Antique information recognition system and method based on self-attention mechanism

Abstract

The present invention discloses an antique information recognition system and method that introduces a self-attention mechanism, which relates to the field of image processing technology. The system comprises: constructing a self-attention recognizer and embedding it in an antique information platform; uploading an antique scan image according to the multi-threaded port of the antique information platform, pre-processing the antique scan image and importing it into the self-attention recognizer; performing self-attention element triggering and decision-making based on recognition needs to determine the information recognition result; wherein the recognition step based on the self-attention recognizer includes: performing fuzzy texture-based image recognition and block segmentation, performing directional recognition under element-concurrent self-attention on the block image, aggregating and reorganizing the concurrent recognition results, and determining the information recognition result. The present invention solves the technical problems of low accuracy, inability to effectively extract global features, and inability to process complex damaged areas in the prior art of antique image recognition, thereby achieving the technical effect of improving antique image recognition accuracy and comprehensively extracting feature information.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an antique information recognition system and method that introduces a self-attention mechanism. Background Art

[0002] Currently, antique image recognition technology faces challenges such as low accuracy, an inability to effectively extract global features, and an inability to process complex damaged areas. Traditional methods typically rely on local feature extraction, ignoring the correlation between global information and complex regions within an image. This results in suboptimal recognition results when performing comprehensive recognition on antique images. Summary of the Invention

[0003] The present application provides an antique information recognition system and method that introduces a self-attention mechanism, which is used to solve the technical problems of low accuracy, inability to effectively extract global features and process complex damaged areas in the existing technology in antique image recognition.

[0004] In view of the above problems, this application provides an antique information recognition system and method that introduces a self-attention mechanism.

[0005] In a first aspect of the present application, a system for identifying antique information using a self-attention mechanism is provided, the system comprising:

[0006] The identifier deployment module is used to introduce multiple self-attention dimensions, perform hierarchical structured deployment, build a self-attention identifier and embed it in the antique information platform; the information identification module is used to upload antique scans according to the multi-threaded port of the antique information platform, pre-process the antique scans and import the self-attention identifier, perform self-attention meta-triggering and decision-making based on recognition needs, and determine the information recognition results; wherein, the recognition steps based on the self-attention identifier include: performing fuzzy texture-based image recognition and block segmentation, performing meta-concurrent self-attention directional recognition on the block images, aggregating and reorganizing the concurrent recognition results, and determining the information recognition results, wherein the meta-concurrent self-attention at least includes block feature elements, inter-block relationships, and cross-modal relationships.

[0007] The second aspect of the present application provides an antique information recognition method that introduces a self-attention mechanism, the method comprising:

[0008] Multiple self-attention dimensions are introduced, hierarchical and structured deployment is carried out, a self-attention recognizer is constructed and embedded in the antique information platform; according to the multi-threaded port of the antique information platform, the antique scan image is uploaded, the antique scan image is preprocessed and imported into the self-attention recognizer, and self-attention meta-triggering and decision-making are carried out based on recognition needs to determine the information recognition result; wherein, the recognition step based on the self-attention recognizer includes: by performing fuzzy texture-based image recognition and block segmentation, performing meta-concurrent self-attention directional recognition on the block image, aggregating and reorganizing the concurrent recognition results to determine the information recognition result, wherein the meta-concurrent self-attention at least includes block feature elements, inter-block relations, and cross-modal relations.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application introduces multiple self-attention dimensions, performs hierarchical structured deployment, builds a self-attention recognizer and embeds it in an antique information platform; uploads antique scans according to the multi-threaded port of the antique information platform, pre-processes the antique scans and imports the self-attention recognizer, performs self-attention meta-triggering and decision-making based on recognition needs, and determines the information recognition results; wherein, the recognition steps based on the self-attention recognizer include: by performing image recognition and block segmentation based on fuzzy texture, performing directional recognition under meta-concurrent self-attention on the block images, aggregating and reorganizing the concurrent recognition results, and determining the information recognition results, wherein the meta-concurrent self-attention at least includes block feature elements, inter-block relations, and cross-modal relations. The present invention solves the technical problems of low accuracy, inability to effectively extract global features and process complex damaged areas in the existing technology in antique image recognition. By introducing multiple self-attention dimensions, performing hierarchical structured deployment and concurrent self-attention mechanism, the technical effects of improving antique image recognition accuracy and comprehensively extracting feature information are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0012] Figure 1 A schematic diagram of the structure of an antique information recognition system that introduces a self-attention mechanism according to an embodiment of the present application;

[0013] Figure 2 A flow chart of the method for identifying antique information using a self-attention mechanism provided in an embodiment of the present application.

[0014] Description of the accompanying drawings: identifier deployment module 11, information identification module 12. DETAILED DESCRIPTION

[0015] This application provides an antique information recognition system and method that introduces a self-attention mechanism to solve the technical problems of low accuracy, inability to effectively extract global features and process complex damaged areas in the existing technology in antique image recognition. By introducing multiple self-attention dimensions, hierarchical structured deployment and concurrent self-attention mechanism, the technical effect of improving the accuracy of antique image recognition and comprehensively extracting feature information is achieved.

[0016] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0017] It should be noted that any variations of the terms "include" and "have" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices.

[0018] Example 1, as Figure 1 As shown, the embodiment of the present application provides an antique information recognition system that introduces a self-attention mechanism, and the system includes:

[0019] The identifier deployment module 11 is used to introduce multiple self-attention dimensions, perform hierarchical structured deployment, build a self-attention identifier and embed it in the antique information platform.

