Building decoration image analysis method and system

Abstract

The invention relates to the technical field of image analysis, and discloses a building decoration image analysis method and system, and the method comprises the steps: carrying out the visual analysis processing of an inputted building decoration image, and extracting and quantifying an aesthetic feature vector from the overall visual performance; then, constructing an associated database of aesthetic feature vectors and decoration element labels, and performing neighbor matching in the database according to the extracted aesthetic feature vectors to obtain the closest aesthetic feature vectors of the professional gallery images; and finally, inversely solving a decoration element combination rule according to a neighbor matching result, and generating a corresponding decoration element combination suggestion. According to the method, limitation of a traditional method on discrete element identification and co-occurrence frequency statistics is broken through through deeper aesthetic feature analysis, so that redefinition of a relationship between elements in design style evolution and change of an overall atmosphere created by the redefinition are better captured; and a more prospective and personalized design reference is provided for a designer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image analysis, in particular to a building decoration image analysis method and system. BACKGROUND

[0002] In the field of interior design, auxiliary systems analyze a large number of building decoration pictures to extract the combination rules of decoration elements, providing design reference for designers. Such systems usually collect and process professional interior design pictures, use image recognition technology to convert visual information into data, and then count the co-occurrence frequency of different decoration elements to summarize universally applicable combination rules. For example, the system may find that "Nordic style sofa" is often paired with "light-colored wood grain floor" and "green plants". When the designer inputs the requirements, the system will generate a preliminary design scheme or material list according to these rules, providing efficient data support in the form of effect pictures.

[0003] However, this analysis method relying on past data and fixed rules has increasingly highlighted its limitations in the face of constantly changing design trends. Design styles are not static, but evolve over time, culture, and aesthetic concepts. For example, today's "modern minimalism" may incorporate more natural materials for warmth and environmental considerations. However, the rules summarized by the system based on old data often only repeat classic combinations from the past, making it difficult to capture these subtle design trend changes. When designers attempt to explore emerging styles such as "wabi-sabi", the system-generated solutions may appear out of place, lacking fashion sense and understanding of new aesthetic concepts, simply piecing together elements that fail to reflect the unique use of material texture, spatial white space, and light changes in the style, as well as the deep pursuit of the beauty of imperfection. SUMMARY

[0004] The present application provides a building decoration image analysis method and system to solve the technical problems of existing interior design auxiliary systems in the face of constantly changing design trends, difficulty in capturing subtle design trend changes, inability to understand deep design intentions, and low recognition accuracy when dealing with non-standard data, resulting in chaotic and unpredictable analysis results. In a first aspect, to solve the above technical problems, the present application provides a building decoration image analysis method, comprising: performing visual analysis and processing on the input building decoration image, extracting and quantifying an aesthetic feature vector from the overall visual performance of the building decoration image, the aesthetic feature vector including color saturation distribution, texture complexity, light softness and contrast, spatial white space ratio, and line and shape features; constructing a correlation database of the aesthetic feature vector and decoration element annotation, the correlation database storing the aesthetic feature vector of the professional gallery image and its corresponding decoration element annotation; According to the aesthetic feature vector, a nearest neighbor matching is performed in the association database to obtain an aesthetic feature vector of a professional gallery image closest to the aesthetic feature vector; According to the nearest neighbor matching result, a reverse solution decoration element combination rule is generated, the reverse solution decoration element combination rule including analyzing decoration elements shared by the nearest neighbor professional gallery images and combination manners of the decoration elements, and generating a decoration element combination suggestion corresponding to the decoration element combination rule.

[0005] Preferably, the input architectural decoration image is subjected to visual analysis processing, and an aesthetic feature vector is extracted and quantified from the overall visual performance of the architectural decoration image, including: The architectural decoration image is subjected to adaptive brightness and contrast adjustment; The adjusted architectural decoration image is subjected to semantic segmentation to separate main physical structure regions in the adjusted architectural decoration image; A center skeleton line is extracted for each main physical structure region; In the main physical structure region segmented by the semantic segmentation, color values of pixels are extracted, and relative color relationships between main color clusters are calculated to generate the aesthetic feature vector; The aesthetic feature vector is normalized.

[0006] The extraction of the color values of the pixels in the main physical structure region segmented by the semantic segmentation and the calculation of the relative color relationships between the main color clusters include: Distribution characteristic evaluation is performed on the color values of the pixels in the main physical structure region segmented by the semantic segmentation, the distribution characteristic evaluation including calculation of a number of histogram peak values of the color values of the pixels and distances between adjacent peak values; According to the distribution characteristic evaluation result, a number of main color clusters is dynamically determined; The main color clusters are identified according to the number of main color clusters; In the identified main color clusters, the relative color relationships between the main color clusters are calculated.

[0007] Preferably, the adaptive brightness and contrast adjustment of the architectural decoration image includes: The architectural decoration image is subjected to illumination component and reflection component separation to obtain the illumination component and the reflection component; The illumination component is subjected to dynamic range compression to obtain a dynamic range compressed illumination component; The reflection component is subjected to detail enhancement to obtain a detail enhanced reflection component; The dynamic range compressed illumination component and the detail enhanced reflection component are fused to obtain the adjusted architectural decoration image.

[0008] Preferably, the adjusted architectural decoration image is semantically segmented to isolate the main physical structure regions in the adjusted architectural decoration image, comprising: Multi-scale feature extraction is performed on the adjusted architectural decoration image to obtain visual information of the adjusted architectural decoration image at different resolutions; The multi-scale features are fused to generate a feature representation containing contextual information; Based on the fused feature representation, the pixels of the adjusted architectural decoration image are classified to identify and isolate the main physical structure regions; According to the classification result, a segmentation mask of the main physical structure region is generated.

[0009] Preferably, after classifying the pixels of the adjusted architectural decoration image based on the fused feature representation, it further comprises: In the pixel classification process, a structure boundary smoothing constraint is introduced to suppress the response to structure surface texture, pattern or digital enhancement effect; The structure boundary smoothing constraint is introduced in the pixel classification process to suppress the response to structure surface texture, pattern or digital enhancement effect, comprising: Divide the pixel classification result into local regions to identify physical structure regions with different boundary smoothing requirements; For each of the physical structure regions, calculate the local gradient distribution characteristics of the boundary, which represent the boundary sharpness or boundary blur of the physical structure region; According to the local gradient distribution characteristics, dynamically determine the constraint strength of the structure boundary smoothing of each of the physical structure regions, the constraint strength is negatively correlated with the boundary sharpness and positively correlated with the boundary blur.

[0010] Preferably, the center skeleton line is extracted for each of the main physical structure regions, comprising: Perform geometric analysis based on region shape features on the main physical structure region to identify irregular boundary segments and complex geometric features in the main physical structure region; According to the irregular boundary segments and the complex geometric features, dynamically adjust the local smoothness parameter of the skeleton line extraction, the local smoothness parameter is used to control the sensitivity of the skeleton line to local details; In the skeleton line extraction process, the local smoothness parameter is applied to the irregular boundary segments to suppress the interference of surface details; The distance transformation is performed on the main physical structure region, and the central skeleton line is obtained based on the distance transformation result and the dynamically adjusted local smoothness parameter.

[0011] Preferably, the region shape feature-based geometric analysis is performed on the main physical structure region to identify irregular boundary segments and complex geometric features in the main physical structure region, including: A multi-scale curvature analysis is performed on the boundary of the main physical structure region to obtain curvature variation information of the boundary of the main physical structure region at different scales; The curvature variation information at different scales is fused to generate a curvature feature representation containing local details and global trends; Based on the fused curvature feature representation, an adaptive threshold is set to distinguish geometric noise from the irregular boundary segments or the complex geometric features; According to the distinguishing result, the irregular boundary segments and the complex geometric features are identified.

[0012] Preferably, the distance transformation is performed on the main physical structure region, and the central skeleton line is obtained based on the distance transformation result and the dynamically adjusted local smoothness parameter, including: The distance transformation is performed on the main physical structure region to obtain the distance transformation result; The central skeleton line is obtained according to the distance transformation result and the dynamically adjusted local smoothness parameter, and the central skeleton line includes connectivity and hole structure of the main physical structure region.

