Intelligent identification and language conversion method for traditional building-riding facade component based on hierarchical coding system

Abstract

The invention discloses a hierarchical coding system-based intelligent recognition and language conversion method for traditional building-riding facade components, which realizes automatic coding and standardized description generation by constructing a six-level hierarchical coding system and training a special component recognition model, and lays a data foundation for evolution analysis of large-scale building-riding buildings. The method comprises the following steps: S1, constructing a six-level standard coding system; s2, collecting, preprocessing and labeling images, and constructing a labeling data set; s3, constructing a building-riding part identification model, and training and identifying the building-riding part identification model by using the existing data set and the annotation data set; s4, allocating coding information of a standard coding system for the recognition result of S3, extracting multi-dimensional features, performing component correlation analysis, and constructing a component feature database; s5, standard component description texts of three levels are automatically generated based on the steps S3 and 4; s6, constructing a structured knowledge graph of the apartment facade components based on the steps S3 to S5; and S7, integrating the processing capabilities in the steps S2 to S6 into a complete system.

Description

Technical Field

[0001] The present invention belongs to the field of digital protection of architectural cultural heritage and artificial intelligence technology, and in particular refers to a method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system. By constructing a standardized component coding system and training an artificial intelligence model, the automatic recognition, coding and standardized description generation of arcade facade components are achieved. Background Art

[0002] As an important cultural heritage, the facade features of traditional arcade buildings are the key basis for distinguishing arcade types from different periods and regions. However, existing technologies have the following deficiencies in the identification and standardized expression of arcade facade components: (1) Lack of a systematic component coding system Different researchers use different naming and classification methods for the same component. For example, "Shouliang" is called "lintel beam" or "menlin beam" in different regions, which makes it difficult to systematically organize architectural knowledge and cannot support large-scale computer processing and evolutionary analysis.

[0003] (2) Facade component recognition relies on manual experience Traditional methods require professionals to identify various components on the facade one by one, which is inefficient and highly subjective. Especially when faced with large-scale surveys of hundreds of arcades, manual identification methods are difficult to ensure consistency and accuracy.

[0004] (3) Lack of standardized language expression for component features Existing recording methods are mostly descriptive texts, lacking a structured coding system and cannot be directly used for computer analysis, which seriously restricts the subsequent architectural evolution analysis, protection and restoration work.

[0005] (4) Insufficient application of artificial intelligence technology Although deep learning has made breakthroughs in the field of image recognition, in terms of architectural component recognition, due to the lack of standardized training data and specialized recognition models, existing technologies have difficulty in accurately identifying the complex components of arcade facades.

[0006] Therefore, there is an urgent need for a method based on a standardized coding system to realize the automatic identification and coding of arcade facade components by training artificial intelligence models, providing a basis for subsequent building evolution analysis. Summary of the Invention

[0007] The main purpose of the present invention is to provide a method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system to solve the problems existing in the prior art. By constructing a six-level hierarchical coding system, training a special component recognition model, and realizing automatic coding and standardized description generation, it lays a data foundation for large-scale arcade building evolution analysis.

[0008] In order to achieve the above object, the solution of the present invention is: A method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system, comprising: Step S1. Construct a standard coding system with six levels: building system level coding, regional subsystem coding, category coding, subcategory coding, specific building coding, and component coding. Component coding uses a position-type-attribute sub-coding composite structure, with position-type forming the basic component coding. Step S2. Systematically collect images according to preset image acquisition standards to obtain a sufficient number of arcade facade images, and then pre-process the collected arcade facade images, and then manually annotate the pre-processed arcade facade images to construct an annotation dataset; Step S3. Build a model for identifying arcade components. Use existing datasets and annotated datasets to train and identify these components. The network architecture of this model uses Mask R-CNN as its foundational framework. Its backbone network uses ResNet-101 combined with a Feature Pyramid Network (FPN). This network extracts multi-scale features through bottom-up and top-down feature fusion, adapting to the size differences of arcade components, from small ornaments to large pediments. Furthermore, a spatial relationship module is added to the Mask R-CNN framework. This module uses a graph convolutional network to model the spatial constraints between different components. Step S4. For the recognition result of step S3, assign the coding information of the standard coding system and extract multi-dimensional features, perform component association analysis, and build a component feature database; Step S5. Based on the recognition results from step S3 and the encoded information from step S4, three levels of standardized component description text are automatically generated, achieving accurate conversion from visual features to natural language. These descriptions include a basic level description of less than 20 words, a standard level description of less than 50 words, and a specialized level description of less than 100 words. Step S6. Integrate the recognition result of step S3, the encoding information of step S4, and the description text of step S5 to construct a structured knowledge graph of the arcade facade components; Step S7. Integrate the various processing capabilities in steps S2 to S6 into a complete system, which specifically includes: implementing the image preprocessing capability of step S2, implementing the component intelligent recognition capability of step S3, implementing the automatic encoding and feature extraction capability of step S4, implementing the multi-level description generation capability of step S5, and implementing the knowledge graph construction capability of step S6, to achieve end-to-end automated processing through a unified system architecture.

[0009] The standard coding system in step S1 is specifically: (1) Building system-level coding, uniformly using CTQL as the top-level identifier for traditional arcades; (2) Regional sub-system codes are set according to the geographical areas where the arcades are distributed; (3) Group codes, used to distinguish regional characteristics; (4) Sub-group coding, accurate to the specific historical street; (5) Specific building code, which assigns a number of digits to each arcade; (6) Component code, its sub-code is: (1) Position: According to the structural logic of the arcade facade, R represents the roof, B represents the body, and F represents the seat; (2) Type, which is used to reflect the detailed classification of the location; (3) Attributes, including at least one of the material, age, and functional characteristics of the component, where M represents material, T represents age, and U represents function.

[0010] The image acquisition standards preset in step S2 are as follows: (1) The axis of the camera lens is perpendicular to the main plane of the building facade when shooting; (2) The shooting distance should be controlled within 1.5 to 2 times the building height; (3) Use a standard lens with an equivalent focal length of 35 to 50 mm; (4) Choose uniform lighting conditions for shooting; (5) The image resolution must be no less than 4000×3000 pixels.

[0011] The pre-processing process in step S2 includes: First, perspective correction is performed by detecting the vertical and horizontal lines in the facade, calculating the perspective transformation matrix, and correcting the oblique image to the standard orthographic projection; Secondly, color standardization is performed, using the standard color card placed during shooting to calculate the color correction matrix; Then, adaptive histogram equalization technique is used to enhance the image contrast; Finally, scale normalization is performed to adjust all images to a uniform pixel density.

[0012] In step S2, when constructing the annotated dataset, at least five architectural history experts with more than 10 years of experience are invited to participate in the annotation. The pre-processed arcade facade image is annotated on the annotation software, and the component outlines, component types, and material properties and preservation conditions are accurately selected on the image to finally obtain the annotated dataset. To ensure the annotation quality, a cross-validation mechanism is adopted during the annotation process. Each component is independently annotated by at least two experts, and any disagreements are reached through discussion.

[0013] The training strategy of the arcade component recognition model in step S3 adopts the following three-stage progressive method: In the first stage, the COCO dataset is used for pre-training to enable the arcade component recognition model to learn basic visual feature extraction capabilities; In the second stage, domain adaptation is performed using a building image dataset to enable the arcade component recognition model to understand the unique visual patterns of building components. In the third stage, the labeled dataset of step S2 is used for fine training, focusing on optimizing the arcade component recognition model's ability to recognize arcade-specific components.

