Art design style migration system based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an art design style migration system based on deep learning, which comprises an image perception analysis module, a style constraint calculation module, a detail structure mapping module, a scale corresponding screening module and a region fusion regulation and control module. According to the method, the texture, the structure and the edge features of the image are accurately analyzed, key details and structure areas are identified, so that the flexibility of style migration is improved, style application is more accurate, the problem of excessive stylization or content imbalance is avoided, the spatial features of the content areas are effectively identified by analyzing the detail structure levels of the image, and the image quality is improved. The method comprises the following steps of: selecting and matching an optimal style domain according to a scale corresponding relation, enhancing multi-scale style adaptability, improving expressive force of an image under multi-style conversion, ensuring natural fusion of contents and styles, improving a visual effect and keeping harmony between the styles and original contents, and ensuring the accuracy of style mapping, screening and matching an optimal style domain according to a scale corresponding relation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an art design style transfer system based on deep learning. Background Art

[0002] The field of artificial intelligence (AI) encompasses the ability to emulate, extend, and expand human intelligence, enabling computer systems to possess capabilities such as perception, reasoning, learning, and decision-making. Core areas within this field include data acquisition and representation, feature extraction and modeling, intelligent reasoning and analysis, and model training and optimization. It is widely used in areas such as image recognition, natural language processing, machine translation, and intelligent control. In the field of image processing and generation, AI technology can achieve functions such as image style transfer, image synthesis, and image reconstruction. This relies on training neural network models with large amounts of sample data, building a systematic approach capable of identifying and reproducing image style features.

[0003] Among them, the art design style transfer system refers to an image style transfer method based on deep neural networks. It completes the image style transfer process by extracting feature information such as texture, color, and lines from the target style image and mapping it to the input design image. The system mainly covers technical matters such as image feature extraction, style feature matching, content preservation constraints, and style transfer synthesis. Specifically, it uses the output of the intermediate layer of the convolutional neural network as the basis for style and content features, uses the feature reconstruction loss function to calculate the difference between style and content, and updates the input image through backpropagation until the style fusion goal is achieved. The system uses a pre-trained image recognition network for feature extraction and combines the similarity between the content image and the style image in a multi-level feature space for style reconstruction.

[0004] Existing technologies rely on shallow feature extraction based on convolutional neural networks when processing image style transfer, and lack in-depth analysis of image details and structure. This method has difficulty in accurately grasping the balance between content and style when processing complex images, which can easily lead to loss of details or overly abrupt styles, affecting the naturalness of the image. In particular, when faced with subtle texture changes in images, existing methods lack sufficient scale matching and fine style adjustment, resulting in inconsistent style transfer results. Specifically, the edge processing, texture frequency, and scale contrast of the image are not processed accurately enough, resulting in distorted or overly rigid style expressions in the details. Since existing methods mainly rely on single-level feature matching, they ignore the fusion of styles at different scales, resulting in a single image style expression. This deficiency makes the style-transferred image fail to achieve the expected high-quality conversion effect and cannot give full play to the advantages of mutual mapping between different styles. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the existing technology and propose an art design style transfer system based on deep learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions: an art design style transfer system based on deep learning includes: The image perception and parsing module obtains the target image and style image input files, extracts the image texture distribution frequency and edge gradient change rate, and annotates the stable and variable content areas in the image by comparing the main structural directions and contour continuity of the region, thereby generating an image content sensitivity map. The style constraint calculation module calls the image content sensitivity map, divides the style image into detail blocks, identifies the texture difference rate and morphological boundary offset, selects the area that meets the threshold as the style coverage unit, performs content-style bidirectional mapping comparison, and generates a style mapping association group; The detail structure mapping module extracts the spatial structure hierarchy of the target image content area based on the style map association group, corresponds it to the directional offset of the style image texture group, determines the density change segment, delineates the activity area and the structure reference area, and obtains the image detail structure layer table; The scale correspondence screening module calls the image detail structure hierarchical table, selects the scale range of the target image structure unit, identifies the multi-scale style domain of the style image, compares the overlap ratio of edge consistency and texture frequency difference, screens the best matching style domain, and generates a scale correspondence list.

[0007] As a further solution of the present invention, the image content sensitive map includes the main direction distribution density of the structure, the contour continuity level identification, the regional stability label, and the texture response feature map; the style mapping association group includes the style coverage area number, the texture difference comparison item, the morphological boundary offset record, and the mapping corresponding index table; the image detail structure layered table includes the structure level number, the texture direction mapping relationship, the density change segment label, the structure stability parameter, and the active area positioning mark; the scale correspondence relationship list includes the scale matching range set, the style domain selection list, the texture frequency overlap coefficient, and the edge feature overlap rate index.

[0008] As a further solution of the present invention, the image perception and analysis module includes: The texture mapping extraction submodule obtains the target image and style image input files, extracts the grayscale gradient values and color channel offset values in the corresponding area, compares the average amplitude of the grayscale gradient values in the area with the standard deviation of the color channel offset values, and generates a texture migration tendency value map; The structural continuity assessment submodule identifies the edge gradient change rate and the length of the contour connected segment in the target image area based on the texture migration tendency value map, compares the density distribution of the edge gradient change rate with the variation range of the connected segment length, and obtains the structural compatibility distribution coefficient; The sensitive area calibration submodule extracts the contour closure and direction consistency values in the image based on the structural compatibility distribution coefficient, screens the regional aggregation distribution of the contour closure and direction consistency values, calibrates the areas with strong style response and the areas with stable style matching, and generates an image content sensitive map.