[0013] In a second aspect, the present application provides a building decoration image analysis system, comprising: An input end for performing visual analysis and processing on an input building decoration image, extracting and quantifying an aesthetic feature vector from the overall visual performance of the building decoration image, the aesthetic feature vector containing color saturation distribution, texture complexity, light softness and contrast, space white space ratio, and line and shape features; A construction end for constructing an association database of the aesthetic feature vector and decoration element label, the association database storing the aesthetic feature vector of a professional gallery image and its corresponding decoration element label; according to the aesthetic feature vector, performing a nearest neighbor matching in the association database to obtain the aesthetic feature vector of the professional gallery image closest to the aesthetic feature vector; A matching end for inversely solving decoration element combination rules according to the nearest neighbor matching result, the inverse decoration element combination rules including analyzing the decoration elements and their combination modes common to the nearest neighbor professional gallery images; and generating a decoration element combination suggestion corresponding to the decoration element combination rules.

[0014] The application provides a building decoration image analysis method and system. Aesthetic feature vectors are extracted and quantified from the overall visual performance of an input building decoration image through visual analysis and processing. The aesthetic feature vectors include color saturation distribution, texture complexity, light softness and contrast, space white space ratio, and line and shape features. Then, a correlation database of aesthetic feature vectors and decoration element labels is constructed, and nearest neighbor matching is performed in the database according to the extracted aesthetic feature vectors to obtain the aesthetic feature vectors of the closest professional gallery images. Finally, the combination rules of decoration elements are inversely solved according to the nearest neighbor matching results, and corresponding decoration element combination suggestions are generated. The application aims to break through the limitations of traditional methods on discrete element recognition and co-occurrence frequency statistics through deeper aesthetic feature analysis, better capture the redefinition of the'relationship' between elements in design style evolution and the changes in the 'overall atmosphere' created thereby, and provide designers with more forward-looking and personalized design references. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a building decoration image analysis method flowchart provided by an embodiment of the application; Figure 2 is another building decoration image analysis method flowchart provided by an embodiment of the application; Figure 3 is another building decoration image analysis method flowchart provided by an embodiment of the application; Figure 4 is a building decoration image analysis system structure schematic diagram provided by an embodiment of the application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0017] Referring to Figure 1 , a building decoration image analysis method flowchart provided by an embodiment of the application includes the following steps: S1, performing visual analysis and processing on an input building decoration image, and extracting and quantifying aesthetic feature vectors from the overall visual performance of the building decoration image. The aesthetic feature vectors include color saturation distribution, texture complexity, light softness and contrast, space white space ratio, and line and shape features; S2, construct a database associating the aesthetic feature vector with the decoration element annotation, the database storing the aesthetic feature vector of the professional gallery image and its corresponding decoration element annotation; S3, according to the aesthetic feature vector, perform a nearest neighbor matching in the database to obtain the aesthetic feature vector of the professional gallery image closest to the aesthetic feature vector; S4, according to the nearest neighbor matching result, reverse the decoration element combination rule, the reverse decoration element combination rule including analyzing the decoration elements and their combination modes common to the nearest neighbor professional gallery images; and generating a decoration element combination suggestion corresponding to the decoration element combination rule.

[0018] The "aesthetic feature vector" mentioned in the present application is a quantitative representation for describing the visual aesthetic properties of architectural decoration images. It contains multiple dimensions, such as "color saturation distribution" for describing the brightness of colors in the image and their distribution in the image; "texture complexity" for quantifying the fineness and repetition pattern of textures in the image; "light softness and contrast" for evaluating the transition smoothness of light and shadow in the image and the difference degree of light and dark areas; "space blank proportion" for measuring the proportion of blank areas in the image not occupied by main decoration elements, which is usually related to the breathing and balance in design; "line and shape features" for describing the direction, thickness of lines and geometric shape composition in the image. These features together form a multi-dimensional vector that can comprehensively and objectively reflect the overall aesthetic style of the image.

[0019] The "association database" is a storage structure containing a large number of aesthetic feature vectors of professional gallery images and their corresponding "decoration element annotations". "Decoration element annotation" refers to the classification and description of specific decoration elements (such as sofas, coffee tables, lamps, wall materials, etc.) in the image. By associating the aesthetic feature vector with the decoration element annotation, the database can establish a mapping relationship between visual aesthetics and specific design elements.

[0020] "Nearest neighbor matching" is a data retrieval technique that aims to find one or more vectors closest to the query vector (i.e. the aesthetic feature vector extracted from the input architectural decoration image) in the feature space in the database. The distance can be measured by Euclidean distance, cosine similarity, etc. Through nearest neighbor matching, the system can identify the professional gallery image most similar to the input image in terms of aesthetic style.

[0021] The "reverse decoration element combination rule" refers to in-depth analysis of similar images after obtaining the near neighbor matching results to reveal the common decoration elements and their combination methods. This is not just a simple statistical co-occurrence frequency of elements, but more importantly, to understand how these elements interact and jointly create a specific aesthetic style. Finally, the "decoration element combination suggestion" is based on the reverse rule to provide specific design guidance to users, such as recommending specific style furniture matching, color scheme or material selection.

[0022] In specific implementation, the building decoration image analysis method of the present application can be carried out according to the following steps: First, the input building decoration image is subjected to visual analysis processing to extract and quantify the aesthetic feature vector from the overall visual performance of the image. For example, a deep learning model such as a convolutional neural network (CNN) can be used to extract features from the image. This model can be pre-trained on a large set of labeled building decoration image data to enable it to learn and recognize visual aesthetic features in the image. During processing, the image can be input into a pre-trained feature extraction network, which will output a high-dimensional vector, which is the aesthetic feature vector. The aesthetic feature vector can include color saturation distribution, texture complexity, light softness and contrast, space white space ratio, and line and shape features. For example, color saturation distribution can be obtained by calculating the mean and variance of the saturation of different color regions in the image; texture complexity can be quantified by Gabor filter set or local binary pattern (LBP) method; light softness and contrast can be evaluated by analyzing the brightness histogram and edge information of the image; the space white space ratio can be obtained by identifying the main object area through image segmentation technology, and then calculating the pixel ratio of the non-object area; the line and shape features can be extracted by Hough transform or edge detection algorithm to extract the lines and geometric shapes in the image, and then quantitatively describe them.

[0023] Secondly, construct an association database of aesthetic feature vectors and decoration element annotations. This database can be a structured database, such as a relational database or a NoSQL database. In the construction process, a large number of professional gallery images can be collected, and each image can be annotated with decoration elements manually or semi-automatically. For example, for a Nordic-style living room picture, decoration elements such as "light-colored wood grain floor", "gray fabric sofa", "simple pendant lamp", and "green plants" can be annotated. At the same time, visual analysis is performed on these professional gallery images to extract and quantify their aesthetic feature vectors. Then, the aesthetic feature vector of each professional gallery image and its corresponding decoration element annotation are stored as a record in the association database. For example, the aesthetic feature vector can be stored using a vector database, and the decoration element annotation can be stored using a relational database, and the two can be associated through a unique identifier.

[0024] Thirdly, according to the extracted aesthetic feature vector, perform nearest neighbor matching in the association database to obtain the aesthetic feature vectors of the nearest professional gallery images. In nearest neighbor matching, various similarity measurement methods can be used, such as Euclidean distance, cosine similarity, or Manhattan distance. For example, when a new architectural decoration image is input, its aesthetic feature vector is first extracted. Then, compare this vector with the aesthetic feature vectors of all professional gallery images in the association database to calculate their distances or similarities. Finally, select the K aesthetic feature vectors of the nearest professional gallery images as the nearest neighbor matching result. To improve matching efficiency, approximate nearest neighbor (ANN) algorithms such as KD-tree, local sensitive hashing (LSH), or Faiss library can be used.