[0020] In this embodiment, the primary task of the Recognizer Deployment Module 11 is to build and deploy a self-attention recognizer with high-precision, multi-dimensional analysis capabilities to support image recognition and feature determination tasks within the antique information platform. First, by introducing multiple self-attention dimensions, the Recognizer Deployment Module 11 establishes an attention mechanism system that covers different perceptual levels and information structures. This mechanism covers fuzzy texture attention, feature attention, relational attention, cross-modal attention and global attention. Fuzzy texture attention is used to capture the fuzzy areas in antique images caused by age, wear and aging, and enhance the perception of fuzzy features; feature attention focuses on identifying the most representative and recognizable image areas in antiques, such as specific patterns, carvings, material traces, etc.; relational attention is used to model the spatial relationship or semantic connection between local features in the image, such as the positional dependence or style combination between components; cross-modal attention enables the recognizer to fuse visual information with existing text descriptions, expert labels or database knowledge in the platform to achieve a linked understanding between image information and non-image information; and global attention integrates image information from a holistic perspective to enhance the recognizer's understanding of the overall style, configuration, layout and other macro features of antiques.

[0021] On the basis of the establishment of the attention mechanism, the recognizer deployment module 11 deploys various attention mechanisms in a hierarchical and structured manner to improve functional coordination efficiency and hierarchical perception capabilities. First, the first recognition layer is constructed based on fuzzy texture attention. This layer is responsible for parsing the fuzzy areas of antique images and preliminary block processing, providing high-quality local input for subsequent layers; then, feature attention, relationship attention, and cross-modal attention are deployed in parallel in the second recognition layer. In this layer, the recognizer simultaneously processes local feature extraction, regional relationship modeling, and image-text information fusion to form multi-channel concurrent perception capabilities; finally, the third recognition layer is deployed based on the global attention mechanism, and the results of the first two layers are aggregated and reorganized, and the recognition decision results are output based on the overall characteristics. A unified self-attention recognizer structure is constructed between the three recognition layers through a fully connected method to ensure that different attention mechanisms are both independent and fused and collaborative in function, thereby improving the overall recognition accuracy and generalization ability.

[0022] The practical deployment of a self-attentional recognizer relies on a thorough training process and diverse data support. During the training phase, a dataset containing a large number of antique images is prepared. These images cover a variety of materials, historical periods, and are captured from different angles and resolutions, ensuring comprehensive and representative training data. In addition to image data, cross-modal data such as expert annotations, literature, text descriptions, and expert reviews are also included to train the cross-modal attention mechanism and enhance the recognizer's ability to integrate multi-source information. Before training, all image data undergoes standardized preprocessing, including image denoising, brightness alignment, and resolution scaling, to ensure that the input data meets the model's processing requirements. During training, the self-attentional recognizer optimizes the weight parameters within the self-attention structure through multiple rounds of iterative learning, enabling the different attention mechanisms to learn appropriate weight distributions and triggering paths based on the training samples. Fuzzy texture attention learns to distinguish texture details in noisy images, feature attention learns to identify the most distinctive image regions across different antiques, relational attention trains the model to understand spatial or structural relationships between local images, and cross-modal attention builds semantic connections between images and text, improving the accuracy and interpretation capabilities of multi-dimensional recognition. The training process also includes training for self-attention meta-trigger mechanisms tailored to recognition needs. This allows the recognizer to autonomously activate relevant recognition layers or attention nodes based on the recognition task, thereby achieving a recognition process with on-demand activation and flexible responses. Throughout the training process, the model is continuously evaluated and tuned. Through cross-validation, precision evaluation, and recall monitoring, we ensure that the recognizer maintains stable and reliable performance after training.

[0023] After the self-attention recognizer completes training and optimization, the recognizer deployment module 11 integrates it into the antique information platform in an embedded manner to realize an integrated hardware and software operating environment.

[0024] Furthermore, in the system provided by the embodiment of the application, the introduction of multiple self-attention dimensions also includes:

[0025] Introduce fuzzy texture attention and deploy the block mode to set the first attention node; introduce feature attention and set the second attention node; introduce relational attention and set the third attention node; introduce cross-modal attention and set the fourth attention node; introduce global attention and set the fifth attention node.

[0026] In an embodiment of the present application, fuzzy texture attention is first introduced and processed through a blocking mode. Fuzzy texture attention is used to identify fuzzy areas in an image due to aging, wear or other external factors. In order to effectively extract the texture features of these fuzzy areas, the image is divided into multiple small blocks, and the blocking method is selected according to the regularity or irregularity of the texture. In the case of regular textures, non-uniform geometric blocking, such as irregular rectangles or polygons, is used to better adapt to the texture structure in the image; for irregular textures, uniform geometric blocking, such as squares or rectangles, is used to ensure that the size and shape of each block are consistent. In this way, fuzzy texture attention can focus on the key texture areas in each small block and enhance the recognition of details in complex images. This step is set as the first attention node, which is responsible for processing the fuzzy texture areas in the image.

[0027] Next, feature attention is introduced and set as the second attention node. This mechanism helps the model focus on key feature areas in the image, particularly those local features that are most important for recognition. For example, carvings, patterns, or unique markings in antique images are crucial for determining the authenticity and historical value of the objects. By weighting these key features, feature attention enables the model to prioritize and accurately identify these local areas, even in complex or noisy backgrounds. This mechanism ensures that the system can efficiently extract the most important information in the image, thereby improving recognition accuracy.

[0028] Relational attention is then introduced and set as the third attention node. This mechanism helps the model understand the spatial or semantic relationships between different regions in an image. By analyzing the interdependencies and spatial relationships between local regions, relational attention can identify the structural relationships between multiple parts in an image. For example, in images of ancient sculptures, different parts of the sculpture (such as the hands, face, and base) may have certain spatial and structural connections. Relational attention can identify these dependencies and use them as a basis for overall recognition. In this way, the model can capture the associations between local features, thereby enhancing its understanding of the overall structure of the image.

[0029] Next, we introduce cross-modal attention, setting it as the fourth attention node. Cross-modal attention enhances the model's recognition capabilities by combining images with other types of information (such as text descriptions and tags). Images themselves may contain features that are difficult to identify directly visually, but textual information associated with the image (such as expert labels and item descriptions) can provide important contextual clues for recognition. By combining image information with this textual information, the cross-modal attention mechanism enables the model to rely not only on visual information during recognition but also obtain supplementary information from other modalities, improving its understanding of complex scenes.