[0009] As a further solution of the present invention, the style constraint calculation module includes: The atlas partitioning submodule calls the image content sensitive atlas, extracts image structure blocks and color clustering areas, extracts edge density and brightness gradient, performs morphological aggregation and block division on local areas, and performs style image structure division to generate a style detail block partition map; The texture recognition submodule calls the style detail block partition map, extracts the partition texture frequency distribution and boundary morphology data, identifies the grayscale change rate and boundary displacement rate, and screens according to the style texture difference threshold and boundary morphology change judgment standard to obtain a texture mapping area set; The mapping generation submodule extracts the target image content features and style image features based on the texture mapping area set, compares the color distribution vector and texture response tensor in the area, calculates the style matching strength value between the areas, and uses the matching strength as the basis to perform feature mapping index combination to generate a style mapping association group.

[0010] As a further solution of the present invention, the detail structure mapping module includes: The structure layer extraction submodule extracts the spatial arrangement matrix and contour trend map of the content area in the target image based on the style map association group, identifies the texture direction offset vector in the style image, compares the distribution changes of the two at the same spatial position, arranges the image area according to the gradient difference, and generates a style guidance space hierarchy map; The active area identification submodule extracts the regional texture density map and texture direction offset map based on the style-guided spatial hierarchy map, determines the local density gradient segment, and screens the high-fluctuation segment area based on the offset amplitude difference. It calculates the structural response distribution index and marks the corresponding position as the style active structure area or the stable structure area to obtain the image detail structure layered table.

[0011] As a further solution of the present invention, the scale corresponding screening module includes: The structure scale calibration submodule calls the image detail structure hierarchical table to extract the hierarchical distribution of high-frequency contour areas and low-frequency flat areas in the target image, divides the structural unit boundaries according to the convolution kernel response intensity gradient value, and generates the structural unit scale interval; The style domain parsing submodule extracts the directional gradient histogram amplitude at different scales based on the structural unit scale interval, detects the texture co-occurrence matrix change rate between adjacent scales, and establishes a multi-scale style domain set; The style matching calculation submodule calls the multi-scale style domain set, calculates the space-frequency domain joint matching coefficient, compares the overlap ratio threshold of each style domain, and generates a scale correspondence list.

[0012] As a further solution of the present invention, the system also includes a regional fusion control module: The regional fusion control module calls the scale correspondence list, counts the color balance, boundary response and style strength of each unit image block during the fusion process, combines the content and style weights for cross-judgment, sets the content retention threshold and style activation factor, and generates an art design content and style fusion layout; The art design content style fusion layout includes a color balance layer, a boundary response distribution map, a style intensity adjustment factor, and a content retention weight setting table.

[0013] As a further solution of the present invention, the regional fusion control module includes: The color balancing submodule extracts the RGB channel values of the image block, calls the scale correspondence list to determine the regional scale level, compares the color distribution of the image block according to the level color reference value, adjusts the channel values that deviate from the reference, and obtains the color distribution weight; The boundary recognition submodule extracts the edge gradient value and grayscale difference of the image block based on the color distribution weight, selects the key boundary area, compares and identifies the response intensity with the gradient difference ratio of the adjacent area, and obtains the boundary continuity measurement map; The style fusion submodule calls the boundary continuity measurement map and the style strength value and content priority value of each image block, combines the content retention threshold and the style activation factor, assigns styles to the image blocks according to the regional scale, and generates an art design content style fusion layout.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, by accurately analyzing the texture, structure, and edge feature information of the image, a more detailed mapping between content and style is achieved. By accurately marking the stable content area and the volatile content area, key details and structural areas can be identified in the design image, thereby improving the flexibility of style transfer. The in-depth analysis of the sensitivity of image content makes the style application process more accurate and avoids the problem of over-stylization or imbalance between style and content. By analyzing the structural level of image details, the spatial structural characteristics of the content area can be effectively identified and corresponded to the texture differences of the style image, thereby ensuring the accuracy of details in the style mapping process and avoiding the ambiguity or inaccuracy of the original style features. In the processing process, the optimal style domain is screened and matched according to the scale correspondence, which enhances the adaptability of multi-scale styles, and can better balance the edge consistency and texture accuracy, improve the expressiveness of the image under multiple style conversions, and maintain the natural fusion between content and style. It not only improves the visual effect after migration, but also ensures the harmony between style changes and the original content of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a system flow chart of the present invention; Figure 2 This is a flow chart of the image perception and analysis module in the present invention; Figure 3 This is a flow chart of the style constraint calculation module in the present invention; Figure 4 This is a flowchart of the detail structure mapping module in the present invention; Figure 5 This is a flow chart of the scale correspondence screening module of the present invention; Figure 6 This is a flow chart of the regional fusion regulation module in the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0017] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0018] See also Figure 1 , the art design style transfer system based on deep learning includes: The image perception and parsing module obtains the target image and style image input files, extracts the image texture distribution frequency and edge gradient change rate, and annotates the stable and variable content areas in the image by comparing the main structural directions and contour continuity of the region, thereby generating an image content sensitivity map. The style constraint calculation module calls the image content-sensitive atlas to divide the style image into detail blocks, identify texture difference rates and morphological boundary offsets, select areas that meet the threshold as style coverage units, perform content-style bidirectional mapping comparison, and generate style mapping association groups; The detail structure mapping module extracts the spatial structure hierarchy of the target image content area based on the style map association group, corresponds it to the directional offset of the style image texture group, determines the density change segment, delineates the activity area and the structural reference area, and obtains the image detail structure layer table; The scale correspondence screening module calls the image detail structure hierarchical table, selects the scale range of the target image structure unit, identifies the multi-scale style domain of the style image, compares the overlap ratio of edge consistency and texture frequency difference, selects the best matching style domain, and generates a scale correspondence list; The regional fusion control module calls the scale correspondence list, and during the fusion process, it counts the color balance, boundary response, and style intensity of each unit image block. It combines the cross-judgment of content and style weights, sets the content retention threshold and style activation factor, and generates an art design content-style fusion layout.

