Antique age determination method and system based on artificial intelligence and big data

Abstract

The present application relates to the technical field of antique identification, and in particular, the present application discloses an antique age determination method and system based on artificial intelligence and big data, which comprises the following steps: collecting antique images and automatically identifying category labels, accurately positioning the initial identification area, combining process feature detection and multi-dimensional confidence evaluation, screening out the most representative target identification area to reduce the risk of misjudgment, and constructing a multi-level age detection model, using multiple feature analysis modules to extract age characteristics of different feature dimensions, and calculating the importance coefficient of each module through a feature correlation module, and fusing multi-dimensional age characteristics to determine the age determination result of the antique to be identified, thereby significantly improving the accuracy and reliability of antique age determination.

Description

Technical Field

[0001] The present invention relates to the technical field of antique identification, and in particular to a method and system for determining the age of antiques based on artificial intelligence and big data. Background Art

[0002] In the field of antique identification, traditional methods of dating mainly rely on the experience of experts and manual detection tools. These methods are not only time-consuming and labor-intensive, but also easily affected by subjective factors, making it difficult to ensure the accuracy and consistency of the identification results.

[0003] With the advancement of technology, computer vision and machine learning techniques are increasingly being applied to antique identification, aiding in identification through image processing and feature extraction. However, existing technologies often only provide single-dimensional feature analysis, such as those based solely on texture or shape features in an image, which cannot fully reflect the characteristics of an antique. Furthermore, existing technologies lack the ability to integrate features and perform multi-dimensional analysis, making it difficult to accurately determine the age of an antique. Therefore, a new technical solution is needed that can integrate features from multiple dimensions to improve the accuracy and efficiency of antique identification. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for determining the age of antiques based on artificial intelligence and big data.

[0005] First, a method for determining the age of antiques based on artificial intelligence and big data is provided, comprising:

[0006] Acquire a captured image of the antique to be inspected, determine a category label of the captured image by performing category judgment on the captured image, and determine a plurality of initial identification regions matching the category label from the captured image;

[0007] Performing process feature detection on each of the initial identification areas to obtain process feature data of each of the initial identification areas;

[0008] generating a multi-dimensional detection condition for performing a confidence test on the initial identification region, and determining the confidence of the plurality of initial identification regions using a confidence test model based on the multi-dimensional detection condition, the plurality of initial identification regions and their corresponding process feature data;

[0009] selecting a target identification area of ​​the captured image from the plurality of initial identification areas according to the confidence level;

[0010] The captured image of the target identification area is input into the age detection model, which includes multiple feature analysis modules and feature association modules with processing priorities from high to low, and executes the following in the age monitoring model:

[0011] perform feature extraction on the collected image to determine an age feature corresponding to each target identification region in the collected image;

[0012] For the age feature corresponding to any target identification region, based on each feature analysis module, determine the derivative parameter of the age feature corresponding to each feature analysis module;

[0013] According to the derivative parameter of each feature analysis module, based on the feature correlation module, determine the importance coefficient of each feature analysis module;

[0014] According to the importance coefficient of each feature analysis module and the derivative parameter of each feature analysis module, determine the fusion feature of the collected image;

[0015] According to the fusion feature of each target identification region, determine the matching degree of the fusion feature of the collected image;

[0016] According to the matching degree, determine the age determination result of the antique to be detected.

[0017] In a second aspect, a server system is provided, comprising a server configured to execute the method provided in the first aspect of the present disclosure.

[0018] Compared with the prior art, the present application provides the following beneficial effects: first, through image acquisition and category judgment, the initial identification region matching the antique category can be accurately positioned, the detection range is effectively reduced, and the efficiency of subsequent analysis is improved. Secondly, through process feature detection and multi-dimensional confidence evaluation, the most representative target identification region is selected, avoiding the misjudgment problem caused by improper region selection in traditional methods. Further, by constructing an age detection model containing multiple feature analysis modules, the system can extract age features from different dimensions and realize multi-feature fusion based on the dynamically calculated importance coefficient, significantly improving the comprehensiveness and accuracy of feature representation. This method innovatively introduces a feature correlation module, which quantifies the contribution of different feature analysis modules and adaptively optimizes the fusion strategy, so that the finally generated fusion feature can more accurately reflect the age attribute of the antique. At the same time, the model trained based on big data can be continuously iterated and optimized, with strong generalization and robustness, providing an efficient and reliable intelligent solution for the antique identification field. BRIEF DESCRIPTION OF DRAWINGS

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0020] Figure 1 A schematic diagram of the steps of a method for determining the age of antiques based on artificial intelligence and big data provided by an embodiment of the present invention;

[0021] Figure 2 A schematic diagram of the server structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0023] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0024] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of a method for determining the age of antiques based on artificial intelligence and big data provided in an embodiment of the present disclosure. The method for determining the age of antiques based on artificial intelligence and big data is introduced in detail below.

[0025] S101 , acquiring an image of an antique to be detected, determining a category label of the image by performing category judgment on the image, and determining a plurality of initial identification regions matching the category label from the image.

[0026] In one embodiment, the antique can be optically imaged using a camera, scanner, or other device to obtain an image. A machine learning classification model is then used to analyze the image's global features, determine the antique's general category, such as "porcelain," "calligraphy," or "painting," and generate a category label. Finally, based on the category label, an object detection algorithm, such as YOLO, is used to identify and select all initial identification regions within the image that match the category label. For example, the initial identification region could be the body of a blue and white porcelain vase or decorative patterns.

[0027] After acquiring the image of the antique to be inspected through the acquisition device, the image is first classified and the corresponding category label is generated; then, based on the label, multiple areas that semantically match the category label are located from the image as the initial range for subsequent detailed analysis.

[0028] In another embodiment, the collected data of the antique to be identified can be obtained by a multimodal sensor or manual input method. The collected data includes image information, material information, process information and text information. For example, the image information is a multi-view image or thermal imaging of the antique to be identified, the material information is a chemical composition report obtained by XRF (X-Ray Fluorescence) analysis and a material analysis report obtained by thermoluminescence detection, the process information is the production process of the antique to be inspected, and the text information is the text obtained by scanning the antique to be identified by OCR (Optical Character Recognition), or the historical background text entered by the user. The specific data type of the collected data is not limited in this embodiment of the application. For different application scenarios or needs, such as the identification of the age of different antiques, the collected data will be different.

[0029] Then, feature extraction is performed on the collected data to obtain global features of the antique to be identified. The global features include at least a first feature and a second feature, wherein the first feature is used to determine the category label of the antique to be identified, and the second feature is used to determine the age of the antique to be identified.

[0030] For example, the first feature is the visual feature information extracted by CNN (Convolutional Neural Networks), which includes the shape, material, and texture of the antique to be inspected. For example, if the shape is characterized as a bottle and the material is characterized as porcelain, then based on the first feature, it is determined that the antique to be identified is a porcelain bottle, and then the category label "porcelain-bottle" is added to the antique to be identified.

[0031] It is worth noting that the category label of the antique to be identified can be one or more, but at least includes a label used to characterize the type of the antique to be identified, such as painting, bottle, bowl, tripod, etc.