[0030] Finally, global attention is introduced and set as the fifth attention node. The purpose of the global attention mechanism is to analyze the image holistically and integrate the features of each local area. Global attention ensures that the model can not only focus on local features when processing an image, but also understand the overall structure of the image by integrating global information. For example, in an image of an antique, the overall layout and the relationship between its parts may provide important clues. Through global attention, the model can integrate all local features to form a comprehensive understanding of the entire image. The global attention mechanism enables the recognizer to understand the overall picture of the image at a higher level, improving its overall image parsing ability.

[0031] Through these steps, the introduced attention mechanisms are set as the first to fifth attention nodes in sequence, ensuring that the system can deeply analyze and understand the image from multiple dimensions such as texture, local features, regional relationships, cross-modal information and global structure, thereby achieving accurate image recognition.

[0032] Furthermore, in the system provided in the embodiment of the application, performing hierarchical structured deployment also includes:

[0033] According to the first attention node, a first recognition layer is constructed; the second attention node, the third attention node and the fourth attention node are deployed in parallel to construct a second recognition layer; according to the fifth attention node, a third recognition layer is constructed; the first recognition layer, the second recognition layer and the third recognition layer are fully connected in three layers to determine the self-attention recognizer.

[0034] In an embodiment of the present application, in the process of constructing a self-attention recognizer, a first recognition layer is first constructed based on the first attention node. This recognition layer focuses on processing the fuzzy texture information in the image and uses a fuzzy texture attention mechanism. The task of fuzzy texture attention is to enhance the features of the fuzzy areas in the image caused by factors such as lighting, wear, and aging. In specific operations, the Canny edge detection algorithm is first applied to enhance the image edges to improve the detail information in the image, especially the boundaries of the fuzzy areas. The image is then divided into multiple small blocks, which are segmented according to the regularity or irregularity of the image texture. For images with regular textures, non-uniform geometric blocks, such as irregular rectangles or polygons, are used to adapt to the complex textures in the image; while for images with irregular textures, uniform geometric blocks, such as squares or rectangles, are used. Each image block is processed by the fuzzy texture attention mechanism, which enhances the detail information of the fuzzy areas in the image through a weighted strategy, making the fuzzy texture features clearer. During the training process, an image dataset with fuzzy area annotations is used. Through supervised learning, the model adjusts parameters according to the loss function to optimize the texture enhancement capability of the fuzzy areas, thereby effectively extracting the potential texture features in the image.

[0035] Next, the second, third, and fourth attention nodes are introduced and deployed in parallel to form the second recognition layer. This layer processes different image feature dimensions through three parallel sub-channels. First, the second attention node uses a convolutional neural network (CNN) to extract local features from the image. The convolution operation scans the image through a series of convolution kernels, automatically learning key local features within the image, such as carvings, patterns, or other details. During training, a dataset of antique images with labeled local features, such as images with patterns or cracks, is used. After processing these local features through the convolution layer, a feature map is formed and passed to downstream processing. Second, the third attention node uses a graph convolutional network (GCN) to model the spatial relationships between image regions. In this step, each image region is treated as a node of a graph, with adjacent regions connected by edges. The graph convolution operation aggregates information between regions. For example, the relative positions of different components in an image (such as a handle and a base) are modeled by the GCN, helping the recognizer understand the dependencies between these regions. Finally, the fourth attention node aligns the image and text description through a joint embedding space, matching image features with the text description (e.g., "Qing Dynasty blue and white porcelain with dragon and floral patterns") in the same semantic space. By using a multi-head attention mechanism, image and text information interact in this space, helping the model to derive the semantic context of the image from the text description. During training, the model learns how to establish effective connections between image content and text descriptions through paired information between images and text descriptions. This parallel deployment of the second recognition layer simultaneously extracts local image features, structural relationships, and semantic information, providing a more comprehensive feature representation for subsequent recognition tasks.

[0036] Subsequently, based on the fifth attention node, the third recognition layer is constructed. This layer uses a self-attention mechanism to model the entire image. The self-attention mechanism calculates the similarity between each region in the image and other regions, using these similarities as weights to integrate information from various parts of the image. Specifically, the image is divided into multiple patches of fixed size, each patch is mapped to a vector, and the correlation between the patches is calculated using the self-attention mechanism. During training, an image dataset containing global labels, such as style or category labels, is used to supervise the model in extracting important information from the global features of the image. For example, when identifying complex sculptures, the third recognition layer can integrate information from all parts of the image (such as the head, base, decoration, etc.) through the global attention mechanism to obtain a holistic understanding of the image. The purpose of this layer is to capture structural features across the entire image and improve the model's ability to perceive global information.

[0037] Finally, the first, second, and third recognition layers are fully connected to form a self-attention recognizer. This process aims to fuse the feature information output by the three layers, integrating the fuzzy texture features, local detail features, spatial relationship features, and global information from different recognition layers. Through the fully connected layers, the output features of each layer are weighted summed or concatenated to form a high-dimensional comprehensive feature representation. Subsequently, a multi-layer neural network is used to integrate all information in the image and ultimately generate a recognition result. Through this process, the final self-attention recognizer is able to not only process local textures, local features, and spatial relationships, but also understand the overall structure of the image, providing efficient and accurate image recognition results.

[0038] Furthermore, the system provided in the application embodiment also includes:

[0039] The blocking mode is determined according to the texture characteristics; if the texture is regular, non-uniform geometric blocking is adopted; if the texture is irregular, uniform geometric blocking is adopted.

[0040] In the embodiment of the present application, the appropriate segmentation mode is first determined based on the texture features. The core of this process is to select the most appropriate image segmentation strategy based on the texture type of the image (regular or irregular), so as to better capture and process the details in the image.