[0014] Preferably, the training process of the arcade component recognition model in step S3 adopts the following optimization measures: (1) Using the focal loss function to solve the problem of component category imbalance, higher loss weights are given to rare component categories with a smaller number; (2) Introducing an online hard example mining mechanism, presetting easily confused component pairs in arcade components based on practical experience, automatically identifying these component pairs during training, and increasing the training frequency of these component pairs according to a preset number or ratio value; (3) Through data enhancement technology, the training samples are expanded three times to improve the generalization ability of the arcade component recognition model.

[0015] Preferably, a post-processing mechanism is established in step S3 to ensure the rationality of the recognition result, specifically: (1) Confidence screening is performed on the detection boxes output by the arcade component recognition model, and only the results with a confidence level higher than 0.8 are retained; (2) Use the non-maximum suppression algorithm to eliminate overlapping detection frames to avoid repeated recognition of the same part; (3) Build a logic rule library based on architectural common sense and verify the rationality of the recognition results, such as checking whether the relative position relationship of the components conforms to the architectural construction logic.

[0016] The process of allocating the standard code in step S4 is as follows: First, determine the basic code based on the identified component type and its position on the arcade facade; By analyzing the visual characteristics of the components, we can infer their possible construction age; Combined with the geographical information of the image shooting location, the region-related code is automatically filled in; Determine the material properties of components through a pre-trained material classification network.

[0017] Preferably, the multidimensional features extracted in step S4 cover the following dimensions: (1) Geometric characteristics; (2) Visual features; (3) stylistic features; (4) Material characteristics.

[0018] Preferably, the steps of component association analysis in step S4 are: (1) Spatial relationship modeling: Based on the bounding box coordinates of the component recognition results, the spatial distance matrix and relative position relationship between the components are calculated to establish a component spatial relationship diagram; (2) Statistical co-occurrence probability: Calculate the co-occurrence frequency of different component types based on a large number of samples and construct a component co-occurrence probability matrix; (3) Co-occurrence pattern mining: Statistically analyze the co-occurrence frequency of different component types in the arcade samples, and use association rule mining algorithms to discover the combination rules of strongly associated components; (4) Verification of age consistency: Based on the inferred age attributes of the components, the age consistency of different components of the same building is verified to identify possible later reconstruction parts; (5) Constructing a component relationship network: Using components as nodes and association strength as edge weights, a weighted undirected graph is constructed for subsequent architectural style analysis and integrity verification.

[0019] Preferably, in step S4, the constructed component feature database is stored in MongoDB; each identified component creates an independent record in the database, which includes a unique identifier, a complete six-level code, precise coordinates in the original image, various extracted feature vectors, recognition confidence scores, and processing timestamp information; by establishing a multidimensional index, fast retrieval based on coding, feature similarity, and spatial location is supported.

[0020] The following quality control mechanism is introduced in step S5: (1) Architectural Grammar Checker: Based on the traditional dependency syntax analysis, it adds architectural professional expression standard checking; (2) Architectural terminology consistency check module: Construct a mapping table of traditional arcade professional terms and uniformly map synonyms into standard terms; (3) Architectural logic validator: establishes architectural domain-specific rules to verify the historical logical correctness of the description; (4) Architectural text readability evaluation algorithm: Based on the characteristics of architectural professional texts, the indicators are set as follows: the density of professional terms does not exceed 30% and the average length of sentences does not exceed 25 words.

[0021] The ontology of the knowledge graph in step S6 defines the following six core entities: (1) Building entity, recording the basic information of the entire arcade; (2) Component entities, which store detailed properties of various building components; (3) Material entities, describing the physical and chemical properties of different building materials; (4) Period entities, which divide historical eras and associate era characteristics; (5) Regional entity, recording geographic spatial information; (6) Craft entities, recording traditional and modern construction techniques; The relationships between entities include: (1) Compositional relationships, describing the whole-part relationships between buildings and components; (2) Positional relationship, recording the spatial distribution of components on the facade; (3) Material relationship, connecting components and the building materials they use; (4) Time relationship, marking the construction and modification history of the component; (5) Process relationship, which describes the manufacturing technology of the components; (6) Evolutionary relationships, tracking the historical changes in component styles.

[0022] Preferably, in step S6, the knowledge graph integrates multi-source data through a knowledge extraction and fusion process, specifically: (1) Automatically extract component entities and basic attributes from the recognition results of step S3 to create the initial nodes of the knowledge graph; (2) Establishing connecting edges between nodes by analyzing the spatial adjacency and functional association between components; (3) Supplement the historical and cultural information of components by integrating external data sources; (4) Use entity alignment technology to identify and merge different representations pointing to the same object to ensure the consistency and completeness of knowledge.

[0023] Preferably, in step S6, the knowledge graph has an inference rule engine to support complex knowledge discovery, specifically: (1) Reasoning mechanism based on description logic; (2) Temporal reasoning, analyzing the evolution path of component styles; (3) Spatial reasoning: discovering the distribution patterns of components.

[0024] Preferably, the knowledge graph in step S6 has a knowledge service interface, providing a variety of access methods, specifically: (1) SPARQL query interface; (2) Natural language query interface; (3) Visual interface; (4) RESTful API.

[0025] The system in step S7 adopts an end-to-end processing flow and an asynchronous processing mechanism to support parallel processing of batch images and ensure efficient operation of the system. Specifically: When the user uploads an image of an arcade facade, the system automatically starts the processing pipeline, first performing image quality inspection and preprocessing, then calling the trained arcade component recognition model to identify the components inside the arcade, then assigning coding information to each component and extracting multi-dimensional features, and then generating three levels of component description text, and finally storing it in the knowledge graph and updating the index.

[0026] Preferably, the system in step S7 adopts the following technical means to optimize system performance: (1) Convert 32-bit floating-point model parameters into 8-bit integer representation through model quantization technology; (2) Use TensorRT to optimize the inference engine and fully utilize the parallel computing capabilities of the GPU; (3) Implement distributed architecture design to support multi-node collaborative processing; (4) Establish an intelligent caching mechanism to cache frequently accessed feature data and query results.

[0027] Preferably, the system in step S7 continuously improves system performance through a continuous optimization mechanism, specifically: (1) Establish a user feedback collection system; (2) Organize expert reviews regularly; (3) Expand the training dataset; (4) Update and optimize the recognition model; (5) Improve the content of the knowledge graph.

[0028] After adopting the above technical solution, the present invention has the following technical effects: (1) The degree of standardization has been greatly improved: the standardized expression of arcade components has been achieved through the six-level coding system, which has completely solved the problems of confusing terminology and inconsistent classification in traditional research; standardized coding has laid the foundation for comparative research across regions and periods.

[0029] (2) Revolutionary improvement in recognition efficiency: The complete recognition, coding, and description generation of a single arcade building takes only 30 seconds, which is more than 100 times more efficient than traditional manual methods. This makes large-scale building surveys and digital archiving possible.

[0030] (3) Knowledge expression is more comprehensive and accurate: The automatically generated three-level description text provides both concise basic information and professional and in-depth detailed analysis, meeting the needs of users at different levels. Standardized language expression facilitates the dissemination and sharing of knowledge.

[0031] (4) Providing a data basis for subsequent research: Standardized coded data can be directly input into the evolution analysis system of arcade buildings, supporting large-scale research on the evolution laws of architectural styles; the knowledge graph provides a rich component library and combination rules for the intelligent design and reconstruction of other related systems.

[0032] (5) Promote the digital transformation of cultural heritage protection: The technical solution provided by the present invention can be extended and applied to the digital protection of other types of traditional buildings, accelerating the digitalization process in the entire field of cultural heritage protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 2 is a system architecture diagram of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to further explain the technical solution of the present invention, the present invention is described in detail below through specific embodiments.

