Intelligent image preprocessing method and system based on AI

Abstract

The invention discloses an AI-based intelligent image preprocessing method and system, and the method comprises the steps: obtaining to-be-processed image data, separating a main image region from a background region, generating a main region mask and a background region mask, and determining an image type label; dividing the main image into sub-regions with independent geometrical characteristics based on the main region mask, and executing pixel-level image preprocessing on each sub-region; constructing a sub-region association graph and performing feature propagation by calculating visual association weights among the sub-regions to obtain consistency correction parameters among the sub-regions; processing the pixels of each sub-region, enabling the adjacent regions to be in smooth transition, and finally synthesizing an optimized main image; and carrying out background color temperature migration on the optimized main image based on the background region mask, and adjusting background ambiguity. The method has the advantages that through the technologies of image segmentation, sub-region division, image convolutional network feature propagation and the like, the image quality is optimized, and natural transition between the main region and the background of the image is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to AI image processing, in particular to an AI-based intelligent image preprocessing method and system. BACKGROUND

[0002] With the wide application of digital images, especially in medical imaging, satellite remote sensing, security monitoring, autonomous driving and other fields, image preprocessing technology has attracted widespread attention. In medical image processing, image preprocessing can help doctors see the lesion area more clearly; in autonomous driving, preprocessing can enhance the image data collected by the camera and improve the recognition accuracy. The purpose of image preprocessing is to improve the analyzability and quality of the image, and to ensure the effectiveness and accuracy of subsequent algorithms such as pattern recognition and machine learning, which is one of the indispensable technologies in modern intelligent systems.

[0003] Current image preprocessing methods and systems on the market rely on manually designed feature extraction and simple image processing techniques, such as fixed filtering algorithms, color space conversion and noise suppression. These methods are difficult to adapt to complex image content and cannot dynamically adjust the processing strategy to cope with different image features. In particular, in image region segmentation, traditional methods mostly use threshold segmentation or edge detection-based algorithms, which are easily affected by image noise and complex background, resulting in insufficient segmentation accuracy. Traditional methods often lack a global perspective when dealing with visual relevance and consistency between different sub-regions, and cannot consider the mutual influence and overall structure between sub-regions. This leads to discontinuity and loss of details in image processing, especially in complex image scenarios, where the boundary transition appears harsh and smooth transition is difficult to achieve. SUMMARY

[0004] To improve existing methods and systems, an AI-based intelligent image preprocessing method and system is provided, which realizes high-quality image optimization by precise image region separation, pixel-level optimization processing and consistency correction, and improves image details, clarity and natural transition, suitable for scenarios requiring high-precision image processing.

[0005] To achieve the above purpose, the technical solution adopted by the present application is: An AI-based intelligent image preprocessing method, comprising: Obtaining image data to be processed, inputting a pre-trained key information extraction neural network model, separating main image regions and background image regions in the image, generating a main region mask and a background region mask, and obtaining an image type label; Based on the main region mask, the main image is divided into sub-regions independent of geometric features, and each sub-region contains a continuous set of pixels; Performing pixel-level image processing operations on each sub-region, including adaptive filtering, color space conversion and noise suppression; Based on the image type label, the visual correlation weight between each sub-region is calculated, the image correlation of each sub-region is obtained, the sub-region correlation graph is constructed, the feature propagation of the sub-region correlation graph is performed through the graph convolution network, and the consistency correction parameter between the sub-regions is obtained. Based on the consistency correction parameter, the pixels of each sub-region are processed to make smooth transition between adjacent sub-regions, and the processed sub-regions are combined to obtain an optimized main image. Based on the background region mask, the color histogram of the optimized main image is obtained to perform background color temperature migration, and the background blur degree is adjusted according to the main region boundary distance.