[0041] When an image exhibits regular textures, non-uniform geometric blocking is used for segmentation. Regular textures typically have distinct, repetitive structures and patterns, such as neat grid-like textures, uniform patterns, or designs, with these features being distributed relatively consistently across the image. For example, in images of antique porcelain, regular textures may appear as uniform geometric patterns or consistent line layouts. To more effectively process such images, non-uniform geometric blocking is used to segment the image into irregularly shaped blocks (such as irregular rectangles and pentagons) to accommodate the more regular details in the image. This blocking method accurately captures the local characteristics of the texture, avoiding the loss of texture features caused by traditional uniform blocking methods.

[0042] For images with irregular textures, uniform geometric blocking is used for segmentation. Irregular textures often manifest as irregular texture structures and forms, such as wear marks in damaged areas or naturally occurring random patterns. These textures lack predictable regularity, and the details in the image vary greatly, so they need to be processed using uniform geometric blocking (such as squares or rectangles). This approach ensures that each block has a consistent shape and size, avoiding the impact of texture irregularities. Uniform blocking allows the model to process each region in the image and extract independent features from each region, while reducing interference caused by inconsistent textures during the blocking process.

[0043] By selecting these block segmentation methods, the texture information of the image can be effectively extracted and optimized according to the characteristics of the image (regular or irregular) after block processing.

[0044] The information recognition module 12 is used to upload the antique scan image according to the multi-threaded port of the antique information platform, pre-process the antique scan image and import it into the self-attention recognizer, and perform self-attention meta-triggering and decision-making based on the recognition needs to determine the information recognition result; wherein, the recognition step based on the self-attention recognizer includes: performing fuzzy texture-based image recognition and block segmentation, performing meta-concurrent self-attention directional recognition on the block image, aggregating and reorganizing the concurrent recognition results, and determining the information recognition result, wherein the meta-concurrent self-attention at least includes block feature elements, inter-block relations, and cross-modal relations.

[0045] In an embodiment of the present application, the information recognition module 12 is responsible for uploading the antique scans, performing preprocessing, and importing the self-attention recognizer to achieve efficient and accurate image recognition. During this process, the multi-threaded port first takes on the image upload task. Through multi-threading technology, the antique information platform can achieve parallel upload of multiple images, thereby effectively improving the transmission speed and processing efficiency of image data. Each thread processes one image, which enables the platform to process multiple antique scans simultaneously, ensuring that the system can still maintain efficient operation when a large amount of data flows in.

[0046] After uploading, the image enters the preprocessing stage. This stage aims to perform operations such as denoising, normalization, and resizing on the image to improve its quality and consistency, making it suitable for subsequent recognition tasks. Image processing techniques include Gaussian blur denoising, histogram equalization (for contrast enhancement), and normalization to ensure that the image content is clear, standardized, and that color and brightness meet standards. After preprocessing, the image is input into the previously trained self-attention recognizer. This recognizer relies on the self-attention mechanism, a technique that captures the relationships between regions in an image. The self-attention mechanism weights each local region based on its correlation with other regions in the image to determine its importance. This allows the model to dynamically focus on discernible parts of the image and enhance its understanding of the image's global structure. As a result, the model not only efficiently processes image details but also understands connections and global patterns between regions.

[0047] After entering the self-attention recognizer, recognition demand, a key factor in the recognition process, guides its meta-triggering and decision-making. Specifically, depending on the image type, the recognizer flexibly adjusts its processing strategy based on the task requirements. For example, for an image of an antique with complex textures, the recognizer may prioritize processing the image's details based on the texture features; whereas for an image with distinct morphological features, it may focus on recognizing the image's shape and structure. This recognition demand-driven mechanism enables the self-attention recognizer to adaptively optimize and adjust based on the image's actual characteristics and task objectives, thereby improving recognition efficiency and accuracy.

[0048] The next steps involve specific recognition steps based on the self-attention recognizer. First, the image undergoes fuzzy texture-based image recognition and block processing. By analyzing the texture features of the image, the image is first divided into multiple small blocks. The block pattern is selected according to the texture type of the image: if the image presents a regular texture, non-uniform geometric block, such as irregular rectangles or polygons, is used to adapt to the texture pattern of the image; if it is an irregular texture, uniform geometric block, such as squares or rectangles, is used to ensure that the scale of each block is consistent, which is convenient for processing areas with irregular textures. This block processing ensures the independence and flexibility of each area of the image during the recognition process, providing important support for subsequent directional recognition.

[0049] After the block processing is complete, the image blocks undergo targeted recognition using meta-concurrent self-attention. Meta-concurrent self-attention involves parallel computation of the features of each image block across multiple parallel channels, weighting the features of each block using a self-attention mechanism. This process not only considers the image's block-level features (such as texture, color, and shape), but also considers inter-block relationships (i.e., the spatial and semantic dependencies between different blocks). For example, in an image, different parts of a pattern (such as ornamentation or cracks) may reside in different blocks. The meta-concurrent self-attention mechanism establishes relationships between blocks, helping the recognizer understand the connections between these parts. Furthermore, cross-modal relationships are processed at this stage. The combination of images and textual descriptions (such as "Ming Dynasty blue and white porcelain with dragon and floral patterns") is processed using a self-attention mechanism, improving the accuracy and richness of image recognition.

[0050] The recognition results of all blocks are then aggregated and reorganized. This process ensures that the feature contribution of each recognition path is fully utilized by weighted fusion of concurrent recognition results. Common aggregation methods include weighted summation, splicing, and gating mechanisms, which can combine the local features, spatial relationship features, and semantic information of the image to generate the final recognition result. Ultimately, through this process, the information recognition module 12 can accurately integrate the outputs from different recognition layers to provide users with clear and accurate antique image recognition results.

[0051] Ultimately, the information recognition results are output to platform users. These results include image classification labels and related attributes, providing reliable data support for antique identification, style analysis, and more.

[0052] Furthermore, the system provided in the application embodiment also includes:

[0053] Through permission constraints, a multi-threaded port is deployed, wherein the multi-threaded port includes at least a user side and an expert side; according to the multi-threaded port, the antique scan image is uploaded; wherein the uploading method includes antique scanning and uploading based on the built-in data acquisition card of the antique information platform, and retrieving and uploading the image document.