[0025] Finally, according to the nearest neighbor matching result, reverse the decoration element combination rule, and generate the corresponding decoration element combination suggestion. After obtaining the K most similar professional gallery images, further analysis can be performed on these images. Reverse the decoration element combination rule, including analyzing the decoration elements shared by the nearest professional gallery images and their combination methods. For example, the most frequently appearing decoration elements in the K images and the common collocation relationship between these elements can be counted. For example, if it is found that in the K images, "wooden dining table" and "rattan dining chair" often appear together, and are usually collocated with "warm-toned pendant lamp", then this can be regarded as a decoration element combination rule. On this basis, generate the decoration element combination suggestion corresponding to the decoration element combination rule. For example, according to the reverse rule, provide the user with the suggestion "recommend using wooden dining table with rattan dining chair, and supplemented with warm-toned pendant lamp, to create a natural and warm dining atmosphere". These suggestions can be presented in the form of text description, image example, or 3D model, etc.

[0026] The present application effectively solves the problem that the traditional building decoration image analysis method is difficult to capture subtle design trend changes and understand deep design intent and aesthetic principles when facing changing design trends. The traditional method relies too much on the identification of discrete "elements" and the statistics of co-occurrence frequency, resulting in schemes that often lag behind trends and lack personalization and foresight.

[0027] Compared with the prior art, the core innovation of the present application is that it extracts and quantifies "aesthetic feature vectors" from the "overall visual performance" of building decoration images, and performs matching and rule inversion based on this. Traditional methods may only be able to identify "L-shaped sofas" and "round coffee tables" and count their frequent co-occurrence. However, by extracting aesthetic dimensions such as color saturation distribution, texture complexity, light softness and contrast, space white space ratio, and line and shape features, the present application can capture "deep design intent" or "aesthetic principles" beyond individual element categories and simple co-occurrence frequency. For example, when a designer tries to explore "wabi-sabi", a traditional system may only be able to simply piece together some "old" element combinations, but cannot understand the unique use of material texture, space white space, and light changes in "wabi-sabi", as well as the deep pursuit of the beauty of imperfection.

[0028] By quantifying these abstract aesthetic features, the system can understand the redefinition of "relationships" between elements and the changes in the "overall atmosphere" created thereby. For example, by analyzing light softness and contrast, the system can distinguish between bright and transparent modern styles and soft and quiet wabi-sabi styles; by analyzing the space white space ratio, the system can understand the use of "empty" in minimalism or wabi-sabi. This neighbor matching based on aesthetic feature vectors can more accurately find professional gallery images similar to the input image in "spirit", thereby inversely solving the decoration element combination rules that truly conform to new trends.

[0029] Thus, the present application can generate more forward-looking and personalized decoration element combination suggestions, which not only contain specific decoration elements, but more importantly provide guidance on how to combine these elements to achieve a specific aesthetic effect. This makes the system more adept at handling cutting-edge design needs, significantly improving the efficiency and satisfaction of designers, and providing a more intelligent and more in line with the times trend of auxiliary tools for the field of interior design.

[0030] In some embodiments of the present application, with reference to Figure 2 , S1 includes: S11, adaptively adjusting the brightness and contrast of the building decoration image; S12, performing semantic segmentation on the adjusted building decoration image to separate the main physical structure area in the adjusted building decoration image; S13, extracting a central skeleton line for each of the main physical structure regions; S14, extracting color values of pixels within the main physical structure regions segmented by semantics and calculating relative color relationships between main color clusters to generate the aesthetic feature vector; S15, normalizing the aesthetic feature vector.

[0031] Specifically, adaptive brightness and contrast adjustment is performed on the architectural decoration image, aiming to optimize the visual quality of the image and make it more suitable for subsequent analysis and processing. The adjustment process can dynamically adjust according to the overall brightness distribution and contrast of the image to avoid overexposure or underexposure and enhance the detail performance of the image.

[0032] Among them, the semantic segmentation of the adjusted architectural decoration image to separate the main physical structure regions refers to classifying each pixel in the image into a predefined semantic category, such as walls, floors, ceilings, furniture, etc., so as to identify the main physical structures in the image. Through semantic segmentation, the boundaries of different decoration elements can be accurately defined, providing structured regional information for subsequent feature extraction.

[0033] Further, extracting a central skeleton line for each of the main physical structure regions refers to extracting a skeleton line representing the shape and topological structure of each separated physical structure region through image processing algorithms. The central skeleton line can simplify complex shapes and capture their core geometric features, which is helpful for analyzing line and shape features.

[0034] Within the main physical structure regions segmented by semantics, extracting color values of pixels and calculating relative color relationships between main color clusters refers to analyzing the color information of pixels within each identified physical structure region. Through clustering and other methods, the main color clusters in the image are identified, and the relative relationships between these color clusters, such as hue, saturation, brightness, etc., are further analyzed to quantify the color saturation distribution as an aesthetic feature.

[0035] Finally, normalizing the aesthetic feature vector refers to unifying the extracted aesthetic feature vectors (such as color saturation distribution, texture complexity, light softness and contrast, spatial white space ratio, and line and shape features) to the same numerical range or scale. Normalization processing can eliminate the differences in dimensions and numerical ranges between different features, ensuring that each feature has the same weight in subsequent matching and analysis, improving the accuracy and stability of analysis.

[0036] The scheme of the present application refines the original architectural decoration image analysis process into a series of ordered sub-steps, realizing the accurate extraction and quantification of aesthetic feature vectors. First, the adaptive brightness and contrast adjustment step ensures the optimal visual analyzability of the image, laying the foundation for subsequent feature extraction. Second, the semantic segmentation step decomposes the image into regions with clear physical meaning, enabling subsequent feature extraction to target specific structures and avoiding confusion that may arise from global analysis. By extracting the central skeleton line, line and shape features in the architectural decoration image can be effectively captured, which are crucial for evaluating design style. Extracting pixel color values in specific physical structure regions and calculating the relative color relationships between major color clusters can accurately quantify color saturation distribution, which is a key component of aesthetic features. Finally, normalizing all aesthetic feature vectors ensures that different features can be fairly compared and utilized in subsequent correlation database matching and decoration element combination rule reverse analysis processes, thereby improving the robustness and accuracy of the entire analysis method.

[0037] Through the above technical solutions, more detailed and comprehensive visual analysis of architectural decoration images can be performed. Adaptive brightness and contrast adjustment effectively improves the quality of the image, making subsequent feature extraction more accurate. The application of semantic segmentation enables the extraction of aesthetic features to focus on key physical structures in the image, avoiding interference from irrelevant background information, thereby improving the representativeness and effectiveness of the features. The extraction of the central skeleton line provides a structured basis for the quantification of line and shape features. The calculation of major color clusters and their relative color relationships makes the quantification of color saturation distribution more detailed and accurate. The normalization of aesthetic feature vectors ensures the comparability of features of different dimensions, significantly improving the accuracy and stability of subsequent nearest neighbor matching. Overall, these refined steps work together to enable the aesthetic feature vectors extracted from architectural decoration images to more accurately and comprehensively reflect the overall visual performance of the images, providing a high-quality data foundation for subsequent decoration element combination rule reverse analysis.

[0038] In particular, with reference to Figure 3 , S14 includes: S141, evaluating the distribution characteristics of pixel color values in the main physical structure region segmented by semantics, including calculating the number of histogram peaks of the pixel color values and the distance between adjacent peaks; S142, dynamically determining the number of major color clusters according to the distribution characteristic evaluation results; S143, identifying the major color clusters according to the number of major color clusters; S144, calculating the relative color relationships between the identified major color clusters.

[0039] Specifically, the distribution characteristic evaluation of pixel color values refers to analyzing the distribution of pixel color values in the color space to reveal their inherent structure and regularity. Among them, calculating the number of histogram peak values of pixel color values aims to identify the types or number of main colors in the image, and each peak value usually corresponds to a main color. The distance between adjacent peak values is used to measure the difference or transition smoothness between these main colors. The greater the distance, the more significant the color difference, and the smaller the distance, the softer the color transition. The purpose is to provide a quantitative basis for subsequent color cluster identification, ensuring accurate preliminary judgment of image color composition.

[0040] Further, according to the distribution characteristic evaluation results, the number of main color clusters is dynamically determined. This means that the system does not use a preset fixed number of color clusters, but adjusts the number of color clusters to be identified according to the color complexity and distribution characteristics of the image itself. For example, if the evaluation results show that the number of histogram peak values is large and the distance between peak values is large, more color clusters may need to be identified; conversely, if the number of peak values is small and the distance between peak values is small, fewer color clusters can be identified. This dynamic determination method can better adapt to different architectural decoration image color styles and avoid under-segmentation or over-segmentation problems caused by a fixed number.