[0019] The image content sensitive map includes the main direction distribution density of the structure, the contour continuity level identification, the regional stability label, and the texture response feature map. The style mapping association group includes the style coverage area number, the texture difference comparison item, the morphological boundary offset record, and the mapping corresponding index table. The image detail structure layered table includes the structure level number, the texture direction mapping relationship, the density change segment label, the structure stability parameter, and the active area positioning mark. The scale correspondence list includes the scale matching range set, the style domain selection list, the texture frequency overlap coefficient, and the edge feature overlap rate index. The art design content style fusion layout includes the color balance layer, the boundary response distribution map, the style intensity adjustment factor, and the content retention weight setting table.

[0020] See also Figure 2 , the image perception and parsing module includes: The texture mapping extraction submodule obtains the target image and style image input files, extracts the grayscale gradient values and color channel offset values in the corresponding area, compares the average amplitude of the grayscale gradient values in the area with the standard deviation of the color channel offset values, and generates a texture migration tendency value map; When extracting corresponding regions between the target image and the style image, we first need to identify key feature regions within the image and calculate the grayscale gradient and color channel offset values for these regions. To extract the grayscale gradient, we need to calculate the grayscale difference for each pixel within the region. Edge detection is performed using the Laplacian or Sobel operator. This yields the gradient value for the region, indicating the degree of change. The color channel offset calculates the deviation in RGB values between the target image and the style image in color space. This offset can be obtained by comparing the difference in RGB channel values for each pixel in the image. For example, if the RGB value of a pixel in the target image is (100, 150, 200), while the RGB value of the same position in the style image is (110, 145, 195), the color channel offset value is (10, -5, -5). After obtaining the grayscale gradient value and color channel offset value, they need to be analyzed and processed to calculate the average amplitude of the grayscale gradient value and the standard deviation of the color channel offset value in the area. The average amplitude of the grayscale gradient value can be obtained by calculating the average absolute change in the grayscale values of all pixels in the area. For example, if the calculation is performed on a certain area of the target image, the average amplitude is 2. The standard deviation of the color channel offset value is obtained by calculating the standard deviation of the RGB offset values of all pixels. If the average grayscale gradient value of the target area is high and the standard deviation of the color offset value is large, it means that the texture migration tendency in this area is strong, otherwise it is weak. Finally, a texture migration tendency value map is obtained.

[0021] The structural continuity assessment submodule identifies the edge gradient change rate and the length of the contour connected segment within the target image region based on the texture migration tendency value map, compares the density distribution of the edge gradient change rate with the variation range of the connected segment length, and obtains the structural compatibility distribution coefficient; Based on the generated texture migration tendency value map, the structural compatibility of the target image region is assessed by calculating the edge gradient change rate and the length of the contour connected segments within the target image region. The edge gradient change rate is calculated by analyzing the gradient changes at the image edge and using gradient calculation methods such as Sobel or Canny edge detection to obtain the gradient map of the region. The rate of grayscale change at the edge is then calculated. For example, if the gradient at the edge of a horizontal region increases from 0.5 to 1.5, the change rate is (1.5-0.5) / (position change). Assuming the position change is 1, the change rate is 1. The contour connected segment length is calculated by analyzing the connected points on the image contour, finding the set of continuously connected points in the contour and calculating its length. This can be achieved using the connected component labeling algorithm. For the edges and contours in the target image, the evaluation process requires calculating the density distribution of the edge gradient change rate and comparing the change range of the connected segment length. If the edge gradient change rate of a certain area in the image is high and the length of the connected segment is short, it indicates that the area has good structural compatibility. Through the evaluation value, the structural compatibility distribution coefficient can be finally obtained, which can be used to measure the structural matching degree between the area and the style image.

[0022] The sensitive area calibration submodule extracts the contour closure and directional consistency values in the image based on the structural compatibility distribution coefficient, screens the regional aggregate distribution of contour closure and directional consistency values, calibrates areas with strong style response and stable style matching, and generates an image content sensitive map; The closure and directional consistency values of the image's contours are extracted and combined with indicators for screening, ultimately identifying regions with strong style responses and stable style matching. The closure of the image's contours is calculated. Contours with high closure indicate relatively complete object edges within the region, which can provide better texture transfer results. Closure is calculated based on edge connectivity, determined by analyzing closed paths on the contours. Directional consistency is assessed by measuring the directional changes of contour segments within the region. Small directional changes indicate high consistency. By analyzing these indicators, regions with strong style responses and stable style matching can be identified. For example, if the contour closure of the target region is 0.85 and the directional consistency value is 0.9, it indicates that the region exhibits strong responsiveness during style transfer and stable style matching. Therefore, it can be labeled as a region with strong style responses. A content-sensitive map of the image is generated, which contains information on the strength and stability of the style response of each region, providing an important reference for further texture transfer.

[0023] See also Figure 3 , the style constraint calculation module includes: The atlas partitioning submodule calls the image content-sensitive atlas to extract image structural blocks and color clustering areas, extract edge density and brightness gradients, perform morphological aggregation and block division on local areas, and perform style image structure division to generate a style detail block partition map. Structural blocks and color clustering areas are obtained from the style image, and edge density and brightness gradient are extracted as basic parameters. The division of image structural blocks can be extracted based on the contour boundary and color mean difference in the image. For example, in a work of art, there is a clear boundary between the building edge and the sky background. The edge gradient of the area can be extracted as the structural boundary benchmark, and then the edge density is calculated. The edge density can be defined as the proportion of pixels with pixel gradient amplitude greater than the threshold within a unit area. The threshold is set to 30. If the number of pixels greater than 30 in the unit block area exceeds 40% of the total number of pixels, it is determined to be a high edge density area. The brightness gradient can be statistically analyzed by the grayscale difference of each pixel. The Sobel operation is used to obtain the gradient in each direction, and the brightness gradient map is accumulated. The brightness gradient interval value is determined by the mean and variance, and the image area with a brightness gradient greater than 10 is extracted as the object of further analysis. In actual works, for example, the facial area and clothing area of the character in the image have relatively clear brightness levels, and the effectiveness of their boundary division can be judged by the brightness gradient. After extracting the structural blocks and the brightness difference areas, the image is aggregated in local areas. Using spatial continuity as the criterion, adjacent areas with similar brightness and structural distribution are divided into unified blocks. The spatial neighborhood similarity threshold is set to the Euclidean distance of the color vector less than 12, that is, if the Euclidean distance between the three-dimensional color vectors of two blocks is less than 12, they are regarded as the same structural unit. Then the block operation is performed to draw the boundary line and determine the geometric center and bounding box of each image block. The bounding boxes are numbered and their color mean and brightness mean are recorded for subsequent comparison and judgment. Each image block obtained through the above process constitutes a structural partition map, which serves as the basic data for subsequent processing in style transfer to generate a style detail block partition map.