[0035] refer to Figure 1 As shown, the present invention discloses a method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system, comprising: Step S1. Construct a standard coding system with six levels, namely: building system level code, regional sub-system code, category code, sub-category code, specific building code, and component code. Component coding adopts a position-type-attribute sub-coding composite structure. Position-type constitutes the basic coding of the component, which is used to reflect the actual position of the component on the arcade (similar to coordinates).

[0036] This step establishes a standard coding system from macro to micro to achieve standardized expression of arcade facade components; the design of the coding system follows the principles of hierarchical progression, unique identification, and easy expansion.

[0037] Specifically, in the above standard coding system: (1) Building system-level coding, uniformly using CTQL as the top-level identifier for traditional arcades; (2) Regional sub-lineage coding: MN represents the Minnan sub-lineage, LN represents the Lingnan sub-lineage, XJP represents the Singapore sub-lineage, MLXY represents the Malaysian sub-lineage, etc., according to the geographical areas where the arcades are distributed; (3) Group code, used to distinguish regional characteristics. For example, under the Minnan subfamily, XM represents the Xiamen group, QZ represents the Quanzhou group, and ZZ represents the Zhangzhou group. Under the Lingnan subfamily, CS represents the Chaoshan group, GZ represents the Guangzhou group, HN represents the Hainan group, and GX represents the Guangxi group. (4) Sub-group coding, accurate to the specific historical streets, such as LCQ-ZS for Quanzhou Licheng District-Zhongshan Road, AX-LY for Anxi-Luoyang Street, etc.; (5) Specific building code, which assigns a number of digits to each arcade (usually three digits, but not limited to this); (6) Component code, its sub-code is: (1) Position: According to the structural logic of the arcade facade, R represents the roof, B represents the body, and F represents the seat; (2) Type, which is used to reflect the detailed classification of the location, for example: in the roof part, R01 is set to represent the parapet R01, R02 is set to represent the pediment, R03 is set to represent the eaves, etc.; in the body part, B01 is set to represent the arcade column, B02 is set to represent the arcade lintel, B03 is set to represent the second floor wall, B04 is set to represent the window, B05 is set to represent the door, etc.; in the seat part, F01 is set to represent the column base, F02 is set to represent the ground, F03 is set to represent the steps, etc.; (3) Attributes, including at least one of the material, age and functional characteristics of the component. M is set to represent material, T is set to represent age, and U is set to represent function. For example, in the material attribute code, M01 is set to represent wood, M02 is set to represent brick and stone, M03 is set to represent concrete, M04 is set to represent metal, M05 is set to represent ceramics, etc. In the age attribute code, T01 is set to represent the early Qing Dynasty, T02 is set to represent the middle Qing Dynasty, T03 is set to represent the late Qing Dynasty, T04 is set to represent the early modern period, T05 is set to represent the late modern period, and T06 is set to represent the contemporary period. In the functional attribute code, U01 represents load-bearing, U02 represents decoration, U03 represents maintenance, U04 represents use, and U05 represents connection, etc.

[0038] By combining the above-mentioned codes at all levels, each component of the arcade can obtain a globally unique standard code. For example, the standard code CTQL-MN-QZ-LCQ-ZS-028-R02-M03-T04-U02 can fully express the complex information of "traditional arcade - Southern Fujian sub-system - Quanzhou group - Licheng sub-group - Zhongshan Road - Building 028 - pediment component - concrete material - early modern period - decorative function".

[0039] To ensure the comprehensibility and readability of the code, a bidirectional mapping dictionary between code and natural language can be established according to certain rules. This dictionary can be used to collect professional terms related to arcades in existing technologies. Each code entry can include the standard Chinese name, a set of local colloquial names, the English equivalent, a functional description, and the context of use. This allows for seamless conversion between machine code and human language.

[0040] Step S2. Perform systematic collection according to the preset image collection standard to obtain arcade facade images that meet the quantity requirements, then preprocess the collected arcade facade images, and then manually annotate the preprocessed arcade facade images to construct an annotated dataset.

[0041] This step aims to establish a standardized image acquisition and preprocessing process to obtain high-quality arcade facade images and have them annotated by experienced and knowledgeable humans, thereby obtaining an annotated dataset that can be used for training and ensuring that the model trained in subsequent steps can accurately identify facade components.

[0042] Specifically, in the above step S2, the preset image acquisition standards are as follows: (1) When shooting, the axis of the camera lens should be perpendicular to the main plane of the building facade to ensure that a standard orthographic projection image is obtained; (2) The shooting distance should be controlled at 1.5 to 2 times the building height, ensuring that the facade is fully captured while maintaining sufficient detail clarity; (3) Use a standard lens with an equivalent focal length of 35 to 50 mm to avoid perspective distortion caused by wide-angle lenses; (4) Choose evenly lit conditions such as early morning or evening on a cloudy or sunny day to avoid strong shadows that may interfere with component recognition; (5) The image resolution must be no less than 4000 × 3000 pixels to ensure the identifiability of component details.

[0043] In the above step S2, when performing systematic collection: a certain number of arcades are selected for photography in different areas with arcade buildings in the southern Fujian region, the Lingnan region, the Hainan region, etc. For example, 150 arcades in six representative historical blocks, including Quanzhou Zhongshan Road, West Street, and East Street, Xiamen Zhongshan Road, Gulangyu Island, and Zhangzhou Old Town, are selected in the southern Fujian region; and 150 buildings in 10 arcade blocks, including Guangzhou Beijing Road, Shangxiajiu Street, Jiangmen Sanshisanxu Street, Haikou Zhongshan Road, and Wenchang Wennan Old Street, are selected in the Lingnan region; at least 3 panoramic photos of the front facade and 10 to 15 partial close-up photos are collected for each arcade, among which the partial close-ups focus on photographing characteristic components such as pediments, capitals, arches, and decorative patterns; each arcade is photographed at multiple time points to obtain diverse lighting conditions.

[0044] In the above step S2, the pre-processing process includes: First, perspective correction is performed by detecting the vertical and horizontal lines in the facade, calculating the perspective transformation matrix, and correcting the oblique image to the standard orthographic projection; Secondly, color standardization is performed. Using a standard color card placed during shooting, a color correction matrix is ​​calculated to ensure consistent color performance in images taken under different lighting conditions. Then, adaptive histogram equalization technology is used to enhance the image contrast and make the details of the components in the shadow area clearer; Finally, scale normalization is performed to adjust all images to a uniform pixel density to facilitate subsequent feature extraction and comparison.

[0045] In step S2 above, when constructing the annotated dataset, at least five architectural history experts with more than 10 years of experience were invited to participate in the annotation. The preprocessed arcade facade images were annotated using the annotation software. The component outlines, component types, and material properties and preservation conditions were accurately selected on the images to obtain the annotated dataset. To ensure the quality of the annotation, a cross-validation mechanism was used during the annotation process. Each component was independently annotated by at least two experts, and any disagreements were resolved through discussion.

[0046] Step S3. Construct a recognition model for arcade components, and use existing datasets and labeled datasets to train and recognize the recognition model for arcade components; Figure 1 As shown in the figure, the network architecture design of the arcade component recognition model uses Mask R-CNN as its basic framework, which can simultaneously perform object detection, classification, and instance segmentation tasks. The backbone network of the arcade component recognition model uses ResNet-101 combined with the Feature Pyramid Network (FPN). Through bottom-up and top-down feature fusion, it can effectively extract multi-scale features and adapt to the size differences of arcade components, from small decorations to large pediments. In addition, the basic framework of Mask R-CNN also innovatively adds a spatial relationship module. The spatial relationship module uses a graph convolutional network (GCN) to model the spatial constraints between different components, such as "pediment must be located at the top of the facade" and "column base must be located at the bottom of the column." Therefore, the introduction of spatial relationships significantly improves the arcade component recognition model's ability to judge the rationality of component placement, thereby reducing misidentification.

