimization parameters. Based on the determined optimization level and the scene adaptation category of the source image, the fused image is intelligently optimized in multiple dimensions.

9. The image intelligent processing system based on generative images according to claim 8, characterized in that, The integrated image intelligent optimization module communicates with the image processing effect verification module. The image processing effect verification module constructs an image processing effect verification system to determine whether the processed image is qualified. The qualified processed image is used as the final output result of the system. The specific operation process is as follows: Based on the scene adaptation category of the source image, the effect verification standard of the processed image is determined. The calculated P value is compared with the corresponding qualified threshold. If the P value is not less than the corresponding qualified threshold, the fused and optimized image is judged to be qualified; otherwise, the fused and optimized image is judged to be unqualified.

10. An image intelligent processing method based on generative images, employing the image intelligent processing system based on generative images as described in any one of claims 1-9, characterized in that, Includes the following steps: Step 1: Source image parsing and transmission; Step 2: Hierarchical image generation; Step 3: Quality quantification and judgment; Step 4: Source and source image fusion and reconstruction; Step 5: Fusion image optimization; Step 6: Effect verification and output.

Images