[0032] The second feature includes the texture feature, color feature, aging feature, and text feature of the antique to be identified, and the category label of the antique to be identified is "porcelain-bottle" based on the first feature. For example, the second feature characterizes that the surface of the porcelain bottle has a blue and white texture and the color is indigo. The aging feature is that the porcelain bottle has rust spots on the surface and the text on the bottom of the porcelain bottle is "Yongle". Therefore, based on the second feature, it can be judged that the antique to be identified is Ming Dynasty blue and white porcelain, and its age is probably during the Yongle period of the Ming Dynasty (1403-1424).

[0033] By adding category labels to antiques to be appraised, the identification points (appraisal areas) of the antiques can be quickly determined, thereby improving the efficiency of determining the age of the antiques to be appraised.

[0034] S102: Perform process feature detection on each of the initial identification areas to obtain process feature data of each of the initial identification areas.

[0035] Among them, the craft feature data include quantifiable features such as the surface texture of antiques (such as the glaze crackle of porcelain), decorative patterns (such as the Taotie pattern on bronze ware), material properties (such as the glossiness of jade), and production traces (such as the patterns of pottery body throwing).

[0036] S103, generating a multi-dimensional detection condition, and determining the confidence of the multiple initial identification areas through a confidence detection model based on the multi-dimensional detection condition, the multiple initial identification areas and their corresponding process feature data.

[0037] The multi-dimensional detection condition is used to perform confidence detection on the initial identification area.

[0038] In one embodiment, the confidence detection model is a random forest neural network model, which is used to map multimodal data into confidence values ​​between 0 and 1.

[0039] It is worth noting that the identification areas used to determine the age of each category of antiques are different. For example, the identification areas of blue and white porcelain may be the bottom signature and the bottle mouth, and the identification areas of bronze ware may be the surface texture of the body, etc. Moreover, even for the identification areas corresponding to antiques of the same category, the confidence levels of the identification areas corresponding to different feature dimensions are different. For example, for antiques classified as porcelain, in the dimension of craftsmanship features, the interior of the porcelain is the identification area with the highest confidence level, in the dimension of material features, the surface of the porcelain is the identification area with the highest confidence level, and in the dimension of text features, the signature on the bottom of the porcelain is the identification area with the highest confidence level.

[0040] The identification area is a specific part of the antique being identified that is valuable for determining its age, such as the inscription area on a bronze artifact or the bottom of a porcelain artifact. The feature dimension represents the type of features that are valuable for determining the age of each identification area. For example, if the identification area is the mouth of a porcelain bottle, the feature dimension might include craftsmanship and color features. If the target identification area is a scroll painting, the feature dimension might include material and texture features.

[0041] For example, the category label of the antique to be identified is "porcelain-bottle" and "blue and white porcelain", and the following query results are obtained by querying the historical database: the verification area includes the blue and white decoration area, the bottle bottom inscription area, and the bottom repair area, wherein the blue and white decoration area is used to judge the color style of blue and white, the corresponding feature dimension of the area is blue and white pigment composition, blur degree, and iron rust spot distribution; the bottle bottom inscription area is used to judge the calligraphy characteristics, the corresponding feature dimension of the area is font structure, stroke thickness, and inscription position; the bottom repair area is used to judge the process characteristics, and the corresponding feature dimension of the area is repair process and firestone red color depth.

[0042] Therefore, a multi-dimensional detection condition is generated to determine the confidence of each initial identification area from different feature dimensions to improve the age monitoring efficiency.

[0043] S104, according to the confidence, selecting the target identification area of the collected image from the plurality of initial identification areas.

[0044] According to the confidence ranking, the initial area with high confidence is selected as the target detection area.

[0045] For example, the confidence of three initial areas of a certain blue and white porcelain plate is 0.92 (decoration area), 0.85 (inscription area), and 0.6 (bottom area), and the threshold is set to 0.7. Finally, the first two initial identification areas are selected as the target identification area.

[0046] By determining the target identification area, unnecessary calculation amount is reduced, and the judgment speed of the age of the antique is improved.

[0047] The collected image with the determined target identification area is input into the age detection model, and in the age detection model, the subsequent steps of the present solution are executed, and the age detection model includes a plurality of feature analysis modules with processing priorities from high to low and a feature association module.

[0048] S105, performing feature extraction on the collected image to determine the age characteristics corresponding to each target identification area in the collected image.

[0049] Among them, the age characteristics are used to represent at least one of the process characteristics, color characteristics and aging characteristics that can reflect the age of the antique to be identified.

[0050] For example, the age detection model includes a CNN network, and the CNN network is used to extract features related to the age of the target detection area, such as the aging degree of porcelain glaze color and the layered structure of bronze rust.

[0051] In an embodiment, first, the target identification region of the collected image and its corresponding feature dimension are decomposed into a plurality of specific feature vectors, then the importance weight of each feature dimension is calculated, the weight of each feature dimension represents the influence of the output feature of the feature dimension on the final output result, and finally, the added feature vectors are merged into the final output age feature. The output feature of each feature dimension is used to represent the similarity between the antique to be identified and the historical antique of each age in the feature dimension, and the fused feature is used to represent the similarity between the target identification region corresponding to the feature and the historical antique of each age.

[0052] S106, for any target identification region corresponding to the age feature, based on each feature analysis module, determine the derived parameters of the age feature corresponding to each feature analysis module.

[0053] The age feature is input into a plurality of feature analysis modules, each module performs feature transformation from different dimensions, and outputs derived parameters related to the function of the module. For example, the plurality of feature analysis modules can include a type analysis module for calculating geometric parameters such as aspect ratio and curvature, a decoration analysis module for extracting parameters such as pattern complexity and symmetry, and a material analysis module for outputting parameters such as spectral features and aging indicators.

[0054] The comprehensive feature vector is decomposed into a plurality of professional dimension derived parameters, which facilitates subsequent targeted analysis.

[0055] S107, according to the derived parameters of each feature analysis module, based on the feature correlation module, determine the importance coefficient of each feature analysis module.

[0056] Through the above method steps, the feature correlation module dynamically allocates the importance coefficient according to the discrimination ability of each derived parameter, and highlights the most critical features for age judgment.

[0057] S108, according to the importance coefficient of each feature analysis module and the derived parameters of each feature analysis module, determine the fusion feature of the collected image.

[0058] S109, according to the fusion feature of each target identification region, determine the matching degree of the fusion feature of the collected image.

[0059] In an embodiment, the fusion feature can be compared with the pre-defined age standard feature template to calculate the matching degree. For example, by calculating the cosine value between the corresponding vector of the fusion feature and the preset age feature vector, the matching degree of the age feature of the collected image is determined.

[0060] For example, the characteristic vector of the fusion feature of Class A antiques is [4.9, 3.2, 1.8]. The corresponding standard characteristic vectors of Class A antiques in the historical database are: [5.1, 2.9, 2.1] for the Ming Dynasty and [4.7, 3.5, 1.6] for the Qing Dynasty. Through matching calculation, it is determined that cos Ming Dynasty = 0.94 and cos Qing Dynasty = 0.97. It can be determined that Class A antiques have a higher matching degree with the Qing Dynasty.

[0061] Through the above method, the multi-dimensional feature space is unified, the objective comparison of antiques from different ages is achieved, and the accuracy of the judgment of the age of antiques is improved.

[0062] S110, determining the age determination result of the antique to be detected based on the matching degree.

[0063] In one embodiment, the final age determination result is output based on the degree of matching of the age features of the captured image, combined with a threshold determination. For example, if the highest degree of matching exceeds a set threshold, the corresponding age is output. If the highest degree of matching is less than the set threshold and the difference is not significant (e.g., the difference is less than a preset difference), a "suspected interval" is output, such as "Kangxi period - Yongzheng period." If the degree of matching is less than the preset threshold and the difference from the threshold is greater than the preset difference, "unable to determine" is output.