[0047] Specifically, in step S3 above, the training strategy of the arcade component recognition model adopts a three-stage progressive method: In the first stage, the COCO dataset is used for pre-training to enable the arcade component recognition model to learn basic visual feature extraction capabilities; In the second stage, domain adaptation is performed using a building image dataset to enable the arcade component recognition model to understand the unique visual patterns of building components. In the third stage, the labeled dataset of step S2 is used for fine training, focusing on optimizing the arcade component recognition model's ability to recognize arcade-specific components.

[0048] In step S3 above, the training process of the arcade component recognition model adopts a number of optimization measures to address the technical difficulties in training, specifically: (1) Using the focal loss function to solve the problem of component category imbalance, higher loss weights are given to rare component categories with a smaller number; (2) Introducing an online hard example mining mechanism, based on practical experience, we preset easily confused component pairs in arcade components (e.g., “Baroque pediment” and “Neoclassical pediment”), automatically identifying these component pairs during training, and increasing the training frequency of these component pairs according to a preset number or ratio; (3) Through data enhancement technology (including random cropping, color jittering, geometric transformation, etc.), the training samples are expanded by 3 times to improve the generalization ability of the arcade component recognition model.

[0049] In step S3 above, a comprehensive post-processing mechanism is established to ensure the rationality of the recognition results, specifically: (1) Confidence screening is performed on the detection boxes output by the arcade component recognition model, and only the results with a confidence level higher than 0.8 are retained; (2) Use the non-maximum suppression algorithm to eliminate overlapping detection frames to avoid repeated recognition of the same part; (3) Build a logic rule library based on architectural common sense and verify the rationality of the recognition results, such as checking whether the relative position relationship of the components conforms to the architectural construction logic.

[0050] Step S4. For the recognition result of step S3, assign the coding information of the standard coding system and extract multi-dimensional features, perform component association analysis, and build a component feature database; The process of assigning standard codes is as follows: First, a base code is determined based on the identified component type and its location on the arcade facade. Visual features of the component, such as decorative complexity, geometry, and texture, are analyzed to infer its likely construction age. Geographic information from the image's location is incorporated to automatically populate region-specific codes (i.e., regional subfamily codes, group codes, subgroup codes, and specific building codes). Finally, a pre-trained material classification network is used to determine the component's material properties. This network utilizes a general deep learning architecture and is specifically trained and optimized for the unique material systems of traditional arcade architecture (including Minnan red bricks, cement plaster, washed stone, and carved wood). This network, comprising over 1,000 samples, is used to construct a dataset of arcade building materials, enabling it to accurately identify the unique material types found in arcade architecture.

[0051] Specifically, in step S4 above, the multidimensional features extracted cover the following dimensions: (1) Geometric characteristics, including the absolute size and relative proportions of components, which are calibrated by reference to known dimensions of buildings or standard component sizes; (2) Visual features, extracted using a variety of operators, including local binary patterns for texture analysis, scale-invariant feature transforms for key point detection, and HSV color histograms for color feature extraction; (3) Style features: extracted through a deep style network to generate a 128-dimensional feature vector, which can effectively distinguish different decorative styles such as Southern Fujian traditional, Baroque, neoclassical, and Art Deco; (4) Material features, determined by a pre-trained material classification network that is based on general texture recognition technology but adapted to the domain: a. A dataset dedicated to traditional arcade building materials was constructed, containing 87 common types of arcade building materials, covering traditional materials (such as Southern Fujian red bricks, bluestone, and fir wood) and modern materials (such as cement, reinforced concrete, and washed stone); b. Introduced material aging recognition, distinguishing between new and old materials through features such as texture roughness and color fading, to assist in age determination; c. Build a knowledge base of material-era associations, such as rules like "stonewashing was introduced in the 1920s" and "mosaic decoration became popular in the 1950s," to improve the accuracy of era judgments; d. A material coding mapping table was established to directly output standardized material codes (M01-M05, etc.), achieving seamless integration with the six-level coding system.

[0052] In the above step S4, the steps of component association analysis are: (1) Spatial relationship modeling: Based on the bounding box coordinates of the component recognition results, the spatial distance matrix and relative position relationships (up, down, left, right, included, adjacent, etc.) between components are calculated to establish a component spatial relationship diagram; (2) Statistical co-occurrence probability: Based on a large number of samples, the co-occurrence frequency of different component types is counted to construct a component co-occurrence probability matrix, such as the probability of "Baroque pediment" and "Corinthian capital" appearing at the same time; (3) Co-occurrence pattern mining: Statistically analyze the co-occurrence frequency of different component types in the arcade samples, and use the association rule mining algorithm (Apriori algorithm, minimum support 0.3, minimum confidence 0.7) to discover the combination rules of strongly associated components; (4) Verification of age consistency: Based on the inferred age attributes of the components, the age consistency of different components of the same building is verified to identify possible later reconstruction parts; (5) Constructing a component relationship network: Using components as nodes and association strength as edge weights, a weighted undirected graph is constructed for subsequent architectural style analysis and integrity verification.

[0053] As mentioned above, component association analysis is an important innovation of this step. The present invention not only identifies individual components, but also analyzes the combination patterns of different components on the same facade. Through statistical analysis, it was found that certain component combinations have a strong correlation, such as the co-occurrence probability of "red brick square columns" and "arched corridors" reaching 85%, and the co-occurrence probability of "Corinthian capitals" and "Baroque pediments" reaching 73%; these association rules not only verify the rationality of the recognition results, but also provide an important basis for the subsequent analysis of the evolution of architectural styles.

[0054] In step S4 above, the component feature database constructed is stored in MongoDB, leveraging its excellent support for unstructured data. Each identified component is created as a separate record in the database, containing a unique identifier, a complete six-level code, its precise coordinates in the original image, various extracted feature vectors, recognition confidence scores, and a processing timestamp. Multidimensional indexing supports rapid retrieval based on various methods, including code, feature similarity, and spatial location.

[0055] Step S5. Based on the recognition results of step S3 and the coding information of step S4, three levels of standardized component description text are automatically generated to achieve accurate conversion from visual features to natural language, including a basic layer description of less than 20 words, a standard layer description of less than 50 words, and a professional layer description of less than 100 words; the basic layer description organizes the language in the order of "component name-material properties-location information", which is mainly used for rapid identification and retrieval; the standard layer adds more elements such as era characteristics, regional style, size specifications, decorative features, etc. on the basis of the basic layer, and automatically extracts key information from the feature data when generating the description and organizes the language according to the preset word order rules; the professional layer adds deep information such as process technology, cultural connotation, historical value, etc. on the basis of the standard layer to generate a description text with academic depth. For example: for an arcade column component, the basic layer generates a concise and clear description such as "wooden square column - east side of the arcade", and the standard layer generates a detailed description such as "Quanzhou-style wooden square column in the mid-Qing Dynasty, with a cross-section of 20×20 cm, the capital has simple molding decoration, and is well preserved"; for a Baroque pediment component, the professional layer generates an in-depth description such as "The pediment was built in 1923, using cement plaster technology, decorated with scroll patterns and wreath reliefs in the center, and symmetrical scroll decorations on both sides. The overall style is typical of Nanyang Baroque, reflecting the historical characteristics of commercial buildings in southern Fujian in the early 20th century under the influence of Western culture, and has important architectural history research value."