[0024] The texture recognition submodule calls the style detail block partition map, extracts the partition texture frequency distribution and boundary morphology data, identifies the grayscale change rate and boundary displacement rate, and screens according to the style texture difference threshold and boundary morphology change judgment criteria to obtain the texture mapping area set; Each partitioned area is taken as the basic analysis object, and its corresponding texture frequency distribution and boundary morphology data are extracted. The extraction of texture frequency distribution is based on the grayscale co-occurrence matrix. By counting the spatial distribution of grayscale value pairs in the area, multiple texture feature items such as contrast, entropy, and energy are calculated. For example, for a cloth or stone texture area in a style painting, the contrast is large and the entropy value is obvious, and the texture frequency distribution vector is extracted. The morphological boundary data is measured based on the change of contour curvature and the continuity of the boundary. The polygonal approximation boundary is fitted and the curvature value at each point of the boundary is calculated. The curvature change interval and mean of the area are then counted. During the analysis process, the grayscale change rate is calculated using the mean of the grayscale difference between adjacent pixels. For example, if the pixel grayscales in the area are {90, 95, 102, 108}, and the grayscale change rate is (5+7+6) / 3=6, if the value is greater than the set threshold 5, it is judged to be an area with drastic texture changes. For the boundary displacement rate, the curvature changes of the boundary points in different directions are extracted and analyzed. The difference between adjacent curvature changes is averaged. If the value is greater than 0.3, it is judged as a boundary morphology fluctuation area. The texture difference threshold used in the screening process is set to 1.5 times the standard deviation of the texture energy item. Based on the difference in texture features between different areas, in the current image, the texture energy of area A is 0.21, and that of area B is 0.28. The standard deviation is 0.03, 1.5 times is 0.045, and the difference between the two is 0.07, which is judged as a valid texture difference area. The image block numbers screened by the above judgment constitute a region set that can be used for subsequent mapping, and the depth texture mapping region set is obtained.

[0025] The mapping generation submodule extracts the target image content features and style image features based on the texture mapping region set, compares the color distribution vector and texture response tensor in the region, and uses the formula: ; Calculate the style matching strength value between regions, use the matching strength as the basis to perform feature map index combination, and generate style map association group; in, Indicates the content of the figure color distribution items, Style diagram color distribution items, For the Style map texture response items, For the regional texture gradient terms, is the number of color distribution dimensions, is the number of texture feature dimensions, is the inter-region style matching strength value; The inter-region style matching strength value is a metric used to quantify the overall similarity between the target image content region and the corresponding region in the reference style image in terms of color distribution and texture response. The higher the value, the stronger the coupling between the region and the target style region in terms of color vector structure and texture space characteristics, making it suitable as the main control mapping region in the style transfer operation. This strength value comprehensively considers the degree of difference in color histograms and the combined influence of multi-scale and multi-directional texture response and local gradient changes. Through the weighted integration of color and texture dual channels, it reflects whether the local image region has the ability to express typical style features, thereby providing a computational basis for the subsequent construction of the content style mapping matrix and the generation of the transfer fusion feature map. First, the content area of the target image is calibrated and located, and the color distribution vector and texture response tensor of this area are extracted. The color distribution vector is constructed based on the RGB three-channel histogram. Each channel is divided into 16 intervals, resulting in a total of 48-dimensional vectors, where each dimension represents the proportion of pixels in a specific grayscale interval under that channel. For example, for a certain style map area, the R channel has a pixel proportion of 0.08 in the first interval (0–15) and 0.10 in the second interval. The G channel has a pixel proportion of 0.07 in the first interval and 0.09 in the second interval. The B channel has a pixel proportion of 0.06 and 0.11, respectively. The vector is obtained as follows: , the target content image area uses the same method to obtain the vector To ensure the comparability between vectors, all histogram vectors need to be normalized. The processing method is to divide the number of each pixel by the total number of pixels in the channel, and normalize the three channels separately to obtain the proportional vector within the unit interval, eliminating the influence of resolution on the result. The extraction of texture response tensor is based on Gabor filtering. Four directions (0°, 45°, 90°, 135°) and four scales (σ=1, 2, 3, 4) are set, with a total of 16 groups of filters. Convolution operation is performed on each image area to obtain a 16-dimensional response tensor. , each item represents the texture energy response at the corresponding scale and direction, the unit is the normalized filter intensity value, and the range is normalized to [0, 1]. At the same time, the gradient enhancement factor of the corresponding area is extracted from the image gradient map , defined as the ratio of the mean standard deviation of the gradient direction intensity map, the unit is a dimensionless ratio, through normalization to adapt to the texture tensor matching, the absolute value of the horizontal and vertical gradients of each pixel in the region is averaged to obtain the regional gradient intensity, and then normalized to the [0, 1] interval, expressed as the gradient enhancement response; After all parameters are prepared, perform regional style matching strength calculation; Using the formula: ; Assume that the content graph color distribution vector part is , the color distribution vector part of the style map is , the sum of the absolute differences between the two is set to ; The texture response tensor and gradient enhancement factor terms are , , the sum of the products is: ; Substituting into the formula: ; Final Result This indicates that the matching strength of the current area in terms of color and texture is low. If the system sets the mapping baseline value to 0.08, the area is not suitable for style feature mapping. On the contrary, if it is greater than 0.08, it enters the index matching stage; Indicates the target image Normalized color histogram components, Represents the style image Normalized color histogram components, Style diagram Gabor filter response values in directions and scales, Indicates that the target image is Gradient enhancement factors in directions and scales, 64 is the sum of color dimension and texture dimension, used as the denominator for averaging, The regional style matching strength value is expressed by combining color difference and texture structure to express the style representativeness between regions. It no longer relies on direct matching of local feature points, thus expanding the scope of application of style transfer at the regional level. The result shows that the current region has low style representativeness and is suitable for weak mapping or content feature preservation processing.