[0064] In another embodiment, the age determination result of the antique to be detected can be determined by a probability model, wherein the probability model is a Bayesian classification model. After the fusion features corresponding to each target identification area are input into the probability model, the prior probabilities of the category labels of the antique to be identified in the historical database corresponding to different ages are combined with the fusion features of each target identification area to calculate the conditional probability of the second feature of the antique to be identified in different ages. For example, if the conditional probability of the Xuande period is 0.71, the conditional probability of the Zhengtong period is 0.28, and the conditional probability of the Chenghua period is 0.5, then the age of the antique to be identified can be determined to be the Xuande period. The probability model is used to collect the prior data of the distribution of each age in the historical database, calculate the conditional probability of the antique to be identified belonging to each age, and after obtaining the conditional probability of the antique to be identified, the age area with the highest probability is used as the age judgment result.

[0065] For example, the conditional probability of the antique to be identified is calculated to determine the probability distribution results of each era, that is, the probability of the Xuande era (1426-1435) is 0.78, the probability of the Zhengtong era (1436-1449) is 0.15, and the probability of the Chenghua era (1465-1487) is 0.07. Therefore, the age of the antique to be identified is judged to be the Xuande era (1426-1435).

[0066] Finally, the output age judgment results include the category, age, confidence level and related basis of the antique to be identified. For example, the age judgment results are: Category: Ming Dynasty blue and white porcelain; Age: Xuande period (1426-1435); Confidence level: 78%; Key evidence: 1. The cobalt-iron ratio of the blue and white colorant conforms to the characteristics of "Sumaliqing" (weight 0.33); 2. The double-outline filling technique of the pattern matches the early Xuande period (weight 0.35); 3. The signature on the bottom is a typical feature of the Xuande official style (weight 0.32).

[0067] Through the above method, the present disclosure can, firstly, accurately locate the initial identification area that matches the antique category through image acquisition and category judgment, effectively narrow the detection range, and improve the efficiency of subsequent analysis. Secondly, through process feature detection and multi-dimensional confidence assessment, the most representative target identification area is screened out, avoiding the misjudgment problem caused by improper area selection in traditional methods. Furthermore, by constructing a chronological detection model containing a multi-level feature analysis module, the system can extract chronological features from different dimensions and realize multi-feature fusion based on dynamically calculated importance coefficients, significantly improving the comprehensiveness and accuracy of feature representation. This method innovatively introduces a feature association module, and by quantifying the contribution of different feature analysis modules, it adaptively optimizes the fusion strategy, so that the fusion features finally generated can more accurately reflect the age attributes of antiques. At the same time, the model based on big data training can be continuously iterated and optimized, and has strong generalization and robustness, providing an efficient and reliable intelligent solution for the field of antique identification.

[0068] Optionally, the multiple feature analysis modules include at least a first feature analysis module, a second feature analysis module and at least one other feature analysis module except the first feature analysis module and the second feature analysis module, and each of the feature analysis modules is connected through a fully connected layer.

[0069] In one embodiment, the above step S106 can determine the derived parameters corresponding to each feature analysis module of the age feature by the following method:

[0070] First, based on the first feature analysis module, feature analysis is performed on the collected image to generate derivative parameters of the first feature analysis module.

[0071] Second, based on the second feature analysis module, feature analysis is performed on the derived parameters of the first feature analysis module to determine the derived parameters of the second feature analysis module.

[0072] The first feature analysis module has a higher processing priority than the second feature analysis module, and the second feature analysis module has a higher processing priority than other feature analysis modules thereafter.

[0073] Third, for any of the other feature analysis modules, based on the fully connected layer between the feature analysis module and the feature analysis module before the feature analysis module, the input parameters and output derivative parameters of the feature analysis module before the feature analysis module are merged, and the merged parameters are input into the feature analysis module to generate the derivative parameters of the feature analysis module.

[0074] It's worth noting that feature analysis modules correspond to different feature dimensions, and multiple feature analysis modules have preset processing priorities. These priorities are ranked according to the importance of each dimension's contribution to age determination. The priority of each dimension for each category of antiques is pre-determined based on experience. The preset priorities of the feature analysis modules determine the order of feature extraction, ensuring that high-discriminative features are captured first, thereby improving model efficiency and accuracy.

[0075] Take the bronze artifact to be identified as an example. The target identification area of ​​the bronze artifact is the outer surface of the artifact. The preset priority of the characteristic dimensions of the target identification area is: material characteristics (first characteristic dimension), rust layer structure characteristics (second characteristic dimension), and engraving style characteristics (third characteristic dimension). First, the age characteristics are analyzed by the first characteristic analysis module to obtain the derived parameters of the first characteristic dimension. For example, the copper, tin, and lead ratios are extracted through XRF spectrum data and mapped to a characteristic vector [0.85, 0.12, 0.03]. The characteristic vector indicates that copper accounts for 85%, tin accounts for 12%, and lead accounts for 3%. Then, based on the derived parameters of the first dimension (characteristic Vector), the second feature analysis module further analyzes the secondary key derived parameters. For example, by combining the material composition characteristics, the element distribution of the rust layer (such as the CuCl2 content) is analyzed, or the stacked texture features are extracted through CNN to output the rust feature vector [0.7, 0.3], which represents the degree of oxidation. Finally, the third feature analysis module combines the input parameters of the second feature analysis model (the derived parameters output by the first feature analysis module) and the derived parameters output by the second feature analysis model to generate the derived parameters of the third feature analysis module. The subsequent feature extraction modules all determine the derived parameters in the same way as the third feature extraction module, which will not be repeated here.

[0076] In one embodiment, to determine the derived parameters output by the i-th feature analysis module (i is an integer greater than 2), first, the input parameters of the previous feature analysis module (i-1) (i.e., the derived parameters output by the second feature after the previous i-2 analyses) and the derived parameters output by the previous feature analysis module (i-1) are obtained, and the two are fused, for example, by weighted averaging, concatenation, or residual connection, to obtain the fused parameters. Then, the fused parameters are used as the input parameters of the current dimension (i), and dimensional transformation and nonlinear mapping are performed through operations such as convolution and pooling. The current (i-th) feature analysis module outputs the feature vector of the current dimension (i).

[0077] Since the input of each feature analysis module includes the input parameters and output derivative parameters of the previous feature analysis module, new parameters containing multi-level information are formed, which can retain the original information of the low-level features, avoid feature degradation, and achieve the complementarity and enhancement of cross-dimensional feature information. Accordingly, the fusion parameters obtained according to each feature analysis module will have richer feature information, and then according to the fusion parameters, the age determination results corresponding to the collected data will be more accurate.

[0078] Through this approach, features from high-priority feature dimensions are prioritized, preventing interference from low-priority features in age determination. High-priority features also provide semantic anchors for lower-priority features (e.g., material composition defines the age range of decorative styles), forming a feature reasoning chain from global to local. Furthermore, the preset priority order corresponds to archaeological identification logic (e.g., considering materials first, then craftsmanship), allowing the model's output of age determination to be traced back to specific feature dimensions, consistent with identification practices. Through this hierarchical and cascading design, the system can adaptively integrate feature information from different dimensions, significantly improving the accuracy and robustness of antique dating.