[0006] Preferably, the obtaining of the image data to be processed, the input of the pre-trained key information extraction neural network model, the separation of the main image region and the background image region in the image, the generation of the main region mask and the background region mask, and the obtaining of the image type label specifically include: The image data to be processed is obtained and input into the pre-trained key information extraction neural network model for forward propagation; The image segmentation is performed through the key information extraction neural network model to obtain the main image region and the background image region of the image; According to the output image segmentation result, the main region mask is generated based on the main image region, and the background region mask is generated based on the background image region, the pixel value of the main region is 1, and the pixel value of the background region is 0; Based on the key information extraction neural network model, the image classification is performed at the same time of the image segmentation, and the type label is added to each image.

[0007] Preferably, the main image is divided into sub-regions with independent geometric features based on the main region mask, and each sub-region contains a continuous pixel set, specifically including: Based on the obtained main region mask, the connected domain analysis is performed, and the continuous pixel part in the main region mask is obtained according to the 8-neighbor connectedness; Based on all the obtained connected domain parts, the independence detection is performed to divide the sub-regions, and the shape features of each sub-region are obtained through the geometric feature extraction; Based on the shape features of each sub-region, the merging and segmentation of the regions are performed to enhance the independence of the shape features of each sub-region; Based on the segmented sub-regions, the continuous pixel set in each sub-region is obtained.

[0008] Preferably, the pixel-level image processing operation performed on each sub-region includes adaptive filtering, color space conversion and noise suppression, specifically including: Based on the local characteristics of each sub-region, a suitable filter window size is obtained, and the statistical characteristics of the sub-region are calculated in the neighborhood of each pixel, and the parameters of the filter are dynamically adjusted; Based on the obtained local characteristics of each sub-region, color space conversion processing is performed to convert the color information of the sub-region from the original color space to the target color space, including RGB, HSV, Lab. Each type of noise is suppressed by mean filtering, Gaussian filtering and median filtering.

[0009] Preferably, based on the image type label, the visual correlation weight between each sub-region is calculated, the image correlation of each sub-region is obtained, the sub-region correlation graph is constructed, and the feature propagation of the sub-region correlation graph is performed through the graph convolution network to obtain the consistency correction parameter between the sub-regions, which specifically includes: Based on the image type label obtained by the key information extraction neural network model, the geometric and visual features of each sub-region are extracted; Based on the obtained geometric and visual features of the sub-region, the similarity between the sub-regions is calculated, and the correlation weight is set based on the similarity size; Based on the image correlation of each sub-region, a sub-region correlation graph is constructed, each sub-region is regarded as a node in the graph, the visual correlation between the sub-regions is regarded as the edge between the nodes, and the correlation weight is regarded as the weight of the edge. Based on the graph convolution network, the sub-region feature propagation is performed to enhance the expression of each node to its surrounding environment, and the final representation of each sub-region is obtained. Based on the final representation of each sub-region, the difference between the sub-region features is calculated to obtain the consistency correction parameter.

[0010] Preferably, based on the consistency correction parameter, the pixels of each sub-region are processed to make smooth transition between adjacent sub-regions, and the processed sub-regions are combined to obtain an optimized main image, which specifically includes: Based on the obtained consistency correction parameter, the pixel features of each sub-region are adjusted, including fine-tuning the color and shape geometric features of the sub-region boundary; Based on the adjusted sub-regions, the adjacent sub-regions are smoothly transitioned by Gaussian smoothing; Based on the sub-regions after the smooth transition processing, the sub-regions are merged, and the global consistency is adjusted to obtain an optimized main image region.

[0011] Preferably, based on the background region mask, the color histogram of the optimized main image is obtained for background color temperature migration, and the background blur degree is adjusted according to the main region boundary distance, which specifically includes: Based on the background region, the color histogram of the optimized main image is obtained, and the color temperature of the background region is migrated by adjusting the color channel of the background region based on the color histogram; By calculating the distance between each pixel of the background region and the boundary of the main region, the pixel blur degree is defined, and the background region is blurred; The background region after color temperature migration and blur adjustment is merged with the main image region to obtain a complete image.