[0054] In an embodiment of the present application, in the information identification platform, the access rights of different users and experts when using the platform are ensured through a permission constraint mechanism to ensure the security and rationality of the data. First, the information identification platform supports different types of operations through multi-threaded port deployment. Specifically, the multi-threaded port includes at least a user side and an expert side. On the user side, ordinary users can access and upload image information related to antiques, while on the expert side, expert users can further analyze and identify images and provide professional feedback. In this way, the platform can provide different functional modules for users and experts according to different permissions, while realizing the management and control of uploaded data, ensuring that the data flow and operation of the platform comply with the permission regulations.

[0055] Based on a multi-threaded port, the information recognition platform efficiently supports image uploads. There are two main upload methods. The first is to upload antique scans using the data acquisition card built into the antique information platform. The data acquisition card is part of the platform's internal hardware, used to convert physical antique scans into digital images. The data acquisition card connects the antique scanner to the platform, captures images, and uploads them to the platform in real time. During the image acquisition process, the acquisition card uses a high-resolution scanner to perform detailed scans of the antiques and uploads the image data to the platform in real time via a high-speed transmission channel, ensuring high-fidelity image quality. This upload method is often used in scenarios requiring high-precision image acquisition, especially for highly detailed objects such as antiques.

[0056] The second upload method is through image file retrieval. This method is often used to retrieve and upload antique images from existing image libraries or databases. Users or experts can use the platform's document retrieval function to select digital images related to antiques and upload them through the platform's interface. These images may originate from external databases, archives, or other platforms and can be quickly imported into the platform through the image file interface. This method offers convenient image upload and is suitable for integrating and managing antique images already in digital format.

[0057] The two uploading methods each have their own advantages. Data acquisition card uploading can provide extremely high image quality and a fine scanning process, which is particularly suitable for obtaining high-resolution original images; while image document retrieval and uploading is more convenient, and can quickly integrate existing digital image resources and improve the efficiency of platform data processing.

[0058] Through these steps, the platform provides flexible and efficient upload channels based on different needs and permission constraints, ensuring that antique image data can enter the system safely and promptly for subsequent identification and analysis.

[0059] Furthermore, the system provided in the embodiment of the application, which triggers and makes decisions based on self-attention elements based on identification needs, also includes:

[0060] Recognition requirements are received, and the recognition requirements are converted into self-attention units, wherein the self-attention unit identifier has a focus direction; based on the self-attention unit, the self-attention recognizer is layer-triggered and node-triggered within the layer.

[0061] In an embodiment of the present application, the information recognition module 12 receives the user's recognition requirements and converts them into self-attention elements, so that the self-attention recognizer can perform targeted processing on the image according to the task requirements. The recognition requirement is a specific request from the user for the image content, such as identifying the texture features, age identification or material analysis of antiques. After receiving the recognition requirement, the requirement is parsed through natural language processing (NLP) technology and converted into self-attention elements, where each self-attention element has a clear "focus", that is, it indicates the image feature area that the recognizer should pay attention to. For example, if the user requires the recognition of the texture features of the image, the "focus" of the self-attention element will instruct the model to prioritize the texture area of the image.

[0062] Based on these self-attention elements, the corresponding layers and nodes of the self-attention recognizer are triggered. The first recognition layer is triggered first, which primarily processes blurred textures in the image. Using texture-based image recognition technology, the blurred areas in the image are analyzed and identified, and then the image is divided into multiple small blocks using a block-based technique for subsequent processing. This process utilizes the self-attention mechanism to focus on key blurred areas in the image, accurately restoring and enhancing the texture information in these areas.

[0063] The segmented image is then input into the second recognition layer, which triggers concurrent segmented image directional recognition based on the second, third, and fourth attention nodes. In this stage, each segment of the image is processed through multiple parallel channels, focusing on local features, spatial relationships, and cross-modal information. A lateral interaction compensation mechanism is introduced to enhance the interaction between different channels, ensuring that the various image features are fully integrated and compensated during processing. After processing, the concurrent recognition results are passed to the next stage.

[0064] Finally, the concurrent recognition results are fed into the third recognition layer, where they are aggregated and reorganized based on historical and modern characteristics. This process analyzes the correlation between historical and modern features in the image, integrating the various parts of the image into a complete recognition output to determine the final information recognition result.

[0065] Furthermore, in the system provided by the embodiment of the application, layer triggering and intra-layer node triggering of the self-attention identifier are performed based on the self-attention element, further comprising:

[0066] Trigger the first recognition layer, identify the fuzzy texture of the antique scan image, and determine the block image; import the block image into the second recognition layer, trigger the concurrent block image directional recognition based on the second attention node, the third attention node and the fourth attention node, and introduce lateral interaction compensation to output the concurrent recognition result; import the concurrent recognition result into the third recognition layer, aggregate and reorganize it based on ancient and modern characteristics, and determine the information recognition result.

[0067] In the information recognition module of this embodiment, the image first enters the first recognition layer. This layer's primary task is to identify blurred textures within the image and determine image segments. Because antique images often contain blurred areas due to age, wear, or aging, these areas require special processing to extract effective features. In this layer, a convolutional neural network is used to extract blurred texture features from the image. Specifically, a CNN uses multiple layers of convolution kernels to scan the image, effectively enhancing image details, particularly edges and texture features, when processing blurred areas. During this process, techniques such as Canny edge detection are used to further enhance image boundaries, making the features of blurred areas more prominent. After image processing is complete, the image is segmented into multiple small blocks for subsequent analysis. The segmentation method is adjusted based on the regularity or irregularity of the texture. If the image exhibits regular textures, non-uniform geometric segmentation, such as irregular rectangles or polygons, is used to better accommodate repetitive textures in the image. If the image exhibits irregular textures, uniform geometric segmentation, such as squares or rectangles, is used to ensure that each block has a consistent size, facilitating subsequent processing and feature extraction. Through this block division method, each small block in the image can be analyzed independently, thus providing high-quality local input for the subsequent recognition layer.