[0079] Optionally, the above step S107 includes the following exemplary implementations:

[0080] In step 1, the derived parameters of the feature analysis module with the lowest processing priority are input into the feature association module to determine a reference feature vector of the feature association module.

[0081] It is worth noting that the lowest priority modules (such as micro-aging analysis) often contain key details for dating and are suitable as reference benchmarks.

[0082] Step 2: Determine the associated feature vector corresponding to each feature analysis module based on the feature association module according to the derived parameters of each feature analysis module.

[0083] Step 3: Determine the importance coefficient of each feature analysis module based on the feature association module according to the reference feature vector and the associated feature vector corresponding to each feature analysis module.

[0084] In one embodiment, a reference feature vector is generated based on the derived parameters output by the feature analysis module with the lowest priority (reference feature vector = W ref ×Derived parameter+b ref , W ref is the weight matrix, b ref is a bias term). Then, a linear transformation is performed on the derived parameters output by each feature analysis module to generate associated feature vectors comparable to the reference feature vector. For example, three feature analysis modules are used to sequentially analyze the target identification region, generating derived parameters along three dimensions. Each feature dimension corresponds to an associated feature vector, i.e., associated feature vectors 1-3. Finally, the similarity between the reference feature vector and associated feature vectors 1-3 is calculated using dot product or cosine similarity. The importance coefficient is sofmax[similarity1, similarity2, ..., similarityn].

[0085] Through the above scheme, the importance coefficient of each dimension is automatically adjusted according to the feature analysis module with the lowest processing priority, so as to achieve adaptive multi-dimensional feature distribution differences of different antiques. By referring to the feature vector, a global analysis standard is established, so that the derived parameters output by the low-priority feature analysis module can guide the distribution of importance coefficients of high-priority features. When the derived parameters output by a certain feature analysis module contain noise (such as traces of restoration), its importance coefficient can also be automatically reduced, thereby improving the anti-interference ability in the process of judging the age of antiques.

[0086] Optionally, the above step S108 includes the following exemplary implementations:

[0087] Step 1: Input the derived parameters of each feature analysis module into the feature association module to determine the feature expression vector of each feature analysis module.

[0088] In one embodiment, for any derivative parameter output by the feature analysis module, a feature expression vector of the feature analysis module is generated through a feature conversion function.

[0089] Among them, the feature conversion function is used to convert the original features into more discriminative feature expression vectors through nonlinear mapping or feature transformation algorithms. The feature conversion functions include: activation functions (such as ReLU, Swish), normalization functions (such as Layer Norm), and feature mapping networks (such as CNN / MLP).

[0090] Step 2: Adjust the importance of the feature expression vector of each feature analysis module through the importance coefficient of each feature analysis module, and determine the fusion feature corresponding to the acquired image by merging each feature expression vector after importance adjustment.

[0091] In one embodiment, a fully connected layer is used to map the derivative parameters of different dimensions to a unified space, determine the feature expression vector of each feature analysis module, ensure that the dimensions of all expression vectors are consistent, and use the importance coefficient as a scaling factor to enhance or suppress the feature expression vector of each module, and then sum them up to obtain the final fusion feature.

[0092] For example, the dating of porcelain is illustrated. The collected data includes multi-angle high-definition images of a blue-and-white porcelain plum vase (including decoration, base mark, and foot), an XRF spectrum analysis report (showing the cobalt composition: 7.2% cobalt, 0.8% manganese, and 6.5% iron), and the OCR-recognized text and image of the base mark "Made in the Xuande Period of the Ming Dynasty." The feature analysis modules for this porcelain in the target inspection area are, in order: blue-and-white color features, decoration style features, calligraphy features, and vessel type features. The derived parameters for each feature analysis module are shown in Table 1 below:

[0093] Table 1

[0094] Then, a feature conversion function is used to generate a feature expression vector for each feature analysis module. For example, for the derived parameters of the blue-and-white color feature, the LeakyRLU activation function is used to amplify the difference in cobalt and iron content and suppress the noise of manganese content, resulting in the first feature expression vector = LeakyReLU(1.5 × original feature - 0.1) = [0.95, 0.05, 0.25]. For the derived parameters of the decorative style feature, the feature weight of the "diffuse rust spot" feature is enhanced through a feature mapping network (MLP), resulting in the second feature expression vector = MLP([0.68, 0.22, 0.1]) = [0.85, 0.1, 0.05].

[0095] The importance coefficient of each feature analysis module may be set in advance based on experience and may be obtained by querying a historical database.

[0096] Finally, the importance coefficient of each feature analysis module and the feature expression vector of each feature analysis module are weighted and summed to obtain the fusion feature corresponding to the target identification area.

[0097] For example, fusion feature = ∑(importance coefficient j ×Feature expression vector j ), where j represents the j-th feature analysis module.

[0098] Through the above scheme, subtle feature differences are amplified, the sensitivity of age distinction is improved, and interference factors in the collected data are avoided. The feature expression vector is adjusted through the importance coefficient, which can be directly mapped to the expert identification logic, simulating the "focus amplification" and "comprehensive weighing" process of key features by cultural relics identification experts, and realizing high-precision judgment of the age of antiques by artificial intelligence.

[0099] Optionally, the feature association module includes multiple feature processing paths for parallel processing.

[0100] In one embodiment, steps S107 to S108 may be performed by the following exemplary embodiments:

[0101] First, for each feature processing path of the feature association module, according to the derived parameters of each feature analysis module, based on the feature processing path, the importance coefficient of each feature analysis module is determined, and the importance of the derived parameters of each feature analysis module is adjusted according to the importance coefficient of each feature analysis module, and each derived parameter after the importance adjustment is merged to determine the merged parameter corresponding to the feature processing path.

[0102] Among them, each feature processing path of the feature association module corresponds to different types of feature associations, such as the association between shape and pattern, and the association between material and craftsmanship.

[0103] Then, the merging parameters of each feature processing path of the feature association module are fused to determine the fusion parameters, and the fusion features corresponding to the collected image are determined based on the fusion parameters.

[0104] In one embodiment, the feature association module may be a multi-dimensional feature fusion network including multiple feature processing channels.

[0105] It is worth noting that each channel has a different weighted transformation matrix to focus on different feature analysis modules (for example, the weight of material features in the material channel accounts for a larger proportion, and the weight of process quality features in the pattern channel accounts for a larger proportion), so as to avoid errors caused by the dominant judgment of a single feature analysis module.

[0106] In one embodiment, the above step S108 may determine the fusion features corresponding to the captured image in the following manner:

[0107] For each feature processing channel of the multi-dimensional feature fusion network, the derived parameters of each feature analysis module are weighted summed by the weight corresponding to the feature processing channel to obtain an enhanced feature representation of the feature processing channel.

[0108] The enhanced feature representations of all feature processing channels are spliced ​​and integrated, and the integration results are subjected to feature compression and semantic mapping to generate fused features corresponding to the target identification area.

[0109] Feature compression involves reducing the number of feature analysis modules through dimensionality reduction (e.g., pooling and convolution) to preserve key information. Semantic mapping involves mapping the compressed features to a chronological semantic space, such as the Shang and Zhou dynasties, the Tang and Song dynasties, and the Ming and Qing dynasties.

[0110] In one embodiment, for each feature processing channel, the contribution ratio of each dimensional feature is adjusted according to the attention weight, thereby generating an enhanced feature representation of each feature processing channel.