[0056] In step S5 above, the description text generation process combines rule templates and deep learning methods: for common standard components, predefined language templates are used to ensure the standardization and accuracy of the description; for complex or special components, a large language model that has been fine-tuned in the architectural field is called, and domestic mainstream models such as Alibaba Tongyi Qianwen, Baidu Wenxin Yiyan, Zhipu ChatGLM or DeepSeek are used as the basis. LoRA is fine-tuned through the collection of 100,000 architectural professional documents to enhance the model's ability to understand architectural professional knowledge and generate more flexible and expressive description text.

[0057] A quality control mechanism is introduced in step S5 to ensure the professionalism and readability of the generated description. Although some existing natural language processing technologies are used, the present invention innovatively combines and improves these technologies to specifically adapt to the characteristics of architectural description texts. It constructs an architectural professional terminology library, a domain-specific logical rule library, and readability evaluation standards, forming a complete architectural description text quality control system. Specifically, (1) Architectural Grammar Checker: Based on traditional dependency syntax analysis, it adds architectural professional expression standard checks, such as spatial logic verification such as "the capital is located above the column body", to ensure that the description conforms to common sense of architectural construction; (2) Architectural terminology consistency check module: A mapping table of traditional arcade professional terms containing more than 3,000 entries was constructed, and synonyms such as "shouliang / lintel beam / menlin beam" were uniformly mapped to standard terms to ensure the standardization of descriptions; (3) Architectural logic validator: 50+ architectural domain-specific rules have been established, such as “Baroque decoration appeared no earlier than 1890” and “RC structure appeared no earlier than 1910”, to verify the historical logic correctness of the description; (4) Architectural text readability evaluation algorithm: Based on the characteristics of architectural professional texts, the professional term density is set to no more than 30%, and the average sentence length is set to no more than 25 words, to ensure that it is easy to understand while maintaining professionalism.

[0058] Step S6. Integrate the recognition results from step S3, the encoding information from step S4, and the description text from step S5 into a structured knowledge graph of the arcade facade components to support complex queries and knowledge reasoning. The knowledge graph ontology defines six core entities: (1) Building entity, recording the basic information of the entire arcade; (2) Component entities, which store detailed properties of various building components; (3) Material entities, describing the physical and chemical properties of different building materials; (4) Period entities, which divide historical eras and associate era characteristics; (5) Regional entity, recording geographic spatial information; (6) Craft entities, recording traditional and modern construction techniques; The relationships between entities include: (1) Compositional relationships, describing the whole-part relationships between buildings and components; (2) Positional relationship, recording the spatial distribution of components on the facade; (3) Material relationship, connecting components and the building materials they use; (4) Time relationship, marking the construction and modification history of the component; (5) Process relationship, which describes the manufacturing technology of the components; (6) Evolutionary relationships, tracking the historical changes in component styles. Each relationship has a clear attribute definition, such as the composition relationship contains attributes such as quantity, position, and importance.

[0059] In step S6 above, the knowledge graph integrates multi-source data through the knowledge extraction and fusion process, specifically: (1) Automatically extract component entities and basic attributes from the recognition results of step S3 to create the initial nodes of the knowledge graph; (2) Establishing connecting edges between nodes by analyzing the spatial adjacency and functional association between components; (3) Supplement the historical and cultural information of components by integrating external data sources such as historical documents and expert knowledge; (4) Use entity alignment technology to identify and merge different representations pointing to the same object to ensure the consistency and completeness of knowledge.

[0060] In step S6 above, the knowledge graph has an inference rule engine that supports complex knowledge discovery, specifically: (1) The reasoning mechanism based on description logic can deduce implicit information from explicit knowledge; for example, the rule “if a component uses Baroque decoration and is located in southern Fujian, then the construction period is likely to be 1890-1930” can help infer missing chronological information; (2) Temporal reasoning, analyzing the evolutionary path of component styles, such as the evolutionary sequence of "Doric order → Corinthian order → Composite order"; (3) Spatial reasoning: discovering the distribution patterns of components, such as “arcades in commercial streets mostly adopt open designs, while arcades in residential areas mostly adopt closed designs.”

[0061] In step S6 above, the knowledge graph has a knowledge service interface that can provide a variety of access methods, specifically: (1) SPARQL query interface, which supports complex graph query operations and allows professional users to write precise query statements; (2) Natural language query interface, which converts users' natural language questions into structured queries through semantic parsing technology; (3) Visual interface, using force-directed graph and other technologies to intuitively display the relationship network between components; (4) RESTful API, supports the integration of third-party applications, and provides standardized data output in JSON format.

[0062] Step S7. System integration and practical application The various processing capabilities in steps S2 to S6 above are integrated into a complete system, which specifically includes: image preprocessing capability (step S2), component intelligent recognition capability (step S3), automatic encoding and feature extraction capability (step S4), multi-level description generation capability (step S5), and knowledge graph construction capability (step S6), achieving end-to-end automated processing through a unified system architecture.

[0063] Specifically, the system in step S7 above adopts an end-to-end processing flow and an asynchronous processing mechanism to support parallel processing of batch images and ensure efficient operation of the system. Specifically: When the user uploads an image of an arcade facade, the system automatically starts the processing pipeline, first performing image quality inspection and preprocessing, then calling the trained arcade component recognition model to identify the components inside the arcade, then assigning coding information to each component and extracting multi-dimensional features, and then generating three levels of component description text, and finally storing it in the knowledge graph and updating the index.

[0064] The system in step S7 above uses the following technical means to optimize system performance: (1) Convert 32-bit floating-point model parameters into 8-bit integer representations through model quantization technology, increasing the inference speed by 3 times while maintaining recognition accuracy; (2) Using TensorRT to optimize the inference engine and fully utilize the parallel computing capabilities of the GPU, the complete processing time of a single image is controlled within 30 seconds; (3) Implement distributed architecture design to support multi-node collaborative processing and meet large-scale data processing needs; (4) Establish an intelligent caching mechanism to cache frequently accessed feature data and query results, significantly improving system response speed.

[0065] The system in step S7 above continuously improves system performance through continuous optimization mechanism: (1) Establish a user feedback collection system to record identification errors and improvement suggestions found during use; (2) Organize expert reviews regularly to conduct professional evaluations of system output results and formulate optimization plans; (3) Expand the training dataset, especially by adding newly discovered rare component samples and edge cases; (4) Update and optimize the recognition model, use new data for incremental learning, and improve the ability to handle complex situations; (5) Improve the content of the knowledge graph and continuously add new component associations and inference rules.

[0066] In practical applications, the system's performance was comprehensively evaluated through large-scale validation testing. Specifically, 1,000 arcade buildings were selected from 10 representative historical arcade districts and subjected to comprehensive recognition, coding, and description generation testing. Test results showed that component recognition accuracy reached 92.5%, with recognition accuracy exceeding 95% for common component types. Coding accuracy reached 96%, with the vast majority of components accurately assigned to level 6 codes. Expert assessment of the rationality of description generation exceeded 90%, demonstrating that the generated text conforms to professional expression standards.

[0067] The present invention achieves important technical innovations in the following aspects: (I) Innovation of the six-level coding system For the first time, a complete coding system has been established, from building systems to specific components, with each component receiving a globally unique identifier. This coding system not only incorporates location and type information but also incorporates multi-dimensional attributes such as material, age, and function, enabling a structured representation of architectural knowledge. Compared to traditional simple classification methods, this coding system offers greater expressiveness and scalability.

[0068] (II) Innovation of specialized component identification models To address the unique characteristics of arcade facade components, a spatial relationship module was innovatively added to the standard object detection network. This module uses a graph convolutional network to model the spatial constraints between components, significantly improving recognition accuracy and rationality. This deep learning approach, incorporating domain knowledge, provides new insights for object recognition in other specialized fields.