[0068] After segmentation, the segmented image is passed to the second recognition layer for further analysis. In this layer, the second, third, and fourth attention nodes are triggered to process multiple image feature dimensions in parallel. The second attention node extracts local features in the image, such as carvings, patterns, and cracks. These areas are crucial for identifying the authenticity, age, and historical value of antiques. The third attention node focuses on modeling spatial relationships, analyzing the dependencies and relative positions between different image regions. For example, in a sculpture image, the spatial relationships between the hands, face, and base of the sculpture are important. The fourth attention node handles cross-modal information fusion, enhancing the model's recognition capabilities by combining the image with other types of information (such as text descriptions, expert labels, and database knowledge). Through multi-channel parallel processing, local features, spatial relationships, and cross-modal information are simultaneously extracted from the image, providing a multi-dimensional feature representation for subsequent recognition tasks.

[0069] A lateral interaction compensation mechanism is introduced between these parallel processing nodes to ensure that the features output by each node can be effectively integrated. The core of lateral interaction compensation is to dynamically adjust the weights based on the contribution and relevance of the output of each node. Specifically, the relevance of each node output to the current recognition task is calculated, and the output of the node is weighted according to these weights. For the recognition of texture features, the output weight of the second attention node is enhanced to ensure that texture information is processed first; for nodes with weaker spatial relationships, their weights are reduced according to their relative contribution to ensure the accuracy and consistency of the final recognition results. For example, if the texture feature is very important for the current image recognition task (such as the identification of patterns on ancient pottery), the output of this node will be given a higher weight, so that it occupies a more important position in the final recognition result.

[0070] Next, after concurrent processing by the second recognition layer, the concurrent recognition results are passed to the third recognition layer. This layer aggregates and reorganizes the outputs of the first two layers, combining both ancient and modern features for comprehensive analysis. Specifically, the results are weighted and fused based on the historical and modern features contained in the image. Historical features include textures, cracks, and signs of aging in antiques, which are crucial for determining the age and provenance of antiques. Modern features, on the other hand, primarily include restoration traces and elements of modern aesthetics in the image. These features are helpful in determining the restoration history and modern aesthetic style of the item. At this layer, the weights of these features are adjusted based on their temporal context and contribution to the final recognition task. A weighted summation method is used to integrate ancient and modern features, ensuring a seamless integration of historical and modern information. Specifically, historical features are assigned higher weights because they are more critical for determining the age of antiques, while modern features (such as restoration traces) are assigned lower weights because they focus more on the current condition of the item rather than its age. During the weighted summation process, features are assigned different weights based on their importance, ultimately outputting a comprehensive recognition result that integrates both historical and modern features.

[0071] Finally, after processing and fusion at all levels, the outputs from the first, second, and third recognition layers are fused through three fully connected layers to generate the final information recognition results. This information recognition result includes multiple dimensions such as the object's age, style, material, texture characteristics, and restoration status. For example, when analyzing an ancient porcelain, the output recognition results include its age (such as "Ming Dynasty blue and white porcelain"), style (such as "blue and white decoration"), material (such as "porcelain clay"), texture (such as "cracks"), history (such as "natural wear"), and restoration status (such as "modern restoration traces"). This information provides valuable data support for antique authentication, style analysis, and historical research. Through meticulous feature extraction and comprehensive analysis, it ultimately provides accurate and comprehensive recognition results.

[0072] Furthermore, the system provided in the application embodiment also includes:

[0073] The information recognition result includes a first recognition result and a second recognition result. The first recognition result is obtained based on the ancient characteristics of the target antique, and the second recognition result is obtained by converting the first recognition result from ancient to modern times.

[0074] In an embodiment of the present application, the information recognition result includes a first recognition result and a second recognition result. The first recognition result is obtained based on the reorganization of the ancient characteristics of the target antique. Specifically, the target antique image is first preprocessed to enhance the image quality and extract historical features. Through technologies such as convolutional neural networks (CNN), features related to age and style in the image, such as texture, cracks, weathering marks, etc., are identified. These historical features help to infer the age and origin of antique items. In order to accurately identify these features, the image is divided into blocks, and the appropriate block division method is selected according to the regularity of the image texture. Each block is processed independently to improve the accuracy of feature extraction. Through these steps, the first recognition result is finally generated, which accurately reflects the historical characteristics in the antique image.

[0075] The second recognition result is obtained by performing a historical-to-modern conversion on the first recognition result. In this stage, the historical features in the first recognition result are fused with modern restoration traces. The purpose of this historical-to-modern conversion is to combine the historical features and modern restoration features in the antique image to produce a more comprehensive recognition result. Through cross-modal information fusion, the historical information in the image is combined with modern data sources such as text descriptions and expert labels to ensure that modern elements such as restoration traces can be identified. Next, a weighted fusion is performed on the historical and modern features, assigning different weights based on their importance in recognition. Through a weighted summation method, historical features generally have a higher weight, while modern restoration features have a relatively lower weight, ensuring that the recognition result reflects both the historical context and the modern status of the object. Finally, a second recognition result is output that incorporates both historical and modern features, providing users with a comprehensive analysis of the object, covering multiple dimensions such as age, style, and restoration status.

[0076] Furthermore, the system provided in the application embodiment also includes:

[0077] If the antique is damaged, the third recognition layer is triggered to perform self-attention restoration calculations on anchor point positioning and stripe patterns to determine fuzzy restoration information; based on the fuzzy restoration information and the first recognition result, the antique is three-dimensionally simulated to determine the simulated restoration of the antique and perform interface visualization.

[0078] In an embodiment of the present application, if the antique is damaged, the third recognition layer is triggered, and the damaged area is determined by anchor point positioning. First, edge detection is performed on the image, and algorithms such as Canny edge detection are used to automatically identify the boundaries of the damaged area. By locating obvious feature points in the image, the starting and ending parts of the damaged area are found, and these points are called anchor points. Anchor points are obvious feature points in the damaged area, such as the starting point of a crack or the edge of a break. They serve as a reference for the subsequent repair process to ensure that the repaired part can be accurately docked with the surrounding area.