[0111] Taking bronze identification as an example, the multi-dimensional feature fusion network includes channel 1 (material analysis channel) and channel 2 (process analysis channel). The weighted transformation matrix corresponding to channel 1 is [0.6, 0.4] (the former in the feature vector represents the probability that the antique belongs to the middle Shang Dynasty, and the latter represents the probability that the antique belongs to the early Western Zhou Dynasty. The feature vectors of the following examples will not be repeated), and the weighted transformation matrix corresponding to channel 2 is [0.3, 0.7]. The derived parameters of the first feature analysis module (material feature) are [0.7, 0.3], and the derived parameters of the second feature analysis module (process feature) are [0.6, 0.4]. Then, the derived parameters of each feature analysis module are weighted and adjusted according to the weight of each feature analysis module. The enhanced feature of channel 1 is expressed as 0.6×[0 .7, 0.3] + 0.4×[0.6, 0.4] = [0.66, 0.34], the enhanced feature representation of channel 2 is 0.3×[0.7, 0.3] + 0.7×[0.6, 0.4] = [0.63, 0.37]. It can be seen that each feature processing channel focuses on different identification directions, and the resulting added feature representations are also different. Then, the enhanced feature representations of channels 1 and 2 are spliced ​​to obtain the splicing result = [enhanced feature representation of channel 1, enhanced feature representation of channel 2]. Finally, the splicing result is convolutionally reduced in dimension, and ReLU activation is applied to enhance the nonlinear expression of the splicing result after dimensionality reduction. Then, the probability results representing the antiques to be identified belonging to different eras are mapped to the age classification space through the fully connected layer to obtain the fusion features corresponding to the collected image.

[0112] Through the above technical solution, different feature analysis modules are focused through different feature processing channels, avoiding the misjudgment of age identification caused by a single feature analysis module. For example, when the bottom pattern of the antique to be identified is forged, its age can still be determined by the color and pattern.

[0113] In one embodiment, the above step S101 may determine a plurality of initial detection areas corresponding to category labels in the captured image by the following exemplary implementation:

[0114] Step 1: determining the associated antique data corresponding to the category label of the antique to be identified by querying the historical database.

[0115] Step 2: Input the associated antique data into a multimodal analysis model to obtain confidence scores for multiple associated verification areas.

[0116] Step 3: Based on the confidence score of each of the associated verification areas and a preset confidence score, determine a plurality of initial detection areas corresponding to the category label in the plurality of associated verification areas.

[0117] For example, the category label of the antique to be identified is "porcelain-bottle". By querying the historical database, multiple historical antiques with the label "porcelain-bottle" are obtained. Among the historical antiques queried, 80% of the historical antiques have the corresponding associated verification areas of the bottle mouth and bottle bottom, and 20% of the historical antiques have the corresponding associated verification areas of the bottle body and bottle bottom. The associated verification areas of each historical antique (associated antique data) are input into the multimodal analysis model respectively to obtain the confidence score of each associated verification area. For example, the bottle body is 0.85, the bottle bottom is 0.92, and the bottle bottom is 0.65. The preset confidence score is set to 0.7. The confidence scores of the bottle body and bottle bottom are greater than 0.7, then the bottle body and bottle bottom are determined to be the initial detection areas.

[0118] Optionally, the associated antique data includes a plurality of historical antiques with the category label, a plurality of associated verification areas corresponding to each of the historical antiques, and a plurality of feature dimensions corresponding to each of the associated verification areas.

[0119] In an exemplary embodiment, the above step 2 can be performed by the following implementation:

[0120] The multiple associated verification areas corresponding to each of the historical antiques are respectively input into the multimodal analysis model, and the confidence score of each associated verification area is determined according to the proportion of each associated verification area compared to all the input associated verification areas.

[0121] In an exemplary embodiment, the above step 3 can be performed by the following implementation:

[0122] The associated verification area having the confidence score greater than the preset confidence score is determined as the initial detection area, and a plurality of feature dimensions are determined based on the associated verification area determined as the initial detection area.

[0123] The confidence score of the associated verification area is used to represent the frequency of the area in the historical samples corresponding to the category. The higher the confidence score, the higher the probability of determining the age of the antique to be identified through the area. The lower the confidence score, the lower the probability of determining the age of the antique to be identified through the area. Different pre-set confidence scores can be set for different antiques. For example, the pre-set confidence score of a category-specific cultural relic (such as an official kiln porcelain) is set to 0.7-0.8, and the pre-set confidence score of a category-diversified cultural relic (such as a folk kiln porcelain) is set to 0.6-0.7.

[0124] Through the above scheme, the method determines the relevant historical data from big data through the category label of the antique to be identified, and automatically focuses on the most critical identification area according to the confidence scores of different associated verification areas, thereby significantly improving the accuracy and intelligent level of antique age determination.

[0125] Optionally, the collected images include a plurality of initial images of the antique to be identified from multiple perspectives, and the collected images include text data.

[0126] In an embodiment, the above step S101 can determine a plurality of initial identification areas matching the category label from the collected images through the following example embodiment:

[0127] Step 1: Generate a multi-dimensional matching condition corresponding to the category label.

[0128] Step 2: Determine the first matching degree between each initial image and the multi-dimensional matching condition.

[0129] Step 3: Perform text reading on each initial image to determine the text part corresponding to each initial image.

[0130] Step 4: Extract a plurality of first keywords from the text data.

[0131] Step 5: Match a second keyword corresponding to the category label from the plurality of first keywords.

[0132] Step 6: Determine the second matching degree between each text part and the second keyword.

[0133] Step 8: For each initial image, if the first matching degree reaches a preset first matching degree and the second matching degree reaches a preset second matching degree, the initial image is determined as an initial identification area matching the category label.

[0134] In an embodiment, the multi-dimensional matching condition includes a text description vector and a visual feature vector.

[0135] Wherein, the text description vector is a semantic vector converted from the category label, and the visual feature vector is a pre-defined typical feature vector on the vision.

[0136] Then, for each of the historical antique, the following double threshold matching mechanism is executed:

[0137] ①Feature extraction is performed on the initial image to generate an image feature vector, and a first matching degree between the antique to be identified and the historical antique is determined according to the cosine similarity between the image feature vector and the visual feature vector.

[0138] ②Target detection is performed on the initial image to obtain a key region with text description in the initial image; keyword extraction is performed on the key region to obtain a plurality of keyword vectors conforming to the category label; and a second matching degree between each keyword vector and the text description vector is determined.

[0139] Wherein, the first matching degree is used to represent the similarity between the overall image features of the initial image and the visual feature vector, and the second matching degree is used to represent the similarity between the text data of the initial image and the corresponding semantic vector converted from the category label.

[0140] Finally, the initial image is determined as the initial identification region of the matching category label when the first matching degree reaches a preset first matching degree and the second matching degree reaches a preset second matching degree.

[0141] The following takes the category label "landscape painting" as an example for illustrative description. First, the category label "landscape painting" is converted into a semantic vector through the BERT model. And the visual feature vector is constructed based on historical data, for example, including the following features: psimahai coining (drawing method), white space composition (composition method), ink color (color).

[0142] Then, the CNN features of the historical painting are extracted, the cosine similarity between the CNN features and the visual feature vector is calculated, and the cosine similarity is determined as the first matching degree. In addition, the key regions in the painting, such as the inscription, seal, and colophon, are identified through target detection, the text content is extracted from these key regions through OCR and converted into keyword vectors, the cosine similarity between the keyword vectors and the text description vector is calculated, and the cosine similarity is determined as the second matching degree.