[0041] On this basis, the main color clusters are identified according to the number of main color clusters. Once the number of color clusters is determined, various clustering algorithms can be used, such as K-means clustering, Mean-Shift clustering, or Gaussian Mixture Model (GMM), to group pixel color values within the main physical structure region. Each group is a main color cluster, representing a representative color in the image. The identified color clusters can effectively simplify complex color information into several core color groups, facilitating subsequent analysis.

[0042] Finally, among the identified main color clusters, the relative color relationships between the main color clusters are calculated. Relative color relationships can be understood as the mutual position, distance, or matching rules of different color clusters in the color space. For example, the hue difference, saturation difference, brightness difference of two color clusters, or their Euclidean distance in a specific color model (such as HSV, Lab) can be calculated. The purpose is to capture the harmony, contrast, or primary and secondary relationships of color matching, which is crucial for evaluating the aesthetic features of architectural decoration images. By analyzing these relative relationships, the color composition and designer's color intent of the image can be better understood.

[0043] The scheme of the present application can accurately capture the types of main colors in the image and the degree of difference between them by evaluating the distribution characteristics of the pixel color values in the main physical structure region segmented by semantic segmentation. It is precisely due to this detailed evaluation that the system can dynamically determine the number of main color clusters, thereby avoiding the limitations brought by the fixed number of color clusters in traditional methods and ensuring the adaptability to images of different color styles. On this basis, by identifying these dynamically determined main color clusters and further calculating the relative color relationships between them, the scheme can quantify the color aesthetic features of the image from multiple dimensions, such as color harmony, contrast, and primary and secondary relationships, providing more detailed and accurate color information for subsequent aesthetic feature vector extraction.

[0044] Through the above technical scheme, more detailed and accurate quantification of the color features of the main physical structure region in the architectural decoration image can be achieved. Specifically, dynamically determining the number of main color clusters enables the color analysis to better adapt to the complexity and diversity of the image itself, avoiding the deviation caused by fixed parameters. At the same time, calculating the relative color relationships between the main color clusters can more comprehensively capture the aesthetic properties of color matching, such as color contrast and harmony, thereby improving the accuracy and representativeness of the aesthetic feature vector. Thus, the present scheme can provide more insightful color information for the analysis of architectural decoration images, thereby improving the accuracy of the reverse solution of decoration element combination rules and the practicality of the recommendations.

[0045] In some embodiments of the present application described above, adaptive brightness and contrast adjustment is proposed for architectural decoration images. However, in actual applications, architectural decoration images often face complex and variable lighting environments, such as strong local shadows, overexposed highlight areas, or insufficient overall light, etc. Traditional adaptive adjustment methods may not be able to optimize brightness and contrast simultaneously under these complex lighting conditions while effectively preserving the details of the image, resulting in problems such as loss of details, color distortion, or unnatural overall appearance in the adjusted image, which in turn affects the accuracy of subsequent aesthetic feature vector extraction. To this end, the present application further proposes an optimization scheme for adaptive brightness and contrast adjustment of architectural decoration images, aiming to improve the adjustment effect through a refined image processing process to better adapt to complex lighting environments.

[0046] The above adaptive brightness and contrast adjustment for architectural decoration images specifically includes: separating the illumination component and the reflection component from the architectural decoration image to obtain the illumination component and the reflection component; performing dynamic range compression on the illumination component to obtain a dynamic range compressed illumination component; performing detail enhancement on the reflection component to obtain a detail-enhanced reflection component; fusing the dynamic range compressed illumination component and the detail-enhanced reflection component to obtain an adjusted architectural decoration image.

[0047] Specifically, the illumination component can be understood as the brightness information in the image produced by direct or indirect illumination of the light source, which mainly reflects the overall light and dark distribution of the image. The reflection component refers to the inherent color and texture information of the object surface in the image, which is independent of the illumination, and mainly reflects the details and material characteristics of the object.

[0048] The purpose of separating the architectural decoration image into the illumination component and the reflection component is to decouple the brightness information and the detail information of the image for independent optimization processing. Among them, performing dynamic range compression on the illumination component means processing the separated illumination component to reduce the range of its brightness value, so as to avoid the appearance of overexposed or underexposed areas in the image. The dynamic range compressed illumination component can make the overall brightness distribution of the image more uniform, and improve the visual comfort.

[0049] In practical applications, performing detail enhancement on the reflection component means processing the separated reflection component to highlight the texture, edge and fine structure of the object in the image. The detail-enhanced reflection component can make the visual details of the image clearer and sharper, and improve the expressiveness of the image. Finally, fusing the dynamic range compressed illumination component and the detail-enhanced reflection component aims to recombine the brightness information and the detail information that have been independently optimized to generate an adjusted architectural decoration image that has good brightness contrast and retains rich details.

[0050] The scheme of the present application realizes independent processing of the brightness information and the detail information of the image by decomposing the architectural decoration image into the illumination component and the reflection component. The illumination component carries the overall light and dark information of the image, and dynamic range compression on it can effectively solve the overexposure or underexposure problem caused by uneven illumination in the image, so that the overall brightness distribution of the image tends to be balanced. At the same time, the reflection component contains the inherent texture and structural details of the image, and detail enhancement processing on it can effectively improve the clarity and sharpness of the image, and avoid the loss of details in the brightness adjustment process. It is this divide-and-conquer strategy that enables the brightness, contrast and details of the image to be optimized simultaneously, thereby overcoming the limitations of traditional methods in which it is difficult to balance brightness and details under complex lighting conditions.

[0051] In some preferred embodiments, the following is described by a specific example. Suppose an image containing indoor architectural decoration is input, the window area of the image is overexposed due to strong light, while the corners of the room are dark due to insufficient light, and the fine texture on the wall is also difficult to identify.

[0052] First, the architectural decoration image is separated into illumination component and reflection component. This can be achieved by algorithms based on Retinex theory, such as Single-Scale Retinex (SSR) or Multi-Scale Retinex (MSR), which decompose the image into low-frequency components representing illumination and high-frequency components representing reflection.

[0053] Second, the separated illumination component is compressed in dynamic range. For example, methods such as logarithmic transformation, gamma correction or histogram equalization can be used to pull the brightness values of overexposed and underexposed areas back into a suitable range, making the overall brightness distribution of the image more uniform, the window area no longer overexposed, and the corner area clear and visible.

[0054] Then, the separated reflection component is enhanced in detail. This can be achieved by techniques such as Unsharp Masking, Difference of Gaussians (DoG) or wavelet transform to highlight details such as wall texture, furniture edges, etc., making them visually clearer and sharper.

[0055] Finally, the illumination component compressed in dynamic range is fused with the reflection component enhanced in detail. The fusion process is usually achieved by simple multiplication or weighted average to recombine the two into a new image. Thus, the adjusted architectural decoration image will have both uniform brightness distribution and rich detail performance, with details in the window area restored, objects in the corner clearly visible, and wall texture lively, providing high-quality visual input for subsequent aesthetic feature extraction.

[0056] In some embodiments of the present application described above, the adjusted architectural decoration image is segmented semantically to separate the main physical structure area. However, in practical applications, architectural decoration images often contain complex structures, diverse textures and different scale information. If a single scale feature or a segmentation method lacking context information is used, it may lead to insufficient segmentation accuracy, making it difficult to accurately identify and separate the main physical structure area in the image, thereby affecting the accuracy of subsequent aesthetic feature vector extraction and analysis. If the above problems are not solved, it may lead to inaccurate quantification of aesthetic feature vectors, which in turn affects the quality of the generated decoration element combination rules and suggestions.

[0057] To this end, the present application further proposes the above-mentioned semantic segmentation of the adjusted architectural decoration image to separate the main physical structure area, comprising: performing multi-scale feature extraction on the adjusted architectural decoration image to obtain visual information of the adjusted architectural decoration image at different resolutions; fusing the multi-scale features to generate a feature representation containing context information; Based on the fused feature representation, the pixels of the adjusted architectural decoration image are classified to identify and separate the main physical structure regions; According to the classification result, a segmentation mask of the main physical structure region is generated.