[0012] Further, an AI-based intelligent image preprocessing system is proposed, which includes: An image preprocessing module: the image preprocessing module is used to input the to-be-processed image data into a pre-trained key information extraction neural network model, separate the main image region and the background region in the image, generate a mask and perform image classification; A region division module: the region division module is based on the main region mask, which performs connected component analysis, divides the main image region into sub-regions with independent geometric features, and extracts the shape features of each sub-region; An image processing module: the image processing module performs pixel-level processing on each sub-region, including adaptive filtering, color space conversion and noise suppression, to optimize the local characteristics of the sub-region; An association calculation module: the association calculation module obtains the association weight between each sub-region, constructs a sub-region association graph and performs feature propagation, and obtains a consistency correction parameter between sub-regions; A sub-region synthesis module: the sub-region synthesis module adjusts the pixels of the sub-region based on the consistency correction parameter and performs smooth transition processing, and finally synthesizes the optimized main image region; A background region adjustment and image synthesis module: the background region adjustment and image synthesis module obtains the color histogram of the optimized main image based on the background region mask, performs background color temperature migration and adjusts the background blur degree, and finally synthesizes a complete image; A processor: the processor is used to process the calculation process of each formula and the construction and calculation process of each model.

[0013] Compared with the prior art, the advantages of the present application are: By combining the pre-trained key information extraction neural network model, the main image area and the background area can be accurately separated, and a clear image structure is provided for subsequent processing. The sub-region division and pixel-level processing operation in the method make each sub-region independent and adaptive, thereby improving the clarity and authenticity of image details. The graph convolution network propagates features between sub-regions to ensure seamless connection between regions and avoid unnatural transitions between regions in traditional methods. In addition, the background region color temperature migration and blur adjustment make the optimized image more natural and real, with higher visual perception quality. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 An AI-based intelligent image preprocessing method is provided for the present application. Figure 2 A region mask diagram is provided for the present application. Figure 3 A sub-region division diagram is provided for the present application. Figure 4 A sub-region image processing diagram is provided for the present application. Figure 5 A consistency correction parameter diagram is provided for the present application. Figure 6 An optimized main image diagram is provided for the present application. Figure 7 A background region processing and merging diagram is provided for the present application. DETAILED DESCRIPTION

[0015] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be made by those skilled in the art.

[0016] The AI-based intelligent image preprocessing system comprises: An image preprocessing module: the image preprocessing module is used for inputting the to-be-processed image data into a pre-trained key information extraction neural network model, separating the main image area and the background area in the image, generating a mask and performing image classification; A region division module: the region division module is based on the main region mask, the module performs connected component analysis, divides the main image area into sub-regions with independent geometric features, and extracts the shape features of each sub-region; An image processing module: the image processing module performs pixel-level processing on each sub-region, including adaptive filtering, color space conversion and noise suppression, to optimize the local characteristics of the sub-region; Correlation calculation module: the correlation calculation module obtains the correlation weight between each sub-region, constructs a sub-region correlation graph and performs feature propagation, and obtains a consistency correction parameter between sub-regions; Sub-region synthesis module: the sub-region synthesis module adjusts the pixels of the sub-region based on the consistency correction parameter, and performs smooth transition processing, and finally synthesizes the optimized main image region; Background region adjustment and image synthesis module: the background region adjustment and image synthesis module obtains the optimized main image color histogram according to the background region mask, performs background color temperature migration and adjusts the background blur degree, and finally synthesizes a complete image; Processor: the processor is used for processing the calculation process of each formula and the construction calculation process of each model.

[0017] Referring to Figure 1 The AI-based intelligent image preprocessing method comprises the following steps: Step one: obtain the image data to be processed, input the pre-trained key information extraction neural network model, separate the main image region and the background image region in the image, generate the main region mask and the background region mask, and obtain the image type label; Step two: based on the main region mask, the main image is divided into sub-regions independent in geometric features, and each sub-region contains a continuous pixel set; Step three: perform pixel-level image processing operations on each sub-region, including adaptive filtering, color space conversion and noise suppression; Step four: based on the image type label, calculate the visual correlation weight between each sub-region, obtain the image correlation of each sub-region, construct a sub-region correlation graph, and perform feature propagation on the sub-region correlation graph through a graph convolution network to obtain a consistency correction parameter between sub-regions; Step five: based on the consistency correction parameter, process the pixels of each sub-region to make smooth transition between adjacent sub-regions, and combine the processed sub-regions to obtain an optimized main image; Step six: based on the background region mask, obtain the color histogram of the optimized main image for background color temperature migration, and adjust the background blur degree according to the main region boundary distance.