[0143] Finally, when the first matching degree of the historical antique is greater than or equal to a preset first threshold (such as 0.8), and the second matching degree of at least one keyword vector is greater than or equal to a preset second threshold (such as 0.9), the historical antique is the target historical antique, for example, the Ming Dynasty Dehua blue and white porcelain vase (the first matching degree is 0.88, and the second matching degree is 0.92) and the Ming Dynasty Chenghua blue and white porcelain cup (the first matching degree is 0.85, and the second matching degree is 0.9) are target historical antiques, and the initial image is determined as an initial identification region of a matching category label.

[0144] This method improves matching accuracy by multi-modal information fusion, combines image visual features and text semantic features, realizes accurate retrieval of historical data, and significantly improves the accuracy and intelligent level of antique age determination.

[0145] Optionally, the collected image includes text data, and after the step S101 is executed, the method further includes:

[0146] First, a third matching degree between each initial identification region and the text data is determined.

[0147] The third matching degree is used to measure the correlation degree of the initial identification region and the image text data, and the higher the value, the stronger the matching of the region visual feature and the text semantic. For example: the matching degree of the mark region and the text "Ming Dynasty Dehua".

[0148] Then, the plurality of initial identification regions are sequentially sorted in order of the third matching degree from high to low to determine the arrangement order.

[0149] Secondly, for the initial identification regions except the last one in the arrangement order, the corresponding adjacent proportion relationship of each initial identification region is determined based on the proportional relationship between the third matching degrees of the adjacent two initial identification regions.

[0150] It is worth noting that the third matching degree ratio (the matching degree of the previous region / the matching degree of the next region) of the adjacent two regions in the arrangement order is used to reflect the change amplitude of the matching degree. For example: 0.95 / 0.85=1.12.

[0151] Finally, in the case that the adjacent proportion relationship corresponding to any initial identification region is greater than a preset proportion threshold, all initial identification regions up to the initial identification region in the arrangement order are retained, and the remaining initial identification regions are deleted.

[0152] The preset proportion threshold is a proportion critical value set by human, which is used to judge whether the matching degree has a sudden drop in correlation degree. For example: when the threshold is set to 2.0, if the adjacent proportion is greater than 2.0, it is considered that the correlation degree of the next region and the text is significantly reduced.

[0153] For example, the image set is a Ming Dynasty blue and white porcelain bottle containing the text "Xuan De Nian Zhi". The initial identification area includes R1: the bottle body twined branch pattern area; R2: the bottom mark area, R3: the background wooden base, and R4: the bottle body stain area. Then, the third matching degree between each initial identification area and the text data is determined to be M1=0.8 (the pattern and the "blue and white" semantic match), M2=0.98 (the mark and the "Xuan De Nian Zhi" are completely matched), M3=0.1 (the base is unrelated to the text), M4 = 0.05 (no correlation with stains); secondly, determine the permutation order = [R2, R1, R3, R4], with adjacent ratios: r1 = 0.98 / 0.8 = 1.22 (R2 → R1, the match decreases gently), r2 = 0.8 / 0.1 = 8 (R1 → R3, the match decreases sharply), and r3 = 0.1 / 0.05 = 2 (R3 → R4, a decrease but not the first mutation point); finally, at the first position where r > 2.0, j = 2 (corresponding to R1 → R3), retain the first two regions {R2, R1} and delete {R3, R4}. The screening result retains the inscription and bottle body decoration areas that are strongly correlated with the "Xuan De Nian Zhi" text, eliminating background interference.

[0154] This solution achieves semantically guided regional precision positioning through cross-modal matching of text and vision and ratio threshold screening. It not only uses text information to improve identification accuracy, but also adapts to different image scenes through dynamic ratio strategies, providing high-quality target area input for subsequent age feature analysis.

[0155] Optionally, the collected image includes text data and multiple initial images, and at least one of the multiple initial images includes an initial identification area.

[0156] In one embodiment, the above step S101 may also determine the category label of the captured image by:

[0157] The initial identification area is subjected to feature extraction by a comprehensive feature extractor to determine an outline vector.

[0158] The comprehensive feature extractor is used to extract features from the text data to determine a description vector.

[0159] The combined vector of the shape vector and the description vector is matched with a preset category label by a category label matcher to determine the category label of the acquired image.

[0160] Among them, the appearance vector is to extract visual features from the initial identification area, such as the shape of the vessel, patterns, etc., and the description vector is to extract semantic features from text data, such as inscriptions, documentary records, etc.

[0161] For example, the initial identification region (e.g., the inscription or decorative pattern on a porcelain artifact) is fed into a comprehensive feature extractor. The comprehensive feature extractor then performs the following steps to extract features from the initial identification region: ① Image preprocessing: Normalize and scale the initial identification region. ② Feature extraction: Extract visual features using a CNN network. ③ Dimensionality reduction: Global average pooling is performed to obtain a fixed-length vector. ④ Output is a shape vector.

[0162] Text data (such as "Made in the Xuande Period of the Ming Dynasty" identified by OCR) is fed into a comprehensive feature extractor. The comprehensive feature extractor then performs the following steps to extract features from the text data: ① Text preprocessing: word segmentation and stop word removal; ② Semantic encoding: encoding into vectors using a pre-trained language model (such as BERT); ③ Pooling: taking the token vector as the overall representation; and ④ Outputting a description vector.

[0163] Finally, the merged vector and the description vector are input into the category label matcher, in which the following processing steps are performed: ① Similarity calculation: for each preset category label Tᵢ (such as "blue and white porcelain"), the similarity sim(V, Tᵢ) is calculated; ② Sorting and decision: the label with the highest similarity is selected as the prediction result; ③ Output category label (such as "Ming Dynasty blue and white porcelain").

[0164] Through this multimodal fusion solution, the system can more comprehensively understand the characteristics of cultural relics, achieve more accurate category determination, and lay the foundation for subsequent chronological analysis.

[0165] Optionally, the captured image includes a signature image, and determining the category label of the captured image by a category label matcher includes the following implementations:

[0166] The combined vector of the appearance vector and the description vector is input into the category label matcher to obtain multiple category confidences of the antique to be identified.

[0167] When the category confidence is not higher than a preset category confidence threshold, the signature image is identified to obtain signature description information.

[0168] Extract keywords from the signature description information and determine a third keyword.

[0169] Category matching is performed in a large database based on the third keyword to determine the category label of the captured image.

[0170] In one implementation, a combined vector of the shape vector (visual features) and the description vector (text semantics) is input into a category label matcher, which outputs predicted probabilities (confidence levels) for multiple categories (e.g., "Ming Dynasty blue and white porcelain" and "Qing Dynasty famille rose porcelain"). If the highest confidence level exceeds a threshold, the category is directly confirmed; otherwise, the signature verification process is triggered.

[0171] For example, the signature verification process is as follows: perform OCR recognition on the signature image (such as the signature on the bottom of porcelain), extract text information, and use the named entity recognition (NER) algorithm to extract key entities (dynasty, reign, kiln, etc.) from the text information to form a third keyword set. For example, if the signature text is "Made in the Chenghua Period of the Ming Dynasty", the third keyword can be obtained: ["Da Ming", "Chenghua", "Nianzhi"].