[0026] See also Figure 4 , the detail structure mapping module includes: The structure layer extraction submodule extracts the spatial arrangement matrix and contour trend map of the content area in the target image based on the style map association group, identifies the texture direction offset vector in the style image, compares the distribution changes of the two at the same spatial location, arranges the image areas according to the gradient difference, and generates a style-guided spatial hierarchy map. Segment the content area corresponding to the style image from the target image, set the area to a 256×256 pixel rectangular block, extract the two-dimensional position coordinates of each pixel point, form a spatial arrangement matrix, and record the boundary direction of the area. Obtain the displacement sequence of the contour line in the image coordinate system through the contour extraction operation, and then construct a local edge direction map based on the grayscale gradient direction between adjacent pixels to form a contour direction map. Obtain the texture offset vector of the style image in the same position mapping area. The offset vector is obtained based on the spatial distribution of the directional difference between pixels in the style texture direction response map. Map all style texture directional offset values to the coordinate system corresponding to the target image structure to complete the position alignment and compare the features of the two images at each spatial position. Distribution changes, perform spatial gradient difference analysis, that is, calculate the gradient change rate of the content image in the direction of adjacent pixels, and compare it with the change amplitude of the offset angle of the corresponding area in the style image. For example, the horizontal gradient of the (i, j) position in the content image is 14, the texture direction of the position in the style image is 60°, and the adjacent one is 75°, then the direction difference is 15°. After normalizing it to the interval [0, 1], participate in the gradient difference comparison, and mark each image block as a different level according to the gradient difference. If the change amplitude exceeds the gradient stratification benchmark value (set to the normalization value 0.25), it is classified as a high-level unit, otherwise it is classified as a low-level unit. After the above analysis and judgment, the spatial structure hierarchy result corresponding to the influence of the style structure is established, and the style-guided spatial hierarchy map is obtained.

[0027] The active region identification submodule extracts the regional texture density map and texture direction offset map based on the style-guided spatial hierarchy map, determines the local density gradient segment, and filters the high-fluctuation segment area based on the offset amplitude difference, using the formula: ; Calculate the structural response distribution index, mark the corresponding position as the style active structure area or stable structure area, and obtain the image detail structure layered table; in, is the structural response distribution index, For participating areas The texture density value, For participating areas The spatial level gradient value of For participating areas The texture direction change rate, For participating areas The texture direction jump amplitude, is the total number of participating regions; The structural response distribution index is a comprehensive indicator used to measure the degree of structural activity exhibited by local regions in an image during the style transfer process. Its value reflects the strength of the superimposed features in the region in terms of texture density changes, spatial level gradient differences, and texture directional mutations. The larger the index, the easier it is for the region to express the changing trend of style characteristics in terms of spatial organization, and the stronger the style superposition potential. During the calculation process, the index comprehensively considers the degree of difference in regional texture density and spatial level, and constructs the overall structural response level through a weighted fusion of directional change rate and mutation amplitude. Ultimately, it serves as a criterion for determining whether a region belongs to a style-active structure area, providing a structural stratification basis for subsequent style transfer mapping. First, spatial structure fragments are extracted from the image according to the region index number. Each fragment is a 64×64 pixel region. The texture response amplitude and gradient level in each region are counted, and the texture response amplitude of each pixel in the region is extracted. The standard texture convolution kernel response matrix is used for pixel-by-pixel convolution. The response amplitude is normalized to the range of 0 to 1, and the response threshold is set to 0.6. The ratio of pixels with a response value greater than 0.6 in the region to the total number of pixels is calculated, which is defined as the texture density value of the region. , if area A has 4096 pixels, of which 1720 pixels have a response value exceeding 0.6, then ; Spatial level gradient value The acquisition is based on the comparison of the average brightness and structural gradient of adjacent regions. The horizontal gradient and vertical gradient of each pixel are synthesized, and their L2 norm is calculated and normalized. The mean difference between adjacent regions is taken as the hierarchical gradient. For example, the mean values of the synthesized gradients between regions 1 and 2 are 0.56 and 0.49 respectively. After normalization, ; Texture direction change rate The variation of the texture direction angle of each pixel is derived from the standard deviation of the direction angle of all pixels in the region and normalized. For example, the standard deviation of the texture direction in region 1 is 18°, and the maximum allowable angle variation is 90°. ; Texture direction jump amplitude Indicates the maximum directional mutation in the area, which is obtained by taking the maximum directional angle change between adjacent pixels and normalizing it. If the maximum mutation in the area is 27°, then ; Set the above parameters as follows: Total number of regions ; ; ; ; ; Substituting into the formula: ; The calculation steps are as follows: , ; , ; , ; , ; The numerator totals: ; The denominator is: ; Final result: ; Structural Response Distribution Index The baseline judgment threshold of 0.45 has been exceeded, so the current fragment area should be identified as a style-active structure area and participate in the subsequent style fusion area construction. By simultaneously introducing the fusion calculation method of texture density difference and directional response jump, the recognition offset caused by relying solely on a single gradient indicator is avoided. This result shows that the calculated area has reasonable structural characteristics of superimposed style features in terms of spatial hierarchy and texture features.