[0018] Referring to Figure 2 As shown in the figure, the image data to be processed is obtained, the pre-trained key information extraction neural network model is input, the main image region and the background image region in the image are separated, the main region mask and the background region mask are generated, and the image type label is obtained. Specifically, it comprises: The image data to be processed is obtained and input into the pre-trained key information extraction neural network model for forward propagation; An image is segmented by a key information extraction neural network model to obtain a main image region and a background image region; According to the output image segmentation result, a main region mask is generated based on the main image region, and a background region mask is generated based on the background image region, the pixel value of the main region is 1, and the pixel value of the background region is 0; Based on the key information extraction neural network model, image classification is performed at the same time of image segmentation, and a type label is added to each image.

[0019] Specifically, load the image data to be processed and convert it into a tensor format suitable for neural network input. The image may come from a file, a camera or a data stream, and needs to be preprocessed to match the training data distribution of the model; The image tensor is input into the pre-trained key information extraction neural network model, and forward propagation calculation is performed. The model simultaneously processes segmentation and classification tasks, and outputs a segmentation probability map and a classification probability vector. Based on the output segmentation probability map, threshold processing is applied to quantize the probability Figure Two value, separate the main image region and the background image region, and the threshold is used to decide the pixel attribution; According to the segmentation result, a binary mask tensor is generated. In the main region mask, the pixel value of the main region is 1, and the pixel value of the background region is 0. In the background region mask, the pixel value of the background region is 1, and the pixel value of the main region is 0. The background region mask can be derived directly from the main region mask; In the forward propagation, the model simultaneously outputs a classification probability vector. Based on this vector, the predicted class of the image is determined by taking the maximum value index, and a type label is added.

[0020] Referring to Figure 3 As shown, based on the main region mask, the main image is divided into sub-regions independent in geometric features, each sub-region containing a continuous pixel set, which specifically includes: Based on the obtained main region mask, a connected domain analysis is performed, and the continuous pixel part in the main region mask is obtained according to the 8-neighbor connectivity; Based on all the obtained connected domain parts, independence detection is performed to divide the sub-regions, and the shape features of each sub-region are obtained through geometric feature extraction; Based on the shape features of each sub-region, the regions are merged and segmented to enhance the independence of the shape features of each sub-region; Based on each sub-region after segmentation, a continuous pixel set in the sub-region is obtained.

[0021] Specifically, a pixel position is given, and its 8-neighborhood is defined as: ; For each pixel in the mask image, if its value is 1 and has not been marked, perform a breadth-first search or depth-first search starting from that pixel, and mark all connected pixels as the same connected component. Each connected component will be assigned a unique label; The goal of independence detection is to determine the boundaries between connected components and ensure that they do not overlap or touch. The detection method can be achieved by checking the adjacency of each connected component. If the boundary of one connected component touches or overlaps with the boundary of another connected component, they belong to the same group of regions; The purpose of geometric feature extraction is to analyze the properties of sub-regions by calculating their shape features. Common shape features include area, perimeter, shape moment, and aspect ratio; Based on geometric features, the connected components are divided by clustering algorithm, ensuring that regions with similar shapes are grouped together. For example, if two sub-regions have similar aspect ratios and areas, they may belong to the same type of region; If two or more sub-regions are very close in geometry and are considered to belong to the same object through independence detection, they can be merged into a larger region. The shape features of the merged region, such as area and perimeter, will be updated based on its constituent parts; if a sub-region is too complex or contains multiple objects, it can be further divided into multiple sub-regions using shape features to determine the appropriate division method; Based on the segmented sub-regions, obtain the continuous pixel set within each sub-region. For each sub-region, find its connected pixel set, i.e., make these pixels remain connected in space.