[0079] After anchor point positioning is complete, self-attention restoration of the stripe pattern is performed. This process performs texture restoration on the damaged area of the image, using a self-attention mechanism to calculate the similarities and relationships between the damaged area and the surrounding areas. The self-attention mechanism restores the texture and details of the damaged area by focusing on the dependencies between the damaged and intact areas. Specifically, by analyzing the texture, shape, and other features of the damaged and surrounding areas in the image, it infers the texture pattern that the missing part should have, and then completes the restoration. This process pays special attention to the stripes and texture patterns in the damaged area, ensuring that the restored image naturally connects to the original image.

[0080] After acquiring the fuzzy restoration information, combined with the first recognition results, a 3D simulation is performed. By mapping the restored texture and shape data into 3D space, depth mapping technology is used to restore the 3D structure of the antique image. Specifically, the details in the fuzzy restoration information are combined with the historical characteristics of the image, and a 3D reconstruction algorithm is used to construct a virtual 3D model of the antique object. This process accurately restores the 3D form of damaged areas, particularly cracks and defects, by extracting information such as the object's shape, size, and surface details, ensuring that the restored model truly reflects the appearance of the antique.

[0081] Finally, the restored 3D model is presented to the user using interface visualization technology. Using OpenGL technology, the 3D restored model is rendered and displayed to the user, who can observe the restored antique by rotating, zooming, and other methods.

[0082] In the embodiments of the present application, in summary, the embodiments of the present application have at least the following technical effects:

[0083] This application introduces multiple self-attention dimensions, performs hierarchical structured deployment, builds a self-attention recognizer and embeds it in an antique information platform; uploads antique scans according to the multi-threaded port of the antique information platform, pre-processes the antique scans and imports the self-attention recognizer, performs self-attention meta-triggering and decision-making based on recognition needs, and determines the information recognition results; wherein, the recognition steps based on the self-attention recognizer include: by performing image recognition and block segmentation based on fuzzy texture, performing directional recognition under meta-concurrent self-attention on the block images, aggregating and reorganizing the concurrent recognition results, and determining the information recognition results, wherein the meta-concurrent self-attention at least includes block feature elements, inter-block relations, and cross-modal relations. The present invention solves the technical problems of low accuracy, inability to effectively extract global features and process complex damaged areas in the existing technology in antique image recognition. By introducing multiple self-attention dimensions, performing hierarchical structured deployment and concurrent self-attention mechanism, the technical effects of improving antique image recognition accuracy and comprehensively extracting feature information are achieved.

[0084] Example 2, based on the same inventive concept as the antique information recognition system that introduces the self-attention mechanism in the previous embodiment, Figure 2 As shown, the embodiment of the present application provides an antique information recognition method that introduces a self-attention mechanism, the method comprising:

[0085] Multiple self-attention dimensions are introduced, hierarchical and structured deployment is carried out, a self-attention recognizer is constructed and embedded in the antique information platform; according to the multi-threaded port of the antique information platform, the antique scan image is uploaded, the antique scan image is preprocessed and imported into the self-attention recognizer, and self-attention meta-triggering and decision-making are carried out based on recognition needs to determine the information recognition result; wherein, the recognition step based on the self-attention recognizer includes: by performing fuzzy texture-based image recognition and block segmentation, performing meta-concurrent self-attention directional recognition on the block image, aggregating and reorganizing the concurrent recognition results to determine the information recognition result, wherein the meta-concurrent self-attention at least includes block feature elements, inter-block relations, and cross-modal relations.

[0086] Furthermore, a multi-dimensional self-attention dimension is introduced, and the method further includes:

[0087] Introduce fuzzy texture attention and deploy the block mode to set the first attention node; introduce feature attention and set the second attention node; introduce relational attention and set the third attention node; introduce cross-modal attention and set the fourth attention node; introduce global attention and set the fifth attention node.

[0088] Furthermore, hierarchical structured deployment is performed, and the method further includes:

[0089] According to the first attention node, a first recognition layer is constructed; the second attention node, the third attention node and the fourth attention node are deployed in parallel to construct a second recognition layer; according to the fifth attention node, a third recognition layer is constructed; the first recognition layer, the second recognition layer and the third recognition layer are fully connected in three layers to determine the self-attention recognizer.

[0090] Furthermore, the method further comprises:

[0091] The blocking mode is determined according to the texture characteristics; if the texture is regular, non-uniform geometric blocking is adopted; if the texture is irregular, uniform geometric blocking is adopted.

[0092] Furthermore, the method further comprises:

[0093] Through permission constraints, a multi-threaded port is deployed, wherein the multi-threaded port includes at least a user side and an expert side; according to the multi-threaded port, the antique scan image is uploaded; wherein the uploading method includes antique scanning and uploading based on the built-in data acquisition card of the antique information platform, and retrieving and uploading the image document.

[0094] Furthermore, the self-attention element triggering and decision-making are performed based on the recognition needs, and the method further includes:

[0095] Recognition requirements are received, and the recognition requirements are converted into self-attention units, wherein the self-attention unit identifier has a focus direction; based on the self-attention unit, the self-attention recognizer is layer-triggered and node-triggered within the layer.

[0096] Furthermore, based on the self-attention element, layer triggering and intra-layer node triggering are performed on the self-attention recognizer, and the method further includes:

[0097] Trigger the first recognition layer, identify the fuzzy texture of the antique scan image, and determine the block image; import the block image into the second recognition layer, trigger the concurrent block image directional recognition based on the second attention node, the third attention node and the fourth attention node, and introduce lateral interaction compensation to output the concurrent recognition result; import the concurrent recognition result into the third recognition layer, aggregate and reorganize it based on ancient and modern characteristics, and determine the information recognition result.