[0058] Specifically, multi-scale feature extraction refers to capturing visual information at different spatial scales by using different sizes of convolution kernels or performing feature extraction on images with different down-sampling rates. The purpose is to ensure that both small decorative details and large architectural structures can be effectively perceived and encoded. For example, pyramid network structures such as FPN or PSPNet can be used to obtain feature maps of images at multiple resolution levels.

[0059] Among them, the fusion of multi-scale features can be understood as integrating features extracted from different scales to generate a more comprehensive and discriminative feature representation. This fusion is usually achieved through splicing, summation or attention mechanisms, etc. The purpose is to combine local detail information with global context information to enhance the understanding of image content. Through fusion, the limitations of single-scale features can be effectively compensated, making the subsequent pixel classification better utilize the overall structure and local texture information of the image.

[0060] In practical applications, based on the fused feature representation, the pixels of the adjusted architectural decoration image are classified, specifically referring to using the decoder part of the deep learning model (such as U-Net, DeepLab, etc.) to map the fused features back to the original image resolution, and to predict the class to which each pixel belongs (for example, wall, floor, ceiling, window, furniture, etc. Main physical structure region). The purpose is to achieve accurate classification at the pixel level and lay the foundation for subsequent separation of main physical structure regions.

[0061] According to the classification result, a segmentation mask of the main physical structure region is generated, which refers to converting the output of pixel classification into a binary or multi-value image mask, where each mask region corresponds to an identified main physical structure region. For example, for the identified "wall" region, a corresponding mask can be generated, in which the wall pixels have a value of 1 and other pixels have a value of 0. The purpose is to provide clear and accurate region boundaries to facilitate further analysis and processing of each independent structure region.

[0062] By the technical solution, the accuracy and robustness of the architectural decoration image semantic segmentation can be improved. Through multi-scale feature extraction, different sizes and details of physical structures in the image can be effectively perceived; through feature fusion, the understanding of the context information of the image is enhanced, so that the segmentation result is more in line with the actual scene semantics; through pixel classification and segmentation mask generation based on the fused features, accurate identification and separation of the main physical structure area are realized. This not only solves the problems of inaccurate segmentation and blurred boundaries of traditional single-scale segmentation methods in complex architectural decoration images, but also provides a high-quality area basis for accurate extraction and quantization of aesthetic feature vectors, thereby improving the overall performance and practical value of the entire architectural decoration image analysis method.

[0063] In some preferred embodiments, the following is described by a specific example. Assume that an architectural decoration image containing a living room, a sofa, a window, and a wall is input. First, multi-scale feature extraction is performed on the adjusted architectural decoration image. A feature extractor based on a convolutional neural network (CNN) can be used, such as a ResNet or VGG network, to extract feature maps at different levels (e.g., from shallow to deep) to obtain visual information at different resolutions. For example, shallow features may contain local details such as edges and textures, while deep features contain higher-level semantic information and global structures.

[0064] Then, the multi-scale features are fused. This can be achieved through a feature pyramid network (FPN), which fuses high-level semantic information with low-level detail information through upsampling, thereby generating a feature representation with rich semantics and details at all scales. For example, high-resolution low-level features can be element-wise added or concatenated with upsampled low-resolution high-level features to generate a feature representation containing context information.

[0065] Then, based on the fused feature representation, the pixels of the adjusted architectural decoration image are classified. This can use a semantic segmentation model, such as U-Net or DeepLabV3+, whose decoder part receives the fused feature representation and classifies each pixel in the image, predicting its probability of belonging to the main physical structure area such as "wall", "sofa", "window", or "floor".

[0066] Finally, according to the pixel classification results, segmentation masks of the main physical structure regions are generated. For example, for each pixel, it is assigned to the class with the highest probability, resulting in a pixel-level classification map. Then, this classification map is converted into a series of binary masks, each corresponding to an identified main physical structure region. For example, a "wall" mask can be generated, in which all pixels classified as walls have a value of 1, and the rest have a value of 0; similarly, a "sofa" mask, a "window" mask, etc. These precise segmentation masks will serve as the basis for subsequent extraction of aesthetic feature vectors, ensuring accurate quantification of the aesthetic properties of each independent structure region.

[0067] In some embodiments of the present application described above, the adjusted architectural decoration image is subjected to semantic segmentation to separate the main physical structure regions in the adjusted architectural decoration image. However, if the structure boundary is not effectively handled during the pixel classification process, the segmentation results may be overly sensitive to subtle features such as structure surface texture, pattern, or digital enhancement, resulting in jagged and inaccurate structure boundaries, affecting the accuracy of subsequent aesthetic feature extraction.

[0068] To this end, the present application further proposes to introduce a structure boundary smoothing constraint in the above-mentioned pixel classification process to suppress the response to structure surface texture, pattern, or digital enhancement effect.

[0069] Specifically, in the pixel classification process, a structure boundary smoothing constraint is introduced to suppress the response to structure surface texture, pattern, or digital enhancement effect. The introduction of a structure boundary smoothing constraint in the pixel classification process to suppress the response to structure surface texture, pattern, or digital enhancement effect includes: Divide the pixel classification results into local regions to identify physical structure regions with different boundary smoothness requirements; For each of the physical structure regions, calculate the local gradient distribution characteristics of the boundary, which represent the sharpness or fuzziness of the boundary of the physical structure region. According to the local gradient distribution characteristics, dynamically determine the constraint strength of the structure boundary smoothing of each of the physical structure regions, which is negatively correlated with the sharpness of the boundary and positively correlated with the fuzziness of the boundary.

[0070] Wherein, the introduction of the structure boundary smoothing constraint refers to applying an additional regularization mechanism when classifying the pixels of the adjusted architectural decoration image, aiming to ensure that the boundaries of the identified main physical structure regions have a certain smoothness, avoiding the fragmentation or inaccuracy of the segmentation results caused by the misidentification of subtle texture, pattern, or digital surface details in the image as structure boundaries.

[0071] The pixel classification result is divided into local regions, which aims to divide the preliminary classified pixel set into multiple independent local regions, so as to perform fine processing on the characteristics of each region, so as to identify which region boundary needs stronger smoothing processing and which region boundary needs to maintain its sharpness.

[0072] The local gradient distribution characteristics of the boundary are calculated for each physical structure region, which aims to quantify the clarity of the region boundary. The local gradient distribution characteristics of the boundary can be understood as the sharpness and range of the change of pixel intensity along the boundary direction. High gradient value usually indicates high boundary sharpness, while low gradient value or wide gradient distribution range indicates high boundary blur.

[0073] The constraint strength of structure boundary smoothing is dynamically determined according to the local gradient distribution characteristics, which aims to realize adaptive boundary smoothing processing. The constraint strength is negatively correlated with the boundary sharpness, which means that for clear and sharp boundaries, the applied smoothing constraint is weak to preserve the original geometric characteristics; and it is positively correlated with the boundary blur, which means that for fuzzy and ambiguous boundaries, the applied smoothing constraint is strong to effectively regularize and smooth them.

[0074] The scheme of the present application effectively solves the problem that the traditional semantic segmentation method is easy to misidentify texture, pattern or digital non-structural details as structure boundaries when processing building decoration images with rich surface details, by introducing structure boundary smoothing constraints in the pixel classification process. By dividing the pixel classification result into local regions and calculating the local gradient distribution characteristics of the boundary for each physical structure region, the system can accurately evaluate the sharpness or blur of the boundary. Based on this, the constraint strength of structure boundary smoothing is dynamically adjusted to ensure that clear structure boundaries can be preserved, while noisy boundaries caused by surface details can be effectively smoothed, thereby generating more accurate and robust semantic segmentation results.

[0075] Through the above technical scheme, the present application can effectively suppress the response to structure surface texture, pattern or digital enhancement effect, thereby avoiding the misidentification of subtle surface details as structure boundaries in the pixel classification process. As a result, the obtained semantic segmentation result has smoother and more accurate physical structure region boundaries, significantly improving the recognition accuracy and robustness of the main physical structure region. This provides more reliable basic data for subsequent aesthetic feature vector extraction (such as color saturation distribution, texture complexity, etc.), ensuring the accuracy and practicality of the final decoration element combination suggestion.