[0022] Referring to Figure 4 As shown, perform pixel-level image processing operations on each sub-region, including adaptive filtering, color space conversion, and noise suppression, which specifically include: Based on the local characteristics of each sub-region, obtain the appropriate filter window size, and calculate the statistical characteristics of the sub-region within the neighborhood of each pixel, and dynamically adjust the parameters of the filter; Based on the obtained local characteristics of each sub-region, perform color space conversion processing to convert the color information of the sub-region from the original color space to the target color space, including RGB, HSV, Lab; Suppress each type of noise through mean filtering, Gaussian filtering, and median filtering.

[0023] Referring to Figure 5 Based on the image type label, calculate the visual correlation weight between each sub-region to obtain the image correlation of each sub-region, construct a sub-region correlation graph, and perform feature propagation on the sub-region correlation graph through a graph convolution network to obtain consistency correction parameters, which specifically include: Based on the image type label obtained by the key information extraction neural network model, the geometric and visual features of each sub-region are extracted; Based on the obtained geometric and visual features of the sub-regions, similarity calculation between sub-regions is performed, and correlation weight setting is performed based on the similarity size; Based on the image correlation of each sub-region, a sub-region correlation graph is constructed, each sub-region is regarded as a node in the graph, the visual correlation relationship between sub-regions is regarded as the edge between nodes, and the correlation weight is regarded as the weight of the edge; Based on the graph convolution network, the sub-region feature propagation is performed to enhance the expression of each node to its surrounding environment relationship, and the final representation of each sub-region is obtained; Based on the final representation of each sub-region, the difference between the sub-region features is calculated to obtain a consistency correction parameter.

[0024] Specifically, after extracting the geometric and visual features, the similarity between each sub-region can be calculated, and the cosine similarity is used to measure the similarity of the visual features and the geometric features between the sub-regions. Based on the calculated similarity, the correlation weight is set to represent the visual correlation between the sub-regions. If the similarity between the sub-regions is high, the correlation weight will be large, indicating that the edge between them in the graph has a strong connection; According to the similarity and correlation weight calculated above, the correlation graph of the sub-regions is constructed, each sub-region in the graph is a node, and the edge between the nodes represents the correlation relationship between the sub-regions, and a weighted graph can be constructed; The graph convolution network is used for node feature propagation and aggregation in the graph, and the node feature propagation is performed through multi-layer convolution operation, so that the feature of each sub-region not only contains its own information, but also contains the information of the neighbor sub-region. Finally, after the propagation of multiple layers of GCN, the final representation of each sub-region is obtained; The consistency correction parameter is obtained by calculating the difference between the final feature representations of the sub-regions. The difference between the sub-region features can be calculated by Euclidean distance.

[0025] Referring to Figure 6 As shown, the pixels of each sub-region are processed based on the consistency correction parameter to make smooth transition between adjacent sub-regions, and the processed sub-regions are combined to obtain an optimized main image, which specifically includes: Based on the obtained consistency correction parameter, the pixel features of each sub-region are adjusted, including fine-tuning the color and shape geometric features of the sub-region boundary; Based on the adjusted sub-regions, the adjacent sub-regions are smoothly transitioned through Gaussian smoothing; Based on the sub-regions after the smooth transition processing, the sub-regions are merged, and the optimized main image region is obtained through global consistency adjustment.

[0026] Specifically, based on the previously calculated consistency correction parameters between sub-regions, we adjust the pixel features of each sub-region. These adjustments include fine-tuning the color and shape geometry of the sub-region boundary. When fine-tuning the color, the consistency correction parameter can adjust the color in the following way, with the formula being:

[0027] wherein, is the color of the sub-region is the color after fine-tuning, is the weight coefficient, representing the degree of influence of the consistency correction parameter on color adjustment, is the consistency correction parameter of the sub-region ; For each pixel in each sub-region, we adjust based on the corrected geometric and visual features; After completing the consistency correction, we need to perform smooth transition between adjacent sub-regions to ensure that the boundaries and transition parts of the image are more natural. Gaussian smoothing is a common smoothing technique used to smooth image or pixel data, reduce noise and optimize transition effects; After performing smooth transition, we next adjust the sub-regions based on global consistency to obtain the optimized main image region. Sub-regions are merged based on the following criteria, including geometric consistency, visual similarity and similarity measure; Global consistency adjustment ensures that the merging of all sub-regions does not result in the loss or distortion of important features in the image, resulting in an optimized main image region with higher visual consistency, geometric accuracy and spatial continuity. Referring to