[0172] Then, the third keyword is matched against predefined classification rules in the historical database, and the category label is modified based on the matching results. For example, if the third keyword contains "Xuan De" and the initial classification confidence is ≥ 0.5, the modified category is "Xuan De blue and white porcelain of the Ming Dynasty".

[0173] Through the above scheme, the misjudgment rate of antiques with different styles across different eras is reduced, and the category labels can be traced back to the specific basis of the signature, thereby improving the confidence level in judging the age of the antiques.

[0174] In one embodiment, the above step S102 includes the following methods:

[0175] Step 1: Determine the material confidence detection value corresponding to each of the initial identification areas through a material evaluation model.

[0176] The material assessment model analyzes the initial identification area and outputs a numerical value reflecting the material reliability of that area, known as the material confidence value. This model can be trained based on information such as spectral analysis and texture features, such as the clay composition of porcelain and the alloy ratio of metal objects.

[0177] Step 2: Determine the color confidence detection value corresponding to each of the initial identification areas through a color evaluation model.

[0178] The color assessment model is used to quantify properties such as the color and gloss of a region, generating a color confidence value. For example, the intensity of the blue and white color of blue and white porcelain and the glossiness of jade can be converted into specific values ​​through the model.

[0179] Step three: determining the aging assessment value corresponding to each of the initial identification areas through an aging assessment model.

[0180] The aging assessment model analyzes the degree of aging in the initial identification area and outputs an aging assessment value. For example, aging phenomena include cracking of porcelain glaze, rust on metal objects, and patina on wood. The model analyzes these characteristics to generate an assessment value.

[0181] Step 4: determining a preset aging value corresponding to the category label in a historical big data database, and determining a fourth matching degree between each of the aging assessment values ​​and the preset aging value.

[0182] In the historical big data database, the preset aging value corresponding to the category label is searched, and then the fourth matching degree between the aging assessment value of each initial identification area and the preset aging value is calculated to measure the degree of consistency between the actual aging characteristics and the typical aging characteristics of the category.

[0183] Step 5: Determine the process feature data corresponding to each of the initial identification areas based on the material confidence detection value, the color confidence detection value, and the fourth matching degree.

[0184] The material confidence test value, color confidence test value and the fourth matching degree are combined to form quantitative data that comprehensively reflects the regional process characteristics, providing a basis for subsequent identification.

[0185] In one embodiment, the following is explained by taking the identification of suspected Song Dynasty Ru kiln porcelain as an example:

[0186] First, the characteristics of the initial identification area were evaluated through three dimensions: ① Material evaluation: The material evaluation model was used to analyze the data such as the alumina and silica content of the porcelain body, and the output material confidence detection value was 0.85, indicating that the material in this area is highly similar to that of Ru kiln porcelain; ② Color evaluation: The color evaluation model analyzed the RGB value and glossiness of the porcelain's sky blue glaze, and obtained a color confidence detection value of 0.78; ③ Aging evaluation: The aging evaluation model was used to observe the morphology, density and other characteristics of the glaze crackle, and the output aging evaluation value was 0.8.

[0187] Then, the preset aging value of Ru kiln porcelain in the Song Dynasty was found in the historical big data database, and the fourth matching degree between the current aging evaluation value and the preset value was calculated to be 0.75.

[0188] Finally, the material confidence detection value, color confidence detection value and the fourth matching degree are combined to obtain the final process characteristic data of the initial identification area, providing a reference for judging the authenticity and age of the porcelain.

[0189] By evaluating from multiple dimensions such as material, color, and aging, we can avoid misjudgment caused by a single factor and more comprehensively reflect the craftsmanship characteristics of antiques. By matching with preset values ​​in the historical big data database and combining the quantitative output of multiple models, we can improve the accuracy of the craftsmanship characteristic data and make the identification results more reliable.

[0190] It should be noted that in each specific embodiment of the present disclosure, when it comes to the need to perform relevant processing based on data related to the characteristics of the target object, such as the target object attribute information or attribute information set, the permission or consent of the target object will be obtained first, and the collection, use and processing of these data will comply with relevant laws, regulations and standards. Among them, the target object can be a user. In addition, when the embodiment of the present disclosure needs to obtain the attribute information of the target object, the separate permission or separate consent of the target object will be obtained through a pop-up window or jump to a confirmation page. After clearly obtaining the separate permission or separate consent of the target object, the necessary target object-related data for the normal operation of the embodiment of the present disclosure will be obtained.

[0191] In the embodiments of the present disclosure, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0192] The method for determining the age of antiques described in the embodiments of this application belongs to the field of artificial intelligence (AI). As a cutting-edge branch of computer science, AI aims to simulate, extend, and expand human intelligence through digital computers or intelligent devices, giving machines the ability to perceive their environment, learn knowledge, and make decisions based on that knowledge. Its core is to build systems that can respond in a human-like intelligent manner.

[0193] Artificial intelligence technology is a comprehensive discipline encompassing a multi-dimensional technical system encompassing both hardware and software. At the foundational level, it encompasses sensors, specialized AI chips, cloud computing, distributed storage, big data processing, and operational / interactive systems, providing the underlying support for algorithm execution and data processing. Software technologies focus on areas such as computer vision, natural language processing, and machine learning / deep learning. Machine learning / deep learning is crucial for accurately determining the age of antiques.

[0194] Computer vision technology plays a key role in this approach. Through image recognition and feature extraction, it analyzes visual information such as the texture, shape, and craftsmanship of antiques. Natural language processing is used to parse textual data such as historical documents and appraisal reports, mining semantic information related to the antique's age. Machine learning and deep learning algorithms can automatically learn and extract age-distinguishing feature patterns from massive amounts of antique images and text, offering greater adaptability and accuracy than traditional methods.

[0195] The artificial intelligence and big data-based antique age determination method provided in the embodiments of the present application can be executed on a server or terminal device, typically a server, such as a data processing server or a model training server. The server can be a standalone physical server, or it can be composed of multiple physical servers forming a server cluster, a distributed system, or a cloud server providing cloud computing services. Terminal devices include, but are not limited to, smartphones, tablet computers, and desktop computers. A direct or indirect connection between the terminal device and the server can be established via wired or wireless communication.

[0196] Data processing and model training are primarily implemented using cloud computing technology. Cloud computing integrates grid computing, distributed computing, and parallel computing technologies, providing computing resources and services on-demand and flexibly scalable over the network. In the context of antique dating, cloud computing can efficiently handle massive amounts of antique image data, historical document data, and the large-scale computing tasks generated by model training. With the development of internet technology, the expansion of data volumes, and the growing demand for digital cultural relic authentication, cloud computing provides a solid technical foundation for intelligent and efficient antique dating, driving the transformation of cultural relic authentication models towards digitalization and intelligence.

[0197] The embodiment of the present invention provides a server 100, which includes a feature extractor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the feature extractor, the server 100 executes the aforementioned method for determining the age of antiques based on artificial intelligence and big data. Figure 2 As shown, Figure 2 This is a block diagram of the structure of a server 100 provided in an embodiment of the present invention. Server 100 includes a memory 111, a feature extractor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, feature extractor 112, and communication unit 113 are electrically connected to each other directly or indirectly.

[0198] An embodiment of the present invention provides a readable storage medium, which includes a computer program. When the computer program is running, it controls the computer device where the readable storage medium is located to execute the aforementioned method for determining the age of antiques based on artificial intelligence and big data.