[0076] In some preferred embodiments, an architectural decoration image is input, which contains a wall with intricate patterns and a window with clear edges.

[0077] Firstly, multi-scale feature extraction and fusion are performed on the adjusted architectural decoration image, and a preliminary classification of pixels is performed based on the fused feature representation to identify wall regions and window regions.

[0078] Next, the preliminary classification results are divided into local regions, and the wall regions and window regions are regarded as two independent physical structure regions.

[0079] Then, the local gradient distribution characteristics of the boundaries of the wall regions are calculated. Since the wall has fine patterns, its boundaries may exhibit certain fuzziness or irregularity in the local area. The local gradient distribution characteristics of the boundaries of the window regions are calculated. Since the edges of the window are usually very clear, the boundaries will exhibit high sharpness.

[0080] Finally, according to the calculated local gradient distribution characteristics, the constraint strength of structure boundary smoothing is dynamically determined. For the wall region, since its boundary fuzziness is high, a strong smoothing constraint is applied to eliminate the subtle irregularities caused by the patterns, so that the overall boundary of the wall exhibits a smooth and continuous effect. For the window region, since its boundary sharpness is high, a weak smoothing constraint is applied to maximize the clarity and geometric accuracy of the window edge. Through this adaptive smoothing process, the final semantic segmentation mask can accurately separate the smooth wall region and the window region with clear edges, avoiding the interference of surface texture on structure boundary recognition.

[0081] In some embodiments of the present application, a center skeleton line is extracted for each main physical structure region. However, in actual architectural decoration images, the main physical structure region often contains irregular boundary segments and complex geometric features, such as fine carvings, arc structures or decorative patterns. If a uniform skeleton line extraction strategy is directly used, it may cause the extracted skeleton line to be too sensitive to these local details, resulting in redundant, inaccurate or broken skeleton lines, which affects the accurate inverse solution of the decoration element combination rule.

[0082] To this end, the present application further proposes that the step of extracting a center skeleton line for each main physical structure region includes: performing geometric analysis based on the shape features of the main physical structure region to identify irregular boundary segments and complex geometric features in the main physical structure region; dynamically adjusting the local smoothness parameter of the skeleton line extraction according to the irregular boundary segments and the complex geometric features, the local smoothness parameter being used to control the sensitivity of the skeleton line to local details; in the skeleton line extraction process, applying the local smoothness parameter to the irregular boundary segments to suppress the interference of surface details; performing distance transformation on the main physical structure region, and obtaining the center skeleton line based on the distance transformation result and the dynamically adjusted local smoothness parameter.

[0083] Specifically, the geometry analysis based on the region shape features of the main physical structure region refers to accurately identifying the non-standard or complex geometric shapes existing in the region by analyzing the geometric properties such as the boundary contour, curvature variation, and topological structure of the region. For example, multi-scale curvature analysis can be used to detect sharp corners, smooth curves, or concave-convex variations of the boundary, so as to distinguish between subtle geometric noise and actual irregular boundary segments or complex geometric features. The purpose is to provide refined region feature information for subsequent skeleton line extraction.

[0084] Among them, dynamically adjusting the local smoothness parameter of skeleton line extraction according to the identified irregular boundary segments and complex geometric features can be understood as adaptively setting the degree of response of the skeleton line algorithm to details according to the local geometric characteristics of the region.

[0085] The local smoothness parameter is used to control the smoothness of the skeleton line in different regions. For example, a lower smoothness can be used in flat or regular regions to preserve structural details, while a higher smoothness can be used in irregular or complex regions to filter out noise and suppress surface detail interference. The purpose is to ensure that the skeleton line can accurately reflect the main structure while avoiding being misled by secondary details.

[0086] In practical applications, during the skeleton line extraction process, the local smoothness parameter is applied to irregular boundary segments to suppress surface detail interference. Specifically, when executing the skeleton line generation algorithm, a dynamically adjusted local smoothness parameter is applied to the irregular boundary region identified by geometry analysis. For example, a local weighting function can be introduced in the distance transformation or morphological operation, or the local convergence condition can be adjusted in the iterative thinning process, so as to impose stronger smoothness constraints on the generation of skeleton line in these regions, and reduce the skeleton line branches caused by local texture or slight concave-convex. The purpose is to ensure that the extracted skeleton line can better represent the macroscopic morphology of the main physical structure rather than the subtle texture of its surface. Further, distance transformation is performed on the main physical structure region, and the center skeleton line is obtained based on the distance transformation result and the dynamically adjusted local smoothness parameter.

[0087] Distance transform is an image processing technique that encodes the distance from each pixel in a binary image to the nearest background pixel. By combining the distance transform result with a dynamically adjusted local smoothness parameter, the position and shape of the skeleton can be determined more accurately. For example, based on the distance transform result, a local maximum-based thinning algorithm can be employed, and an adjusted smoothness parameter can be applied at irregular boundary segments to guide the skeleton to extend along the center of the region while avoiding excessive response to local details. The purpose is to generate a center skeleton that not only reflects the topological structure of the region, but also has good smoothness.

[0088] The scheme of the present application first conducts a fine understanding of the main physical structure region by introducing a geometric analysis based on the shape features of the region, and identifies irregular boundary segments and complex geometric features that may be difficult for traditional methods to handle. It is precisely due to this pre-geometric analysis that the subsequent skeleton extraction process can dynamically adjust the local smoothness parameter according to the actual complexity of the region. During the skeleton extraction process, the appropriate local smoothness parameter is applied to the identified irregular boundary segments, effectively suppressing the interference of surface details and avoiding unnecessary bending or branching of the skeleton due to local noise or fine texture. Finally, by combining the distance transform result with the dynamically adjusted local smoothness parameter, a center skeleton can be generated that not only accurately reflects the topological connectivity of the main physical structure region, but also has good smoothness and robustness.

[0089] Through the above technical scheme, the present application can effectively solve the problem that the traditional skeleton extraction method is prone to produce redundant, inaccurate or overly sensitive to details skeleton when processing architectural decoration images with irregular boundary segments and complex geometric features. Through geometric analysis of the shape features of the region and dynamic adjustment of the local smoothness parameter, the extracted center skeleton can better represent the macroscopic morphology of the main physical structure, reducing errors caused by local detail interference. This significantly improves the accuracy and robustness of skeleton extraction, providing more reliable foundation data for subsequent aesthetic feature vector extraction and inverse solution of decoration element combination rules, thereby improving the overall performance and practicality of the entire architectural decoration image analysis method.

[0090] In some preferred embodiments, it is assumed that an architectural decoration image containing a complex carved window frame needs to be analyzed.

[0091] First, the image is semantically segmented to separate the window frame as the main physical structure region.

[0092] Next, the window frame region is subjected to geometric analysis based on the shape features of the region. For example, by performing multi-scale curvature analysis on the window frame boundary, the fine curves, sharp corners, and irregular concave-convex structures of the carved part can be identified as irregular boundary segments and complex geometric features. According to these identification results, the system dynamically adjusts the local smoothness parameters for skeleton line extraction, for example, setting a higher smoothness parameter at the carved details to avoid the skeleton line excessively tracking each carved texture, and setting a lower smoothness parameter at the straight part of the window frame to maintain its structural integrity.

[0093] Subsequently, the window frame region is subjected to distance transformation, and combined with these dynamically adjusted local smoothness parameters, the skeleton line extraction algorithm is executed. Thus, the obtained central skeleton line will be able to accurately reflect the overall structure of the window frame, while smoothly passing through the complex carved region, avoiding redundant branches caused by detail interference, thereby providing a clear and concise structural representation for subsequent decorative element analysis.

[0094] Specifically, the above-mentioned geometric analysis of the main physical structure region based on the shape features of the region to identify irregular boundary segments and complex geometric features in the main physical structure region can be implemented in the following manner.

[0095] Performing multi-scale curvature analysis on the boundary of the main physical structure region to obtain curvature variation information of the boundary of the main physical structure region at different scales; Fusing the curvature variation information at different scales to generate a curvature feature representation containing local details and global trends; Based on the fused curvature feature representation, setting an adaptive threshold to distinguish between the fine geometric noise and the irregular boundary segment or the complex geometric feature; According to the distinguishing result, identifying the irregular boundary segment and the complex geometric feature.