[0028] , based on the background region mask, the color histogram of the optimized main image is obtained for background color temperature migration, and the background blur degree is adjusted according to the main region boundary distance. Specifically, it includes: Figure 7 Based on the background region, the color histogram of the optimized main image is obtained, and the background region is color temperature migrated by adjusting the color channel of the background region based on the color histogram; By calculating the distance between each pixel of the background region and the main region boundary, pixel blur degree is defined, and the background region is blurred; The background region after color temperature migration and blur adjustment is merged with the main image region to obtain a complete image.

[0029] ​Specifically, the color channel adjustment is performed by calculating the color histogram difference between the background region and the main region; the blur degree of the background region is defined by calculating the distance of each pixel to the boundary of the main region, and for each pixel of the background region, the distance of the pixel to the boundary of the main region is calculated, and the blur degree of the pixel is defined as the distance-based weighted blur degree by a Gaussian function, and the farther the distance, the smaller the blur degree of the pixel; After the color temperature migration and the blur processing, the adjusted background region and the main region are merged to obtain a complete image.

[0030] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above describes specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0031] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments.

[0032] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. An AI-based intelligent image preprocessing method, characterized in that: include: Obtain the image data to be processed, input the pre-trained key information extraction neural network model, separate the main image area and background image area in the image, generate the main area mask and background area mask, and obtain the image type label; Based on the main region mask, the main image is divided into geometrically independent sub-regions, each of which contains a continuous set of pixels; Perform pixel-level image processing operations on each sub-region, including adaptive filtering, color space conversion, and noise suppression; Based on the image type labels, the visual association weights between sub-regions are calculated to obtain the image association of each sub-region, and a sub-region association graph is constructed. The features of the sub-region association graph are propagated through the graph convolutional network to obtain the consistency correction parameters between sub-regions. The pixels of each sub-region are processed based on the consistency correction parameter to achieve a smooth transition between adjacent sub-regions, and the processed sub-regions are combined to obtain an optimized main image; Based on the background area mask, the color histogram of the optimized main image is obtained to perform background color temperature migration, and the background blur is adjusted according to the distance from the main area boundary.

2. The AI-based intelligent image preprocessing method according to claim 1, characterized in that: The steps of obtaining the image data to be processed, inputting the pre-trained key information extraction neural network model, separating the main image area and the background image area in the image, generating the main area mask and the background area mask, and obtaining the image type label specifically include: Obtain the image data to be processed and input it into the pre-trained key information extraction neural network model for forward propagation; Image segmentation is performed through a key information extraction neural network model to obtain the main image area and background image area of ​​the image; According to the output image segmentation results, a main area mask is generated based on the main image area, and a background area mask is generated based on the background image area. The pixel value of the main area is 1, and the pixel value of the background area is 0; Based on the key information extraction neural network model, image classification is performed at the same time as image segmentation, and type labels are added to each image.

3. The AI-based intelligent image preprocessing method according to claim 1, characterized in that: The main image is divided into geometrically independent sub-regions based on the main region mask, and each sub-region contains a continuous set of pixels, specifically including: Perform connected domain analysis based on the obtained main region mask, and obtain the continuous pixel part in the main region mask according to the 8-neighborhood connectivity; Based on all the connected domains obtained, independence detection is performed to divide the sub-regions, and the shape features of each sub-region are obtained by geometric feature extraction; Merge and segment regions based on the shape features of each sub-region to enhance the independence of the shape features of each sub-region; Based on each segmented sub-region, a continuous pixel set in the sub-region is obtained.