[0199] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Based on the above teachings, numerous modifications and variations are possible. These embodiments have been determined and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with different modifications to suit the specific applications contemplated. For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Based on the above teachings, numerous modifications and variations are possible. These embodiments have been determined and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with different modifications to suit the specific applications contemplated.

Claims

1. A method for determining the age of antiques based on artificial intelligence and big data, characterized in that: include: Acquire a captured image of the antique to be inspected, determine a category label of the captured image by performing category judgment on the captured image, and determine a plurality of initial identification regions matching the category label from the captured image; Performing process feature detection on each of the initial identification areas to obtain process feature data of each of the initial identification areas; generating a multi-dimensional detection condition for performing a confidence test on the initial identification region, and determining the confidence of the plurality of initial identification regions using a confidence test model based on the multi-dimensional detection condition, the plurality of initial identification regions and their corresponding process feature data; selecting a target identification area of ​​the captured image from the plurality of initial identification areas according to the confidence level; Inputting the captured image of the target identification area into an age detection model to determine the age determination result of the antique to be detected, the age detection model includes multiple feature analysis modules and feature association modules with processing priorities ranging from high to low, each feature analysis module corresponding to a different feature dimension, and the processing priority is sorted according to the importance of the contribution of the feature dimension corresponding to the feature analysis module to the age determination; The collected image of the target identification area is input into the age detection model to determine the age determination result of the antique to be detected, including: Performing feature extraction on the collected image to determine age features corresponding to each target identification area in the collected image; For the age feature corresponding to any target identification area, based on each of the feature analysis modules, determining the derived parameters of the age feature corresponding to each of the feature analysis modules; Determining the importance coefficient of each feature analysis module based on the feature association module according to the derived parameters of each feature analysis module; Determining the fusion features of the acquired image according to the importance coefficient of each feature analysis module and the derived parameters of each feature analysis module; Determining the matching degree of the fusion features of the acquired image according to the fusion features of each target identification area; Determining the age of the antique to be detected based on the matching degree; The step of determining the importance coefficient of each feature analysis module based on the derived parameters of each feature analysis module and the feature association module includes: Inputting the derived parameters of the feature analysis module with the lowest processing priority into the feature association module to determine a reference feature vector of the feature association module; Determining, according to the derived parameters of each feature analysis module and based on the feature association module, an associated feature vector corresponding to each feature analysis module; Determine, based on the feature association module, the importance coefficient of each feature analysis module according to the reference feature vector and the associated feature vector corresponding to each feature analysis module, and calculate the similarity between the reference feature vector and the associated feature vector by dot product calculation or cosine similarity, where importance coefficient = sofmax[similarity1, similarity2, ..., similarityn]; Determining the fusion features of the acquired image according to the importance coefficient of each feature analysis module and the derived parameters of each feature analysis module includes: Inputting the derived parameters of each feature analysis module into the feature association module to determine the feature expression vector of each feature analysis module; The importance of each feature analysis module's feature expression vector is adjusted using the importance coefficient of each feature analysis module, and the fusion feature corresponding to the acquired image is determined by merging each feature expression vector after importance adjustment.

2. The method according to claim 1, characterized in that The multiple feature analysis modules include at least a first feature analysis module, a second feature analysis module, and at least one other feature analysis module except the first feature analysis module and the second feature analysis module, and each of the feature analysis modules is connected through a fully connected layer; The step of determining, based on each of the feature analysis modules, a derivative parameter of the age feature corresponding to each of the feature analysis modules includes: Based on the first feature analysis module, performing feature analysis on the collected image to generate derivative parameters of the first feature analysis module; performing feature analysis on the derived parameters of the first feature analysis module based on the second feature analysis module to determine the derived parameters of the second feature analysis module, wherein the processing priority of the first feature analysis module is higher than that of the second feature analysis module; For any of the other feature analysis modules, based on the fully connected layer between the feature analysis module and the feature analysis module before it, the input parameters and output derivative parameters of the feature analysis module before it are merged, and the merged parameters are input into the feature analysis module to generate the derivative parameters of the feature analysis module.

3. The method according to claim 1, characterized in that The feature association module includes multiple feature processing paths for parallel processing, each of which corresponds to a different type of feature. Determining the importance coefficient of each feature analysis module based on the derived parameters of each feature analysis module and the feature association module includes: For each feature processing path of the feature association module, determining an importance coefficient of each feature analysis module based on the feature processing path according to the derived parameters of each feature analysis module; Determining the fusion features of the acquired image according to the importance coefficient of each feature analysis module and the derived parameters of each feature analysis module includes: Adjusting the importance of the derived parameters of each feature analysis module according to the importance coefficient of each feature analysis module, merging each derived parameter after the importance adjustment, and determining the merged parameter corresponding to the feature processing path; The merging parameters of each feature processing path of the feature association module are fused to determine the fusion parameters, and the fusion features corresponding to the collected image are determined based on the fusion parameters.

4. The method according to claim 1, wherein The collected images include multiple initial images of the antique to be detected under multiple viewing angles, the collected images include text data, and determining multiple initial identification areas matching the category labels from the collected images includes: Generating multi-dimensional matching conditions corresponding to the category labels; determining a first degree of matching between each of the initial images and the multi-dimensional matching condition; Reading text from each of the initial images to determine a text portion corresponding to each of the initial images; performing keyword extraction on the text data to determine a plurality of first keywords; Matching a second keyword corresponding to the category label among the plurality of first keywords; determining a second degree of matching between each of the text portions and the second keyword; For each of the initial images, when the first matching degree reaches a preset first matching degree and the second matching degree reaches a preset second matching degree, the initial image is determined as an initial identification area matching the category label.

5. The method according to claim 1, wherein The collected image includes text data. After determining a plurality of initial identification areas matching the category labels from the collected image, the method further includes: determining a third degree of matching between each of the initial identification regions and the text data; Sorting the plurality of initially identified regions in descending order of the third matching degree to determine an arrangement order; For the initial identification regions except the last one in the arrangement order, determining the adjacent proportional relationship corresponding to each of the initial identification regions based on the proportional relationship between the third matching degrees of two adjacent initial identification regions; When the adjacent ratio relationship corresponding to any of the initial identification regions is greater than a preset ratio threshold, all the initial identification regions up to the initial identification region in the arrangement order are retained, and the remaining initial identification regions are deleted.

6. The method according to claim 1, characterized in that The collected image includes text data and a plurality of initial images, at least one of the plurality of initial images includes an initial identification area, and determining a category label of the collected image by performing category judgment on the collected image includes: Extracting features from the initial identification area using a comprehensive feature extractor to determine an outline vector; Performing feature extraction on the text data by the comprehensive feature extractor to determine a description vector; The combined vector of the shape vector and the description vector is matched with a preset category label by a category label matcher to determine the category label of the acquired image.

7. The method according to claim 6, characterized in that The collected image includes a signature image, and the class label matcher matches a combined vector of the shape vector and the description vector with a preset class label to determine the class label of the collected image, including: Inputting the combined vector of the shape vector and the description vector into the category label matcher to obtain multiple category confidences of the antique to be detected; When the category confidence is not higher than a preset category confidence threshold, identifying the signature image to obtain signature description information; Extracting keywords from the signature description information to determine a third keyword; Category matching is performed in a large database based on the third keyword to determine the category label of the captured image.

8. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 7.

Images