[0096] Wherein, multi-scale curvature analysis refers to curvature calculation and analysis on the boundary of the main physical structure region at different spatial scales. For example, the boundary can be smoothed at different standard deviations by a Gaussian smoothing filter, and then the curvature at each smoothing scale is calculated. The purpose is to capture the geometric properties of the boundary at different levels of detail, thereby comprehensively understanding its shape. The curvature variation information at different scales can be understood as the boundary curvature values and their distribution calculated at different smoothing scales. These information reflects the bending degree and direction of the boundary in local and global range.

[0097] The curvature change information at different scales is fused, and specifically, feature vector splicing, weighted average or a machine learning-based method can be used for fusion to generate a unified curvature feature representation containing rich context information. The fused feature representation can reflect both the subtle local changes and the overall trend of the boundary, avoiding the loss of details or noise sensitivity problems that may be caused by single-scale analysis.

[0098] Based on the fused curvature feature representation, an adaptive threshold is set, which means that one or more thresholds are dynamically determined according to the statistical characteristics of the fused curvature feature representation or preset rules. For example, a suitable percentile can be selected as a threshold according to the histogram distribution of the curvature feature. The purpose is to adapt to the geometric characteristics of different images and different physical structure regions flexibly, and improve the accuracy of differentiation. The subtle geometric noise generally refers to the small, random boundary fluctuations introduced in the image acquisition or processing process that do not belong to the actual structure features. The irregular boundary segment or complex geometric feature refers to the boundary part with significant bending, sharp corners, concave or convex, and other non-smooth, irregular shapes. Through the adaptive threshold, these noises can be effectively distinguished from the real geometric features.

[0099] According to the differentiation result, the irregular boundary segment and the complex geometric feature are identified, which means that the boundary segment with a curvature feature value higher or lower than a certain threshold is marked as an irregular boundary segment or a complex geometric feature according to the threshold judgment.

[0100] The scheme of the present application can comprehensively capture the geometric characteristics of the boundary at different levels of detail through multi-scale curvature analysis of the boundary of the main physical structure region. By fusing the curvature change information at different scales, a curvature feature representation that contains both local details and reflects global trends can be generated, thereby overcoming the limitations that may be caused by single-scale analysis, such as sensitivity to noise or neglect of important details. On this basis, by setting an adaptive threshold, the differentiation standard can be dynamically adjusted according to the actual curvature feature distribution, so that subtle geometric noise can be effectively filtered out, and the real irregular boundary segment and complex geometric feature can be accurately identified. This method ensures the robustness and accuracy of geometric analysis and provides a reliable basis for the dynamic adjustment of the local smoothness parameter in the subsequent skeleton line extraction process.

[0101] The above technical solution enables accurate identification of irregular boundary segments and complex geometric features in the main physical structure regions. Compared with traditional single-scale or fixed-threshold methods, this application employs multi-scale analysis and adaptive thresholding, significantly improving the accuracy and robustness of boundary geometric feature identification, effectively avoiding misjudging subtle noise as structural features or missing important irregular structures. This provides a more refined and reliable input for the dynamic adjustment of local smoothness parameters during subsequent skeleton line extraction, thereby improving the quality of central skeleton line extraction and the ability to preserve details in architectural and decorative images.

[0102] In some embodiments of this application, when extracting the central skeleton line from the main physical structure region, a distance transformation is performed on the main physical structure region, and the central skeleton line is obtained based on the distance transformation result and dynamically adjusted local smoothness parameters. However, in practical applications, relying solely on distance transformation and local smoothness parameters may not fully guarantee that the extracted central skeleton line accurately reflects the topology of the original main physical structure region, such as its connectivity or internal hole structure. If the skeleton line is topologically inaccurate, it may lead to deviations in subsequent aesthetic feature analysis and inverse derivation of decorative element combination rules.

[0103] In response, this application further proposes the aforementioned method of performing distance transformation on the main physical structure regions, and obtaining the central skeleton line based on the distance transformation results and dynamically adjusted local smoothness parameters, including: A distance transformation is performed on the main physical structure region to obtain the distance transformation result; Based on the distance transformation result and the dynamically adjusted local smoothness parameter, the central skeleton line is obtained, and the topological constraints ensure that the central skeleton line includes the connectivity and hole structure of the main physical structure region.

[0104] Specifically, distance transform is an image processing technique used to calculate the distance of each pixel in an image to the nearest zero pixel (typically the background). The result is usually a distance map, where the value of each pixel represents its distance to a boundary. Dynamically adjusted local smoothness parameters, as described in the above embodiments, are used during skeleton line extraction to control the sensitivity of the skeleton lines to local details based on irregular boundary segments and complex geometric features of the main physical structure regions, thereby suppressing interference from surface details.

[0105] Topological constraints can be understood as a series of rules or conditions applied during skeleton line extraction to preserve the connectivity and hole structure of the original main physical structure regions. For example, in skeletonization algorithms (such as thinning algorithms or distance-transform-based skeleton extraction algorithms), topology preservation operations can be introduced to ensure that the connectivity of the region is not destroyed or holes are accidentally filled or created when iteratively removing pixels. Specifically, this can be done by checking the connectivity of the pixel's neighborhood before and after removal, for example, using Euler number or connected component analysis. The goal is to ensure that the extracted central skeleton line not only reflects the geometric center of the region, but more importantly, accurately preserves the topological features of the original region, thus providing a reliable basis for subsequent structural analysis.

[0106] This application's solution effectively addresses the potential topological distortion problem inherent in traditional methods by introducing topological constraints during the process of obtaining the central skeleton line based on distance transformation results and dynamically adjusted local smoothness parameters. The application of these topological constraints allows for strict monitoring and maintenance of the connectivity and hole structure of the main physical structural regions during skeleton line extraction. For example, when the skeletonization algorithm attempts to remove a pixel, the topological constraints assess whether the removal operation would cause region breakage or hole disappearance; if so, the removal of the pixel is prevented. This mechanism ensures that the generated central skeleton line is geometrically simplified but topologically consistent with the original region, thus avoiding structural misjudgments due to the loss of topological information.

[0107] Through the aforementioned technical solution, the extracted central skeleton line can more accurately and robustly represent the intrinsic structure of the main physical structural areas in architectural decoration images. This topology-preserving skeleton line has a significant positive impact on subsequent aesthetic feature vector extraction, especially the quantification of line and shape features. It ensures accurate identification of structural connectivity and holes, thereby improving the understanding of complex geometric forms and spatial layouts in architectural decoration images and providing a solid data foundation for generating more accurate suggestions for decorative element combinations.

[0108] In some preferred embodiments, a specific example is given below. Suppose there is a U-shaped architectural decoration area containing a noticeable hole. If skeletonization is performed using only distance transformation and local smoothness parameters, in some cases, due to noise or improper algorithm parameter settings, the internal hole may be "filled in" during skeletonization, or the lines connecting to the outside may break at some minor points.

[0109] Specifically, after performing distance transformation on the "hui"-shaped main physical structure region, a distance map can be obtained. When performing skeleton extraction based on this distance map, topological structure constraints are introduced. For example, a topology-preserving thinning algorithm can be adopted. In each iteration of removing boundary pixels, this algorithm checks whether the removal operation will change the Euler number of the image (the Euler number is related to the number of connected components and holes). If removing a certain pixel causes a change in the Euler number, that is, it destroys the connectivity or hole structure, then this pixel will not be removed. In this way, even when there are irregularities or fine structures at the region boundary, the obtained central skeleton line can ensure accurately containing the external connectivity of the "hui"-shaped region and the internal hole structure. Thereby, subsequent analysis of the line and shape features of this skeleton line will be able to accurately identify the complete structure of the "hui"-shape, avoiding misjudgment caused by the loss of topological information.