[0098] Furthermore, the method further comprises:

[0099] The information recognition result includes a first recognition result and a second recognition result. The first recognition result is obtained based on the ancient characteristics of the target antique, and the second recognition result is obtained by converting the first recognition result from ancient to modern times.

[0100] Furthermore, the method further comprises:

[0101] If the antique is damaged, the third recognition layer is triggered to perform self-attention restoration calculations on anchor point positioning and stripe patterns to determine fuzzy restoration information; based on the fuzzy restoration information and the first recognition result, the antique is three-dimensionally simulated to determine the simulated restoration of the antique and perform interface visualization.

[0102] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0103] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0104] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. The antique information recognition system using the self-attention mechanism is characterized by: The system comprises: The recognizer deployment module is used to introduce multiple self-attention dimensions, perform hierarchical structured deployment, build a self-attention recognizer, and embed it in the antique information platform; An information recognition module is used to upload antique scans through the multi-threaded port of the antique information platform, pre-process the antique scans and import them into the self-attention recognizer, trigger and make decisions based on the recognition requirements, and determine the information recognition results; Among them, the recognition steps based on the self-attention recognizer include: By performing image recognition and segmentation based on fuzzy texture, performing directional recognition under meta-concurrent self-attention on the segmented images, aggregating and reorganizing the concurrent recognition results, and determining the information recognition results, wherein the meta-concurrent self-attention at least includes segmentation feature elements, inter-block relations, and cross-modal relations; Among them, multiple self-attention dimensions are introduced, including: Introducing fuzzy texture attention, deploying a block mode, and setting the first attention node; Introduce feature attention and set the second attention node; Introduce relational attention and set the third attention node; Introducing cross-modal attention and setting the fourth attention node; Introducing global attention and setting the fifth attention node; Among them, a layered structured deployment is carried out, including: Constructing a first recognition layer based on the first attention node; The second attention node, the third attention node, and the fourth attention node are deployed in parallel to construct a second recognition layer; Constructing a third recognition layer based on the fifth attention node; The first recognition layer, the second recognition layer and the third recognition layer are fully connected to each other to determine a self-attention recognizer.

2. The antique information recognition system using the self-attention mechanism as claimed in claim 1, characterized in that: Determine the blocking mode according to the texture features; Among them, if it is a regular texture, non-uniform geometric blocking is used, and if it is an irregular texture, uniform geometric blocking is used.

3. The antique information recognition system using the self-attention mechanism as claimed in claim 1, characterized in that: Deploy a multi-threaded port through permission constraints, wherein the multi-threaded port includes at least a user side and an expert side; Uploading the antique scan image according to the multi-threaded port; Among them, the uploading method includes scanning and uploading antiques based on the data acquisition card built into the antique information platform, and retrieving and uploading image documents.

4. The antique information recognition system using the self-attention mechanism as claimed in claim 1, characterized in that: Self-attention meta-triggering and decision-making are guided by recognition needs, including: Receiving a recognition requirement, converting the recognition requirement into a self-attention element, wherein the self-attention element identifier has a focus direction; Based on the self-attention element, the self-attention identifier is layer-triggered and node-intra-layer-triggered.

5. The antique information recognition system using the self-attention mechanism as claimed in claim 4, characterized in that: Based on the self-attention element, layer triggering and intra-layer node triggering are performed on the self-attention recognizer, including: triggering the first recognition layer to recognize the fuzzy texture of the antique scan image and determine the segmented image; Importing the segmented image into the second recognition layer, triggering concurrent segmented image directional recognition based on the second attention node, the third attention node, and the fourth attention node, introducing lateral interaction compensation, and outputting concurrent recognition results; The concurrent recognition results are imported into the third recognition layer, aggregated and reorganized based on ancient and modern characteristics to determine the information recognition results.

6. The antique information recognition system using the self-attention mechanism as claimed in claim 5, characterized in that: The information recognition result includes a first recognition result and a second recognition result. The first recognition result is obtained based on the ancient characteristics of the target antique, and the second recognition result is obtained by converting the first recognition result from ancient to modern times.

7. The antique information recognition system using the self-attention mechanism as claimed in claim 6, characterized in that: The system further comprises: If the antique is damaged, the third recognition layer is triggered to perform self-attention restoration calculations on anchor point positioning and stripe patterns to determine fuzzy restoration information; According to the fuzzy restoration information and the first recognition result, a three-dimensional simulation is performed on the antique, the simulated restoration antique is determined, and an interface visualization is performed.

8. The method for identifying antique information by introducing the self-attention mechanism is characterized by: The method is performed by the antique information recognition system introducing the self-attention mechanism according to any one of claims 1 to 7, comprising: Introducing multiple self-attention dimensions, implementing hierarchical structured deployment, building a self-attention recognizer and embedding it in the antique information platform; According to the multi-threaded port of the antique information platform, the antique scan image is uploaded, the antique scan image is pre-processed and imported into the self-attention recognizer, and the self-attention element is triggered and decided based on the recognition needs to determine the information recognition result; Among them, the recognition steps based on the self-attention recognizer include: By performing image recognition and segmentation based on fuzzy texture, performing directional recognition under meta-concurrent self-attention on the segmented images, aggregating and reorganizing the concurrent recognition results, and determining the information recognition results, wherein the meta-concurrent self-attention at least includes segmentation feature elements, inter-block relations, and cross-modal relations; Among them, a multi-dimensional self-attention dimension is introduced, and the method further includes: Introduce fuzzy texture attention and deploy the block mode to set the first attention node; introduce feature attention to set the second attention node; introduce relational attention to set the third attention node; introduce cross-modal attention to set the fourth attention node; introduce global attention to set the fifth attention node; Wherein, hierarchical structured deployment is performed, and the method further includes: According to the first attention node, a first recognition layer is constructed; the second attention node, the third attention node and the fourth attention node are deployed in parallel to construct a second recognition layer; according to the fifth attention node, a third recognition layer is constructed; the first recognition layer, the second recognition layer and the third recognition layer are fully connected in three layers to determine the self-attention recognizer.

Images