[0028] See also Figure 5 , the scale corresponding screening module includes: The structural scale calibration submodule calls the image detail structure hierarchical table to extract the hierarchical distribution of high-frequency contour areas and low-frequency flat areas in the target image, divides the structural unit boundaries according to the convolution kernel response intensity gradient value, and generates the structural unit scale interval; The structure in the image is segmented at multiple levels to obtain a hierarchical distribution of high-frequency contours and low-frequency regions. This process can be achieved by convolving the image with a specific set of convolution kernels. The convolution kernels respond to edge information in the image, generating high-frequency signals. By calculating the intensity gradient of the response values, the boundaries of the structural units can be demarcated, and the scale intervals of the structural units can be determined. For example, when processing architectural images, the intensity of the convolution kernel response is stronger in edge areas (such as window outlines) and weaker in flat areas (such as smooth parts of walls). The gradient of the response value is used to demarcate specific scale intervals. For example, if the convolution kernel response intensity is 0.8 in the window area and 0.2 in the wall area, and the boundary is set to 0.5, this will help distinguish the building's structural units and determine the scale difference between windows and walls. Based on the response, the scale intervals of the structural units can be generated, ensuring that the structural information in the image can be accurately segmented.

[0029] The style domain parsing submodule extracts the directional gradient histogram amplitude at different scales based on the structural unit scale interval, detects the texture co-occurrence matrix change rate between adjacent scales, and establishes a multi-scale style domain set; The style image is decomposed in the frequency domain to extract histograms of directional gradients at different scales. The image is then converted to the frequency domain via Fourier transform, allowing analysis of texture features within the image. For example, for an image of a carpet with complex texture, frequency domain decomposition is first used to separate high- and low-frequency components. A histogram of directional gradients is then calculated for each component. This histogram reflects the directional variations and frequency distribution of the texture within the image. While processing high-frequency components can capture subtle texture variations within the carpet, low-frequency components better capture the carpet's overall texture pattern. This method analyzes the rate of change of the texture co-occurrence matrix within the image, deriving texture patterns at various scales. This results in a multi-scale style domain containing 32 sets of texture patterns. Each set of texture patterns represents a specific texture feature, such as edge direction, texture smoothness, or noise distribution. These patterns help to analyze the detailed features of the style image at a higher level, providing fundamental data for style matching calculations.

[0030] The style matching calculation submodule calls the multi-scale style domain set and uses the formula: ; Calculate the spatial-frequency domain joint matching coefficient, compare the overlap ratio threshold of each style domain, and generate a scale correspondence list; in, Represents the edge direction continuity strength, is the phase difference of the gradient direction histogram, represents the frequency difference of texture across scales, is the spatial distribution discreteness, is the directional gradient consistency factor, Represents the topological structure similarity, represents the spatial-frequency domain joint matching coefficient; The spatial-frequency joint matching coefficient (JMC) is a comprehensive measure of the similarity between image structure and texture style. It combines image feature information in both the spatial and frequency domains and is calculated using a variety of image feature parameters (such as edge direction strength, gradient direction histogram, texture frequency difference, spatial distribution dispersion, and directional gradient consistency). Information in the spatial domain focuses on structural features of the image, such as edges and shape, while information in the frequency domain focuses on texture features, such as texture detail and directional variations. By combining these two aspects, the JMC provides a more comprehensive description of the overall style and structural consistency of an image, thereby providing a more accurate assessment of style matching. Specifically, the coefficient takes into account the symbiotic relationship and spatial distribution differences of textures at different scales. In the calculation, different features are weighted according to their contribution to produce a value that reflects the global degree of image matching. A higher JMC indicates a high degree of consistency between the image's structure and style, while a lower JMC indicates a significant difference. First, it is necessary to extract relevant parameters from the structural unit scale interval and the multi-scale style domain set and perform a joint calculation based on multiple image feature parameters, including edge direction strength (D), gradient direction histogram phase difference (Φ), cross-scale texture frequency difference (H), spatial distribution dispersion (Ψ), directional gradient consistency factor (Ω), and topological structure similarity (Δ). Each parameter has its own specific calculation method and needs to be obtained through analysis and extraction of image content. The edge direction strength D is calculated by responding to the edge features in the image through convolution operation. The response value of the convolution kernel is larger in the edge area of the image, reflecting the continuity strength of the edge. The gradient direction histogram phase difference Φ is calculated by analyzing the gradient direction of adjacent pixels in the image, reflecting the change in the edge direction. To ensure that the dimensions of each parameter are consistent, all calculated features need to be normalized. Specifically, parameters such as edge direction intensity D, gradient direction histogram phase difference Φ, and cross-scale texture frequency difference H need to be normalized within the interval [0, 1] for weighted calculation and to ensure comparison at the same calculation scale. For example, assuming that the maximum value of D is 0.9 and the minimum value is 0.1, then D in the actual calculation can be normalized using the following formula: ; If D is 0.6, the normalized D is: ; Similarly, the normalization process of the gradient direction histogram phase difference Φ is similar. If the range of Φ is [0, 1] and the value of Φ is 0.8, then this parameter does not need further processing, but this normalized value can still be used when calculating the parameter; Next, we calculate the cross-scale texture frequency difference H. The difference is obtained by comparing the texture frequencies of different scales. The calculation formula of H is: ; in, and Represents the texture frequency of high-frequency and low-frequency areas respectively. For an image, if the texture frequency in the high-frequency area is 0.7 and the texture frequency in the low-frequency area is 0.3, the value of H is 0.4; The spatial distribution discreteness Ψ reflects the spatial distribution difference of the image texture. It is determined by calculating the discrete degree of the texture in each area of the image. For an image, if its texture is more evenly distributed in space, the value of Ψ is smaller, otherwise it is larger. Assuming that the texture of the image is more evenly distributed in space, the value of Ψ is 0.2. The directional gradient consistency factor Ω is a measure of the directional consistency in the image. It can be obtained by calculating the consistency of the directional gradients of each region in the image. Assuming that the Ω value is 0.8, it means that the texture direction in the image has little change. The topological similarity Δ is calculated by analyzing the topological structure in the image and comparing the structural similarities between different regions. Assuming the value of Δ is 0.5, it means that the topological structures of different regions in the image have a certain degree of similarity; After obtaining all the parameters, they can be substituted into the calculation formula of style matching; Substitute the specific values for calculation: D=0.625, Φ=0.8, H=0.4, Ψ=0.2, Ω=0.8, Δ=0.5; First calculate the numerator: ; ; Molecular part: ; Then calculate the denominator: ; The final calculated Q value is: ; The results show that the calculated matching coefficient Q is 2.6237, indicating that the degree of matching between the various structures in the image and the style domain is high. The results show that the matching relationship between the style and structural units of the image is relatively close, which can provide an important basis for subsequent image processing tasks.