[0110] Referring to Figure 2 , the structural diagram of a building decoration image analysis system provided by an embodiment of the present invention includes: An input end, configured to perform visual analysis and processing on the input building decoration image, extract and quantify an aesthetic feature vector from the overall visual performance of the building decoration image, and the aesthetic feature vector includes color saturation distribution, texture complexity, light softness and contrast, spatial blank space ratio, and line and shape features; A construction end, configured to construct an association database between the aesthetic feature vector and the decoration element annotation, and the association database stores the aesthetic feature vectors of professional gallery images and their corresponding decoration element annotations; perform nearest neighbor matching in the association database according to the aesthetic feature vector to obtain the aesthetic feature vector of the professional gallery image that is closest to the aesthetic feature vector; A matching end, configured to reverse-analyze the combination rule of decoration elements according to the nearest neighbor matching result, and the reverse-analysis of the combination rule of decoration elements includes analyzing the common decoration elements and their combination methods of the nearest neighbor professional gallery images; and generating a decoration element combination suggestion corresponding to the combination rule of decoration elements.

[0111] It should be noted that the building decoration image analysis system provided by an embodiment of the present invention is used to execute all the process steps of the building decoration image analysis method in the above embodiment, and the working principles and beneficial effects of the two correspond one by one, so they will not be elaborated here.

[0112] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0113] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for analyzing architectural decoration images, characterized in that, The method includes: Visual analysis processing is performed on the input architectural decoration image to extract and quantify aesthetic feature vectors from the overall visual performance of the architectural decoration image. The aesthetic feature vectors include color saturation distribution, texture complexity, light softness and contrast, spatial white space ratio, and line and shape features. Construct an association database between the aesthetic feature vectors and decorative element annotations, wherein the association database stores the aesthetic feature vectors of professional image library images and their corresponding decorative element annotations; Based on the aesthetic feature vector, nearest neighbor matching is performed in the association database to obtain the aesthetic feature vector of the professional image library image that is closest to the aesthetic feature vector; Based on the nearest neighbor matching results, the rules for combining decorative elements are deduced in reverse. The deduction of the rules for combining decorative elements includes analyzing the decorative elements and their combination methods that are common to images in the nearest professional image library; and generating decorative element combination suggestions corresponding to the rules for combining decorative elements.

2. The method for analyzing architectural decoration images according to claim 1, characterized in that, The step of performing visual analysis processing on the input architectural decoration image, extracting and quantifying aesthetic feature vectors from the overall visual representation of the architectural decoration image, includes: Adaptive brightness and contrast adjustment is performed on the architectural decoration image; Semantic segmentation is performed on the adjusted architectural decoration image to separate the main physical structure regions in the adjusted architectural decoration image; Extract the central skeleton line for each of the main physical structure regions; Within the main physical structure region segmented by the semantic segmentation, the color values ​​of pixels are extracted, and the relative color relationships between the main color clusters are calculated to generate the aesthetic feature vector. The aesthetic feature vectors are normalized.

3. The method for analyzing architectural decoration images according to claim 2, characterized in that, The step of extracting pixel color values ​​within the main physical structure region segmented by semantic segmentation and calculating the relative color relationships between main color clusters includes: The distribution characteristics of pixel color values ​​within the main physical structure region segmented by the semantic segmentation are evaluated. The distribution characteristics evaluation includes calculating the number of histogram peaks of the pixel color values ​​and the distance between adjacent peaks. The number of main color clusters is dynamically determined based on the distribution characteristics assessment results; The primary color clusters are identified based on the number of primary color clusters; Among the identified primary color clusters, the relative color relationships between the primary color clusters are calculated.

4. The method for analyzing architectural decoration images according to claim 2, characterized in that, The adaptive brightness and contrast adjustment of the architectural decoration image includes: The architectural decoration image is separated into illumination and reflection components to obtain the illumination component and the reflection component. The illumination component is dynamically compressed to obtain the dynamically compressed illumination component. The reflection component is enhanced with detail to obtain the enhanced reflection component. The dynamic range compressed illumination component is fused with the detail enhanced reflection component to obtain the adjusted architectural decoration image.

5. The method for analyzing architectural decoration images according to claim 2, characterized in that, The step of performing semantic segmentation on the adjusted architectural decoration image to separate the main physical structure regions in the adjusted architectural decoration image includes: Multi-scale feature extraction is performed on the adjusted architectural decoration image to obtain visual information of the adjusted architectural decoration image at different resolutions; The multi-scale features are fused to generate a feature representation that includes contextual information; Based on the fused feature representation, the pixels of the adjusted architectural decoration image are classified to identify and separate the main physical structure regions; Based on the classification results, a segmentation mask for the main physical structure region is generated.

6. The method for analyzing architectural decoration images according to claim 5, characterized in that, After classifying the pixels of the adjusted architectural decoration image based on the fused feature representation, the method further includes: In the pixel classification process, structural boundary smoothing constraints are introduced to suppress responses to structural surface textures, patterns, or digital enhancement effects; The pixel classification process introduces structural boundary smoothing constraints to suppress responses to structural surface textures, patterns, or digital enhancement effects, including: The pixel classification results are divided into local regions to identify physical structure regions with different boundary smoothness requirements; For each of the physical structure regions, the local gradient distribution characteristics of the boundary are calculated, whereby the local gradient distribution characteristics represent the boundary sharpness or boundary ambiguity of the physical structure region. Based on the local gradient distribution characteristics, the constraint strength for the smoothness of the structural boundary of each physical structural region is dynamically determined. The constraint strength is negatively correlated with the boundary sharpness and positively correlated with the boundary ambiguity.

7. The method for analyzing architectural decoration images according to claim 2, characterized in that, The step of extracting the central skeleton line for each of the main physical structure regions includes: Geometric analysis based on the shape characteristics of the main physical structure region is performed to identify irregular boundary segments and complex geometric features in the main physical structure region; Based on the irregular boundary segments and the complex geometric features, the local smoothness parameters extracted from the skeleton lines are dynamically adjusted. These local smoothness parameters are used to control the sensitivity of the skeleton lines to local details. During the skeleton line extraction process, a local smoothness parameter is applied to the irregular boundary segment to suppress interference with surface details; The distance transformation is performed on the main physical structure region, and the central skeleton line is obtained based on the distance transformation result and the dynamically adjusted local smoothness parameter.

8. The method for analyzing architectural decoration images according to claim 7, characterized in that, The geometric analysis of the main physical structure region based on the region shape features to identify irregular boundary segments and complex geometric features in the main physical structure region includes: Multi-scale curvature analysis is performed on the boundary of the main physical structure region to obtain information on the curvature variation of the boundary of the main physical structure region at different scales. The curvature variation information at different scales is fused to generate a curvature feature representation that includes local details and global trends; Based on the fused curvature feature representation, an adaptive threshold is set to distinguish geometric noise from the irregular boundary segment or the complex geometric feature; Based on the differentiation results, the irregular boundary segments and the complex geometric features are identified.

9. The method for analyzing architectural decoration images according to claim 7, characterized in that, The step of performing a distance transformation on the main physical structure region and obtaining the central skeleton line based on the distance transformation result and the dynamically adjusted local smoothness parameter includes: A distance transformation is performed on the main physical structure region to obtain the distance transformation result; Based on the distance transformation result and the dynamically adjusted local smoothness parameter, the central skeleton line is obtained, which includes the connectivity and hole structure of the main physical structure region.

10. A system for analyzing architectural decoration images, characterized in that, The system includes: The input end is used to perform visual analysis processing on the input architectural decoration image, extract and quantify the aesthetic feature vector from the overall visual performance of the architectural decoration image, and the aesthetic feature vector includes color saturation distribution, texture complexity, light softness and contrast, spatial white space ratio, and line and shape features. The construction end is used to construct an association database of the aesthetic feature vectors and decorative element annotations. The association database stores the aesthetic feature vectors of professional image libraries and their corresponding decorative element annotations. Based on the aesthetic feature vectors, nearest neighbor matching is performed in the association database to obtain the aesthetic feature vectors of the professional image libraries that are closest to the aesthetic feature vectors. The matching end is used to reverse-solve the decorative element combination rules based on the nearest neighbor matching results. The reverse-solved decorative element combination rules include analyzing the decorative elements and their combination methods common to the images in the nearest professional image library; and generating decorative element combination suggestions corresponding to the decorative element combination rules.

Images