4. The AI-based intelligent image preprocessing method according to claim 1, characterized in that: The performing of pixel-level image processing operations on each sub-region, including adaptive filtering, color space conversion, and noise suppression, specifically includes: Obtain the appropriate filter window size based on the local characteristics of each sub-region, calculate the statistical characteristics of the sub-region in the neighborhood of each pixel, and dynamically adjust the filter parameters; Based on the local characteristics of each sub-region, a color space conversion process is performed to convert the color information of the sub-region from the original color space to the target color space, including RGB, HSV, and Lab; Various types of noise are suppressed through mean filtering, Gaussian filtering and median filtering.

5. The AI-based intelligent image preprocessing method according to claim 1, characterized in that: The method of calculating the visual association weights between sub-regions based on the image type labels, obtaining the image association of each sub-region, constructing a sub-region association graph, performing feature propagation on the sub-region association graph through a graph convolutional network, and obtaining consistency correction parameters between sub-regions specifically includes: Based on the labels of each image type obtained by the key information extraction neural network model, feature extraction is performed on the geometric and visual features of each sub-region; Calculate the similarity between sub-regions based on the obtained geometric and visual features of the sub-regions, and set the association weight based on the similarity; Based on the image correlation of each sub-region, a sub-region correlation graph is constructed. Each sub-region is regarded as a node in the graph, the visual correlation between sub-regions is regarded as the edge between nodes, and the correlation weight is regarded as the weight of the edge. Based on the graph convolutional network, sub-region feature propagation is performed to enhance the expression of each node's relationship with its surrounding environment and obtain the final representation of each sub-region; Based on the final representation of each sub-region, the difference between the sub-region features is calculated to obtain the consistency correction parameter.

6. The AI-based intelligent image preprocessing method according to claim 1, characterized in that: The processing of pixels of each sub-region based on the consistency correction parameter to achieve smooth transition between adjacent sub-regions and combining the processed sub-regions to obtain the optimized main image specifically includes: Based on the obtained consistency correction parameters, the pixel features of each sub-region are adjusted, including fine-tuning the color, shape and geometric features of the sub-region boundary; Based on the adjusted sub-regions, the adjacent sub-regions are smoothly transitioned by Gaussian smoothing; The sub-regions after smooth transition processing are merged and the optimized main image region is obtained through global consistency adjustment.

7. The AI-based intelligent image preprocessing method according to claim 1, characterized in that: The method of obtaining the optimized color histogram of the main image based on the background area mask to perform background color temperature migration and adjusting the background blur according to the main area boundary distance specifically includes: Based on the background area, a color histogram of the optimized main image is obtained, and the color temperature of the background area is shifted by adjusting the color channel of the background area based on the color histogram; By calculating the distance between each pixel in the background area and the boundary of the main area, the pixel fuzziness is defined and the background area is blurred; The background area after color temperature shift and blur adjustment is merged with the main image area to obtain a complete image.

8. An AI-based intelligent image preprocessing system, configured to implement the AI-based intelligent image preprocessing method according to any one of claims 1 to 7, characterized in that: include: Image preprocessing module: The image preprocessing module is used to input the image data to be processed into the pre-trained key information extraction neural network model, separate the main image area and background area in the image, generate a mask and perform image classification; Region division module: The region division module is based on the main region mask. This module performs connected domain analysis, divides the main image region into geometrically independent sub-regions, and extracts the shape features of each sub-region; Image processing module: The image processing module performs pixel-level processing on each sub-region, including adaptive filtering, color space conversion and noise suppression, to optimize the local characteristics of the sub-region; Correlation calculation module: The correlation calculation module obtains the correlation weights between each sub-region, constructs a sub-region correlation map and performs feature propagation, and obtains consistency correction parameters between sub-regions; Sub-region synthesis module: The sub-region synthesis module adjusts the pixels of the sub-region based on the consistency correction parameter, performs smooth transition processing, and finally synthesizes the optimized main image region; Background area adjustment and image synthesis module: The background area adjustment and image synthesis module obtains the optimized main image color histogram according to the background area mask, performs background color temperature migration and adjusts the background blur, and finally synthesizes a complete image; Processor: The processor is used to process the calculation process of each formula and the construction calculation process of each model.

Images