[0031] See also Figure 6 , the regional integration control module includes: The color balancing submodule extracts the RGB channel values of the image block, calls the scale correspondence list to determine the regional scale level, compares the color distribution of the image block according to the level color benchmark value, adjusts the channel values that deviate from the benchmark, and obtains the color distribution weight; The RGB channel values of the input image block are extracted. This means that for each pixel in the image, the values of its red (R), green (G), and blue (B) channels are obtained. The values are expressed as integers from 0 to 255. In this way, each pixel in the image block has a set of RGB values, which constitute the color information of the image block. The scale correspondence list is called to determine the regional scale level of the image block. The regional scale level is determined by the size of the image block. For example, if an image block is 5x5 pixels, the scale of the block can be mapped to a category level in the list. The level is used during the processing to determine the position and importance of the image block in the overall image. Based on this level, a corresponding reference color is selected from the color reference value table for comparison. During execution, the RGB channel values of the image block are compared with the reference color. The color difference is calculated to determine the deviation of the color distribution of the image block. If the channel value deviates from the reference color and is higher or lower, the channel value is automatically adjusted to make it closer to the reference value. This is achieved through a specific adjustment algorithm, such as adjusting the RGB value to return it to the set color standard. The adjusted color distribution weight reflects the consistency of the image block with the reference color, providing a more consistent color foundation for subsequent image processing. Suppose an image block has RGB values of (200, 150, 100) and is compared with the reference color (180, 160, 120). After calculating the color difference, linear interpolation or proportional adjustment can be used to reduce the deviation of the RGB channels to ensure a more balanced color distribution of the image block. The color distribution weight is obtained, which reflects the adjustment amplitude and direction of the image block color and provides basic data for further image processing.

[0032] The boundary recognition submodule extracts the edge gradient value and grayscale difference of the image block based on the color distribution weight, screens the key boundary areas, and compares the response intensity with the gradient difference ratio of the adjacent areas to obtain the boundary continuity measurement map; Based on the obtained color distribution weights, edge detection is performed. The difference between the value of each pixel in the image and its surrounding pixels determines its edge strength. A common method for edge detection is to calculate the image's gradient. For example, using the Sobel operator to calculate the gradient of an image block along the x- and y-axes, the gradient magnitude and direction of each pixel are obtained. After edge detection, the image's grayscale difference can be combined to further determine the edge strength and boundary clarity. Specifically, the grayscale difference refers to the difference in grayscale value between a pixel in an image block and its neighboring pixels. The greater the difference, the stronger the presence of a boundary. Regions in an image block with significant edge changes are identified as boundaries. Boundary regions with high response strength are screened and compared with adjacent regions by comparing their gradient difference ratios to identify which boundaries are continuous. A boundary continuity metric map compares the gradient differences and response strength ratios of different regions to determine which parts of the image are continuous boundary lines and which are gaps or blurred areas. For example, assuming that the gradient values of two adjacent image blocks A and B are (5, 6) and (4, 7) respectively, and the grayscale difference is 20 and 30, according to the predetermined ratio standard, mark whether there is continuity between the boundary of image blocks A and B, and finally generate a boundary continuity measurement map for subsequent processing and style fusion.

[0033] The style fusion submodule uses the boundary continuity metric map, the style strength value of each image block, and the content priority value, combined with the content retention threshold and style activation factor, to assign styles to image blocks at the regional scale and generate an art design content style fusion layout; The boundary areas of image blocks are evaluated based on the boundary continuity metric map, which provides a reference for the key areas of style fusion. The style strength value of each image block is combined with its content priority value for processing. The style strength value represents the strength of the image block's expression under the current style. For example, under a certain artistic style (such as Impressionism or cartoon style), the image block will have different style strengths. The content priority value is set according to the content importance or regional division of the image block. For example, some image blocks contain more details and important elements, so they have a higher priority. Combining the content retention threshold and the style activation factor, the image blocks are assigned styles at the regional scale. The content retention threshold determines which content must be retained and which can be replaced with the artistic style during style transfer. The style activation factor determines the degree of application of the style strength. The higher the activation factor, the stronger the style effect. For example, if the style strength value of an image block is 0.8, the content priority is 0.6, the style activation factor is 0.5, and the content retention threshold is 0.7, the system will adjust the appearance of the image block based on the style strength and regional priority while ensuring content retention, generating an art design content-style fusion layout to ensure consistency between the style and content of important areas in the image while maintaining the overall fusion of artistic styles.