[0069] (III) Innovation of intelligent description generation mechanism This system automatically converts image recognition results into standardized text descriptions, addressing the "semantic gap" that has long plagued architectural digitization. By combining rule-based templates and deep learning, it ensures both standardized descriptions and flexibility and expressiveness. The three-tiered description system meets the needs of diverse users.

[0070] (IV) Knowledge Graph-Driven Intelligent Analysis Innovation A comprehensive knowledge graph of arcade components has been constructed, which not only stores static component information but also supports dynamic knowledge discovery through an inference rule engine. This knowledge organization method breaks through the limitations of traditional databases and provides a powerful analytical tool for architectural history research and preservation.

[0071] (V) Field-Specific Innovation of General Technologies This invention innovatively transforms general computer vision technology into a domain-specific one: - Material classification is not a simple application of existing models, but a knowledge system of traditional arcade materials is constructed, including a complete mapping of material types, craft characteristics, and era characteristics; -Established a reasoning chain of "visual features → material recognition → historical period → code generation", achieving a deep integration of technology and domain knowledge; - Through training with over 1,000 specialized samples, the general technology can now identify traditional crafts such as stone washing, axe stone chopping, and stone grinding; -Material recognition results directly serve the generation of six-level codes and the inference of building age, reflecting system integration innovation.

[0072] Through the above scheme, the present invention has the following significant beneficial effects compared with the prior art: (1) The degree of standardization has been greatly improved The six-level coding system has achieved standardized expression of arcade components, completely solving the problems of confusing terminology and inconsistent classification in traditional research; standardized coding has laid the foundation for comparative research across regions and periods.

[0073] (2) Revolutionary improvement in recognition efficiency The complete identification, coding, and description generation of a single arcade building takes just 30 seconds, which is over 100 times more efficient than traditional manual methods. This makes large-scale building surveys and digital archiving possible.

[0074] (3) Knowledge expression is more comprehensive and accurate The automatically generated three-level description text provides both concise basic information and in-depth professional analysis, meeting the needs of users at different levels. Standardized language expression facilitates the dissemination and sharing of knowledge.

[0075] (4) Provide data basis for subsequent research Standardized coded data can be directly input into the evolution analysis system of arcade buildings, supporting large-scale research on the evolution laws of architectural styles; the knowledge graph provides a rich component library and combination rules for the intelligent design and reconstruction of other related systems.

[0076] (5) Promoting the digital transformation of cultural heritage protection The technical solution provided by the present invention can be extended and applied to the digital protection of other types of traditional buildings, accelerating the digitalization process in the entire field of cultural heritage protection.

[0077] The following are specific embodiments of the present invention. [1] Taking the arcade of Li Gan's fragrant oil shop at No. 287, Section 1, Dihua Street, Taipei City as an example, the implementation process and effects of the present invention are described in detail. The building where Li Gan's fragrant oil shop is located is located at No. 287, Section 1, Dihua Street, Datong District, Taipei City. It was built in the 1870s and is an imitation Western-style arcade building with a history of about 150 years. The first-generation owner, Li Jingyue, purchased this building in the 1920s and moved the oil shop from Hsinchu to Dihua Street, specializing in sesame oil and camellia oil. The fragrant sesame oil produced by it was awarded a certificate of merit by the Hsinchu State Industrial Advanced Association in 1926. The building is a two-story arcade structure with a facade width of about 4.5 meters. It is decorated with southern Fujian bricks and cement. It has the typical arcade characteristics of the late Qing Dynasty and is an important historical building on Dihua Street.

[0078] 1. Image Acquisition and Preprocessing Using a SLR camera with a 35mm F1.8 lens, the images were taken from a distance of 7 meters from the building facade. Even lighting conditions were chosen on an overcast day to avoid strong shadows. Three high-resolution 4000×3000 pixel images of the facade and 12 close-up photos were captured, focusing on key features such as the cement plaque-shaped store name sign, the vase-shaped decorative railings, the round columns, the arched pavilion legs, and the wooden window frames.

[0079] During the image preprocessing phase, perspective correction was first performed by detecting vertical lines on the facade, eliminating a skew of approximately 2.3 degrees. Color correction was performed using a standard color chart to ensure the accurate reproduction of the reddish-brown color of the Minnan bricks and the off-white color of the cement decoration. Adaptive histogram equalization was used to enhance the visibility of details in the shadowed areas within the pavilion's footstools. Finally, all images were resized to a standard resolution of 3000×2000 pixels.

[0080] 2. Component Recognition and Automatic Coding The pre-processed image was fed into the trained arcade component recognition model, and the system completed the entire recognition task within 24 seconds. The main components identified and their automatically generated codes are as follows: The cement plaque shop sign received the code "CTQL-TW-NTW-TBS-DH-287-B07-M03-T03-U02," indicating that it was a sign for the arcade at No. 287 Dihua Street in Taipei. Made of concrete, it was built in the late Qing Dynasty and served a decorative purpose. The recognition confidence level was 0.95.

[0081] The two circular columns were coded "CTQL-TW-NTW-TBS-DH-287-F04a-M02-T03-U01" and "CTQL-TW-NTW-TBS-DH-287-F04b-M02-T03-U01," respectively, accurately identifying their masonry structure and load-bearing function. The identification confidence level was 0.96.

[0082] The round arch pavilion leg was coded "CTQL-TW-NTW-TBS-DH-287-F05-M02-T03-U01," and the system correctly identified it as an arched structure made of Minnan bricks. The recognition confidence level was 0.94.

[0083] The vase-shaped decorative railing was given the code "CTQL-TW-NTW-TBS-DH-287-B08-M03-T03-U02", which accurately reflects its decorative properties.

[0084] The three wooden windows on the second floor were coded as follows: the central window "CTQL-TW-NTW-TBS-DH-287-B04a-M01-T03-U04" (left-right opening), and the two side windows "CTQL-TW-NTW-TBS-DH-287-B04b-M01-T03-U04" and "CTQL-TW-NTW-TBS-DH-287-B04c-M01-T03-U04" (top-bottom opening).

[0085] 3. Feature Extraction Results The geometric features extracted by the system show that the building's facade is 4.5 meters wide, the total height of the two floors is 8.5 meters, and the pavilion foot is 2.8 meters deep. The circular columns are 0.35 meters in diameter, dividing the facade into three equal parts. The cement plaque is 1.8 meters wide and 0.6 meters high, located in the center of the facade. The vase railing is 0.9 meters high, and the unit is 0.3 meters wide.

[0086] Visual feature analysis shows that the overall facade color is primarily the reddish-brown color of Southern Fujian bricks (75%), with off-white cement trim (15%) and dark brown wood (10%). Texture features reveal a regular brick joint pattern and a smooth cement surface. The style feature vector, extracted using a deep neural network, scored 0.87 in the imitation Western-style dimension and 0.72 in the traditional Southern Fujian style dimension, reflecting a blend of Chinese and Western influences.

[0087] Material identification accurately determined the use of traditional materials such as southern Fujian red bricks, cement plaster plastic, and wooden doors and windows, which are in line with the architectural technology characteristics of the late Qing Dynasty.

[0088] The material recognition module accurately identifies a variety of materials and their characteristics based on a dedicated training set for arcades: - Minnan Red Brick (Confidence Level 0.96): Identified as traditional kiln-fired red brick by its distinctive orange-red hue (HSV Hue 15-25°) and regular brick joint pattern; - Cement plaster decoration (confidence 0.94): The fine texture of early cement materials was identified, with slight cracks on the surface, consistent with the cement craftsmanship characteristics of the late Qing Dynasty and early Republic of China; - Traditional Chinese fir doors and windows (confidence level 0.92): Based on the vertical wood grain and surface patina, they are identified as Fujian fir and are in good condition. These material identification results automatically generate corresponding codes (M02, M03, M01) and are linked to the architectural technology system of the late Qing Dynasty, providing a reliable basis for dating the buildings.