[0034] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The art design style transfer system based on deep learning is characterized by: The system comprises: The image perception and parsing module obtains the target image and style image input files, extracts the image texture distribution frequency and edge gradient change rate, and annotates the stable and variable content areas in the image by comparing the main structural directions and contour continuity of the region, thereby generating an image content sensitivity map. The style constraint calculation module calls the image content sensitivity map, divides the style image into detail blocks, identifies the texture difference rate and morphological boundary offset, selects the area that meets the threshold as the style coverage unit, performs content-style bidirectional mapping comparison, and generates a style mapping association group; The detail structure mapping module extracts the spatial structure hierarchy of the target image content area based on the style map association group, corresponds it to the directional offset of the style image texture group, determines the density change segment, delineates the activity area and the structure reference area, and obtains the image detail structure layer table; The scale correspondence screening module calls the image detail structure hierarchical table, selects the scale range of the target image structure unit, identifies the multi-scale style domain of the style image, compares the overlap ratio of edge consistency and texture frequency difference, screens the best matching style domain, and generates a scale correspondence list.

2. The art design style transfer system based on deep learning according to claim 1, characterized in that: The image content sensitive map includes the main direction distribution density of the structure, the contour continuity level identifier, the regional stability label, and the texture response feature map; the style mapping association group includes the style coverage area number, the texture difference comparison item, the morphological boundary offset record, and the mapping corresponding index table; the image detail structure layered table includes the structure level number, the texture direction mapping relationship, the density change segment label, the structure stability parameter, and the active area positioning mark; the scale correspondence list includes the scale matching range set, the style domain selection list, the texture frequency overlap coefficient, and the edge feature overlap rate index.

3. The art design style transfer system based on deep learning according to claim 1, characterized in that: The image perception and analysis module includes: The texture mapping extraction submodule obtains the target image and style image input files, extracts the grayscale gradient values and color channel offset values in the corresponding area, compares the average amplitude of the grayscale gradient values in the area with the standard deviation of the color channel offset values, and generates a texture migration tendency value map; The structural continuity assessment submodule identifies the edge gradient change rate and the length of the contour connected segment in the target image area based on the texture migration tendency value map, compares the density distribution of the edge gradient change rate with the variation range of the connected segment length, and obtains the structural compatibility distribution coefficient; The sensitive area calibration submodule extracts the contour closure and direction consistency values in the image based on the structural compatibility distribution coefficient, screens the regional aggregation distribution of the contour closure and direction consistency values, calibrates the areas with strong style response and the areas with stable style matching, and generates an image content sensitive map.

4. The art design style transfer system based on deep learning according to claim 3 is characterized in that: The style constraint calculation module includes: The atlas partitioning submodule calls the image content sensitive atlas, extracts image structure blocks and color clustering areas, extracts edge density and brightness gradient, performs morphological aggregation and block division on local areas, and performs style image structure division to generate a style detail block partition map; The texture recognition submodule calls the style detail block partition map, extracts the partition texture frequency distribution and boundary morphology data, identifies the grayscale change rate and boundary displacement rate, and screens according to the style texture difference threshold and boundary morphology change judgment standard to obtain a texture mapping area set; The mapping generation submodule extracts the target image content features and style image features based on the texture mapping area set, compares the color distribution vector and texture response tensor in the area, calculates the style matching strength value between the areas, and uses the matching strength as the basis to perform feature mapping index combination to generate a style mapping association group.

5. The art design style transfer system based on deep learning according to claim 4 is characterized in that: The detail structure mapping module includes: The structure layer extraction submodule extracts the spatial arrangement matrix and contour trend map of the content area in the target image based on the style map association group, identifies the texture direction offset vector in the style image, compares the distribution changes of the two at the same spatial position, arranges the image area according to the gradient difference, and generates a style guidance space hierarchy map; The active area identification submodule extracts the regional texture density map and texture direction offset map based on the style-guided spatial hierarchy map, determines the local density gradient segment, and screens the high-fluctuation segment area based on the offset amplitude difference. It calculates the structural response distribution index and marks the corresponding position as the style active structure area or the stable structure area to obtain the image detail structure layered table.

6. The art design style transfer system based on deep learning according to claim 5, characterized in that: The scale corresponding screening module includes: The structure scale calibration submodule calls the image detail structure hierarchical table to extract the hierarchical distribution of high-frequency contour areas and low-frequency flat areas in the target image, divides the structural unit boundaries according to the convolution kernel response intensity gradient value, and generates the structural unit scale interval; The style domain parsing submodule extracts the directional gradient histogram amplitude at different scales based on the structural unit scale interval, detects the texture co-occurrence matrix change rate between adjacent scales, and establishes a multi-scale style domain set; The style matching calculation submodule calls the multi-scale style domain set, calculates the space-frequency domain joint matching coefficient, compares the overlap ratio threshold of each style domain, and generates a scale correspondence list.

7. The art design style transfer system based on deep learning according to claim 1, characterized in that: The system also includes a regional fusion control module: The regional fusion control module calls the scale correspondence list, counts the color balance, boundary response and style strength of each unit image block during the fusion process, combines the content and style weights for cross-judgment, sets the content retention threshold and style activation factor, and generates an art design content and style fusion layout; The art design content style fusion layout includes a color balance layer, a boundary response distribution map, a style intensity adjustment factor, and a content retention weight setting table.

8. The art design style transfer system based on deep learning according to claim 7, characterized in that: The regional fusion control module includes: The color balancing submodule extracts the RGB channel values of the image block, calls the scale correspondence list to determine the regional scale level, compares the color distribution of the image block according to the level color reference value, adjusts the channel values that deviate from the reference, and obtains the color distribution weight; The boundary recognition submodule extracts the edge gradient value and grayscale difference of the image block based on the color distribution weight, selects the key boundary area, compares and identifies the response intensity with the gradient difference ratio of the adjacent area, and obtains the boundary continuity measurement map; The style fusion submodule calls the boundary continuity measurement map and the style strength value and content priority value of each image block, combines the content retention threshold and the style activation factor, assigns styles to the image blocks according to the regional scale, and generates an art design content style fusion layout.

Images