[0089] 4. Standardized Description Generation Description of the foundation layer: a two-story brick and wood structure imitating a Western-style building with an oil-lined arcade; Description of standard floor: Arcade building on Dihua Street in Taipei in the late Qing Dynasty, with Minnan brick facade, round arch pavilion feet, cement plaque sign, and vase railing decoration; Professional Floor Description: Built in the 1870s, this imitation Western-style arcade building now houses the Li Gan Incense Oil Shop, boasting a history of approximately 150 years. The facade is constructed of Minnan red bricks. Two circular columns on the first floor divide the facade into thirds, supported by flat stone beams. The pavilion's footing utilizes traditional Minnan brick arching techniques, showcasing the exquisite craftsmanship of the late Qing Dynasty. The central cement plaque and decorative vase railings embody the traditional arcade's blend of Eastern and Western architectural styles. The second-floor wooden windows, with center-opening windows opening left and right and upper and lower windows on the sides, reflect traditional living habits. This building bears witness to the historical transformation of Dihua Street from a late Qing Dynasty commercial port to a modern commercial hub.

[0090] 5. Knowledge Graph Construction The system automatically created the "Li Gan Xiangyou Shop Arcade" building node and established composition relationships with 26 component nodes. Each component node is associated with corresponding material nodes, period nodes, and process nodes. For example, the arch pavilion foot node is connected to the "Southern Fujian Brick" node via the "Material Used" relationship, to the "Late Qing Dynasty" node via the "Construction Period" relationship, and to the "Brick Arch Masonry" node via the "Process Used" relationship.

[0091] Through knowledge reasoning, the system discovered that the building shares similar Western-style architectural features with other late Qing Dynasty merchant houses on Dihua Street, inferring that this style was popular in Taipei after the opening of the port to trade in the late Qing Dynasty. The system also identified the specific functions of the oil company building, such as its need for large storage space and ventilation, which explains the rationality of its internal spatial layout.

[0092] VI. Verification and Evaluation Five architectural history experts were invited to evaluate the system's output. The component recognition accuracy rate was 93%, indicating that the system accurately identified all major components, particularly the arched pavilion feet and vase railings. The coding accuracy rate was 96%, indicating that the coding system accurately reflects the building's period characteristics. The descriptive text was rated 91% reasonable, indicating that the generated description accurately reflects the building's historical value and architectural features.

[0093] The entire processing process takes 24 seconds, including 4 seconds for image preprocessing, 14 seconds for component recognition, 4 seconds for encoding and feature extraction, and 2 seconds for description generation and knowledge graph update. This represents a 100-fold improvement in efficiency compared to the 2-3 hours required by traditional manual methods.

[0094] 7. System application effect statistics In a large-scale test of 1,000 arcade buildings, the system demonstrated stable high performance: - Parts recognition accuracy: 92.5% (common parts types >95%, rare parts types >85%) - Coding accuracy: 96% (position coding accuracy 98%, attribute coding accuracy 94%) - Description generation rationality rate: 90% (95% for basic layer, 92% for standard layer, 85% for professional layer) - Average processing time: 28.3 seconds / building (fastest 22 seconds, slowest 35 seconds) - Knowledge graph scale: 15,000+ nodes, 30,000+ relationship edges - System availability: 99.5% (24 / 7 stable operation) The effectiveness and practicality of the proposed method were fully demonstrated through its practical application in the arcades of the Li Gan Incense Oil Shop. The system not only accurately identified various traditional architectural components, such as arched pavilion legs and vase railings, but also achieved intelligent conversion from image to knowledge through standardized coding and multi-level descriptions. In particular, the recognition of the combination of late Qing Dynasty imitation Western-style buildings and traditional Southern Fujian architectural elements demonstrated the system's precise grasp of the cultural characteristics of traditional arcade architecture. Furthermore, through the construction of a knowledge graph, the system revealed the common characteristics of traditional commercial buildings on Dihua Street, providing digital tools for comparative research on traditional arcade architecture.

[0095] The above embodiments and drawings do not limit the product form and style of the present invention. Any appropriate changes or modifications made by ordinary technicians in the relevant technical field should be deemed to be within the patent scope of the present invention.

Claims

1. A method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system, characterized by include: Step S1. Construct a standard coding system with six levels: building system level coding, regional subsystem coding, category coding, subcategory coding, specific building coding, and component coding. Component coding uses a position-type-attribute sub-coding composite structure, with position-type forming the basic component coding. Step S2. Systematically collect images according to preset image acquisition standards to obtain a sufficient number of arcade facade images, and then pre-process the collected arcade facade images, and then manually annotate the pre-processed arcade facade images to construct an annotation dataset; Step S3. Build a model for identifying arcade components. Use existing datasets and annotated datasets to train and identify these components. The network architecture of this model uses Mask R-CNN as its foundational framework. Its backbone network uses ResNet-101 combined with a Feature Pyramid Network (FPN). This network extracts multi-scale features through bottom-up and top-down feature fusion, adapting to the size differences of arcade components, from small ornaments to large pediments. Furthermore, a spatial relationship module is added to the Mask R-CNN framework. This module uses a graph convolutional network to model the spatial constraints between different components. Step S4. For the recognition result of step S3, assign the coding information of the standard coding system and extract multi-dimensional features, perform component association analysis, and build a component feature database; Step S5. Based on the recognition results from step S3 and the encoded information from step S4, three levels of standardized component description text are automatically generated, achieving accurate conversion from visual features to natural language. These descriptions include a basic level description of less than 20 words, a standard level description of less than 50 words, and a specialized level description of less than 100 words. Step S6. Integrate the recognition result of step S3, the encoding information of step S4, and the description text of step S5 to construct a structured knowledge graph of the arcade facade components; Step S7. Integrate the various processing capabilities in steps S2 to S6 into a complete system, which specifically includes: implementing the image preprocessing capability of step S2, implementing the component intelligent recognition capability of step S3, implementing the automatic encoding and feature extraction capability of step S4, implementing the multi-level description generation capability of step S5, and implementing the knowledge graph construction capability of step S6, to achieve end-to-end automated processing through a unified system architecture.

2. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 1 is characterized in that The standard coding system in step S1 is specifically: (1) Building system-level coding, uniformly using CTQL as the top-level identifier for traditional arcades; (2) Regional sub-system codes are set according to the geographical areas where the arcades are distributed; (3) Group codes, used to distinguish regional characteristics; (4) Sub-group coding, accurate to the specific historical street; (5) Specific building code, which assigns a number of digits to each arcade; (6) Component code, its sub-code is: (1) Position: According to the structural logic of the arcade facade, R represents the roof, B represents the body, and F represents the seat; (2) Type, which is used to reflect the detailed classification of the location; (3) Attributes, including at least one of the material, age, and functional characteristics of the component, where M represents material, T represents age, and U represents function.

3. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 1 is characterized in that In the step S2: The preset image acquisition standards are as follows: (1) The axis of the camera lens is perpendicular to the main plane of the building facade when shooting; (2) The shooting distance should be controlled within 1.5 to 2 times the building height; (3) Use a standard lens with an equivalent focal length of 35 to 50 mm; (4) Choose uniform lighting conditions for shooting; (5) The image resolution must be no less than 4000 × 3000 pixels; The preprocessing process includes: First, perspective correction is performed by detecting the vertical and horizontal lines in the facade, calculating the perspective transformation matrix, and correcting the oblique image to a standard orthographic projection; Secondly, color standardization is performed, using the standard color card placed during shooting to calculate the color correction matrix; Then, adaptive histogram equalization technique is used to enhance the image contrast; Finally, scale normalization is performed to adjust all images to a uniform pixel density; When constructing the annotated dataset, at least five architectural history experts with more than 10 years of experience were invited to participate in the annotation. They annotated the pre-processed arcade facade images using the annotation software, accurately selected the component outlines on the image, selected the component type, and annotated the material properties and preservation status to finally obtain the annotated dataset. To ensure the annotation quality, a cross-validation mechanism was used during the annotation process. Each component was independently annotated by at least two experts, and any disagreements were reached through discussion.

4. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 1 is characterized in that The training strategy of the arcade component recognition model in step S3 adopts the following three-stage progressive method: In the first stage, the COCO dataset is used for pre-training to enable the arcade component recognition model to learn basic visual feature extraction capabilities; In the second stage, domain adaptation is performed using a building image dataset to enable the arcade component recognition model to understand the unique visual patterns of building components. In the third stage, the labeled dataset of step S2 is used for fine training, focusing on optimizing the arcade component recognition model's ability to recognize arcade-specific components.

5. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 4 is characterized in that The training process of the arcade component recognition model in step S3 adopts the following optimization measures: (1) Using the focal loss function to solve the problem of component category imbalance, higher loss weights are given to rare component categories with a smaller number; (2) Introducing an online hard example mining mechanism, presetting easily confused component pairs in arcade components based on practical experience, automatically identifying these component pairs during training, and increasing the training frequency of these component pairs according to a preset number or ratio value; (3) Through data enhancement technology, the training samples are expanded three times to improve the generalization ability of the arcade component recognition model.

6. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 4 is characterized in that In step S3, a post-processing mechanism is established to ensure the rationality of the recognition result, specifically: (1) Confidence screening is performed on the detection boxes output by the arcade component recognition model, and only the results with a confidence level higher than 0.8 are retained; (2) Use the non-maximum suppression algorithm to eliminate overlapping detection frames to avoid repeated recognition of the same part; (3) Build a logic rule library based on architectural common sense and verify the rationality of the recognition results, such as checking whether the relative position relationship of the components conforms to the architectural construction logic.

7. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 1 is characterized in that The process of allocating the standard code in step S4 is as follows: First, determine the basic code based on the identified component type and its position on the arcade facade; By analyzing the visual characteristics of the components, we can infer their possible construction age; Combined with the geographical information of the image shooting location, the region-related code is automatically filled in; Determine the material properties of components through a pre-trained material classification network.

8. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 7 is characterized in that The multidimensional features extracted in step S4 cover the following dimensions: (1) Geometric characteristics; (2) Visual features; (3) stylistic features; (4) Material characteristics; The steps of component association analysis in step S4 are: (1) Spatial relationship modeling: Based on the bounding box coordinates of the component recognition results, the spatial distance matrix and relative position relationship between the components are calculated to establish a component spatial relationship diagram; (2) Statistical co-occurrence probability: Calculate the co-occurrence frequency of different component types based on a large number of samples and construct a component co-occurrence probability matrix; (3) Co-occurrence pattern mining: Statistically analyze the co-occurrence frequency of different component types in the arcade samples, and use association rule mining algorithms to discover the combination rules of strongly associated components; (4) Verification of age consistency: Based on the inferred age attributes of the components, the age consistency of different components of the same building is verified to identify possible later reconstruction parts; (5) Constructing a component relationship network: Using components as nodes and association strength as edge weights, a weighted undirected graph is constructed for subsequent architectural style analysis and integrity verification.

9. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system according to claim 7, characterized in that: In step S4, the constructed component feature database is stored in MongoDB; each identified component creates an independent record in the database, which includes a unique identifier, a complete six-level code, precise coordinates in the original image, various extracted feature vectors, recognition confidence scores, and processing timestamp information; by establishing a multidimensional index, fast retrieval based on coding, feature similarity, and spatial location is supported.

10. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 1 is characterized in that The following quality control mechanism is introduced in step S5: (1) Architectural Grammar Checker: Based on the traditional dependency syntax analysis, it adds architectural professional expression standard checking; (2) Architectural terminology consistency check module: Construct a mapping table of traditional arcade professional terms and uniformly map synonyms into standard terms; (3) Architectural logic validator: establishes architectural domain-specific rules to verify the historical logical correctness of the description; (4) Architectural text readability evaluation algorithm: Based on the characteristics of architectural professional texts, the indicators are set as follows: the density of professional terms does not exceed 30% and the average length of sentences does not exceed 25 words.

11. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 1 is characterized in that The ontology of the knowledge graph in step S6 defines the following six core entities: (1) Building entity, recording the basic information of the entire arcade; (2) Component entities, which store detailed properties of various building components; (3) Material entities, describing the physical and chemical properties of different building materials; (4) Period entities, which divide historical eras and associate era characteristics; (5) Regional entity, recording geographic spatial information; (6) Craft entities, recording traditional and modern construction techniques; The relationships between entities include: (1) Compositional relationships, describing the whole-part relationships between buildings and components; (2) Positional relationship, recording the spatial distribution of components on the facade; (3) Material relationship, connecting components and the building materials they use; (4) Time relationship, marking the construction and modification history of the component; (5) Process relationship, which describes the manufacturing technology of the components; (6) Evolutionary relationships, tracking the historical changes in component styles.

12. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 11, characterized in that In step S6, the knowledge graph integrates multi-source data through knowledge extraction and fusion processes, specifically: (1) Automatically extract component entities and basic attributes from the recognition results of step S3 to create the initial nodes of the knowledge graph; (2) Establishing connecting edges between nodes by analyzing the spatial adjacency and functional association between components; (3) Supplement the historical and cultural information of components by integrating external data sources; (4) Use entity alignment technology to identify and merge different representations of the same object to ensure the consistency and integrity of knowledge; In step S6, the knowledge graph has an inference rule engine that supports complex knowledge discovery, specifically: (1) Reasoning mechanism based on description logic; (2) Temporal reasoning, analyzing the evolution path of component styles; (3) Spatial reasoning: discovering the distribution patterns of components.

13. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 1 is characterized in that The system in step S7 adopts an end-to-end processing flow and an asynchronous processing mechanism to support parallel processing of batch images and ensure efficient operation of the system. Specifically: When the user uploads an image of an arcade facade, the system automatically starts the processing pipeline, first performing image quality inspection and preprocessing, then calling the trained arcade component recognition model to identify the components inside the arcade, then assigning coding information to each component and extracting multi-dimensional features, and then generating three levels of component description text, and finally storing it in the knowledge graph and updating the index.

14. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 13, characterized in that The system in step S7 uses the following technical means to optimize system performance: (1) Convert 32-bit floating-point model parameters into 8-bit integer representation through model quantization technology; (2) Use TensorRT to optimize the inference engine and fully utilize the parallel computing capabilities of the GPU; (3) Implement distributed architecture design to support multi-node collaborative processing; (4) Establish an intelligent caching mechanism to cache frequently accessed feature data and query results.

15. The method for intelligent recognition and language conversion of traditional arcade facade components based on a hierarchical coding system as claimed in claim 13, characterized in that The system in step S7 continuously improves system performance through a continuous optimization mechanism, specifically: (1) Establish a user feedback collection system; (2) Organize expert reviews regularly; (3) Expand the training dataset; (4) Update and optimize the recognition model; (5) Improve the content of the knowledge graph.

Images