An intelligent identification method and system for the aging degree of porcelain glaze

Abstract

The present invention discloses an intelligent identification method and system for the aging degree of porcelain glaze, including: First, a multispectral imaging device is used to collect the reflection characteristics of the target porcelain in the visible light and near-infrared bands, and the original image is processed to generate a standardized porcelain image to be identified, including the overall or partial image. The image is input into a pre-trained target model to obtain the target values of the aging degree grading of each glaze area, and then the aging degree grading result of the target porcelain is determined. Finally, the result is fused with the trace element data detected by X-ray fluorescence spectroscopy to generate an identification report including the age inference and the evaluation of the glaze preservation state, realizing the intelligent and accurate identification of the aging degree of porcelain glaze.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more particularly, to an intelligent identification method and system for the aging degree of porcelain glaze. Background Art

[0002] The traditional identification of the aging degree of porcelain glaze mainly relies on expert experience and limited instrument detection, which has problems such as strong subjectivity, low efficiency, and large influence of human factors on accuracy. Expert identification may produce different conclusions due to differences in personal knowledge and experience, and it is difficult to accurately judge some complex situations. Instrument detection often can only provide data in a single dimension and cannot comprehensively and synthetically evaluate the aging degree of porcelain glaze. Therefore, there is an urgent need for an objective, accurate, and efficient intelligent identification method to meet the growing needs in the field of porcelain identification. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent identification method and system for the aging degree of porcelain glaze.

[0004] In a first aspect, an embodiment of the present invention provides an intelligent identification method for the aging degree of porcelain glaze, including:

[0005] Collect the reflection characteristics of the target porcelain in the visible and near-infrared bands through a multispectral imaging device, perform glaze specular reflection suppression processing on the collected original image, and generate a standardized porcelain image to be identified; the porcelain image to be identified is a porcelain image of a preset image type of the target porcelain, and the preset image type is an overall image or a partial image;

[0006] Load the porcelain image to be identified into a pre-trained target glaze aging degree identification model, and obtain the glaze aging degree target value of the aging degree classification result of each glaze area in the porcelain image output by the target glaze aging degree identification model;

[0007] Determine the target aging degree classification result of the target porcelain according to the glaze aging degree target value of the aging degree classification result of each glaze area in the porcelain image;

[0008] Perform spatio-temporal alignment and fusion on the target aging degree classification result and the trace element data detected by X-ray fluorescence spectroscopy to generate an identification report on the aging degree of porcelain glaze including the inferred result of the age and the evaluation of the glaze preservation state.

[0009] In a second aspect, an embodiment of the present invention provides a server system, including a server, and the server is used to execute the method described in the first aspect.

[0010] Compared with the prior art, the beneficial effects provided by the present invention include: adopting an intelligent identification method and system for the aging degree of porcelain glaze surface disclosed in the present invention, collecting the reflection characteristics of the target porcelain in the visible light and near-infrared bands by using a multispectral imaging device, processing the original image to generate a standardized porcelain image to be identified, including the overall or partial image. Inputting the image into a pre-trained target model to obtain the target values of the aging degree grading of each glaze area, and then determining the aging degree grading result of the target porcelain. Finally, fusing the result with the trace element data detected by X-ray fluorescence spectroscopy to generate an identification report including the age inference and the evaluation of the glaze preservation state, realizing the intelligent and accurate identification of the aging degree of the porcelain glaze surface. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0012] Figure 1 It is a schematic flowchart of the steps of the intelligent identification method for the aging degree of porcelain glaze surface provided by the embodiment of the present invention;

[0013] Figure 2 It is a schematic block diagram of the structure of the computer device provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. The components of the embodiments of the present invention usually described and shown in the drawings here can be arranged and designed in various different configurations.

[0015] The following will describe in detail the specific embodiments of the present invention in conjunction with the accompanying drawings.

[0016] To solve the technical problems in the foregoing background art, Figure 1 It is a schematic flowchart of the intelligent identification method for the aging degree of porcelain glaze surface provided by the embodiment of the present disclosure. The intelligent identification method for the aging degree of porcelain glaze surface will be introduced in detail below.

[0017] Step S201: Collect the reflection characteristics of the target porcelain in the visible and near-infrared bands through a multispectral imaging device, and perform glaze specular reflection suppression processing on the collected original images to generate a standardized porcelain image to be identified; the porcelain image to be identified is a porcelain image of a preset image type of the target porcelain, and the preset image type is an overall image or a partial image;

[0018] Step S202: Load the porcelain image to be identified into a pre-trained target glaze aging degree identification model, and obtain the glaze aging degree target value of the aging degree grading result of each glaze area in the porcelain image output by the target glaze aging degree identification model;

[0019] Step S203: Determine the target aging degree grading result of the target porcelain according to the glaze aging degree target value of the aging degree grading result of each glaze area in the porcelain image;

[0020] Step S204: Perform spatio-temporal alignment and fusion on the target aging degree grading result and the trace element data detected by X-ray fluorescence spectroscopy to generate a porcelain glaze aging degree identification report including the age inference result and the glaze preservation state evaluation.

[0021] In an embodiment of the present invention, exemplarily, the server controls a multispectral imaging device to collect images of a target porcelain. Suppose the target porcelain is a Ming Dynasty blue and white porcelain vase. The multispectral imaging device will scan it in the visible light and near-infrared bands to obtain the reflection characteristics of the target porcelain in these bands. During the collection process, the device will take pictures of the blue and white porcelain vase from different angles and distances to ensure that its surface features are comprehensively captured. For example, first take a picture from directly above to obtain an image of the top of the bottle body, then take pictures of the main part of the bottle body from different side angles, and also take close-up pictures of details such as the bottleneck and the bottom of the bottle. The original images collected will have the problem of glaze reflection, which will affect the accuracy of subsequent identification. After receiving the original images, the server will immediately perform glaze reflection suppression processing on them. Taking an area with relatively serious reflection on the bottle body as an example, the server analyzes the pixel value distribution of this area through a specific algorithm, identifies the high-brightness pixels in the reflective part. Then, according to the pixel characteristics of the surrounding normal area, the pixel values of the reflective area are adjusted to reduce its brightness and make it more consistent with the brightness of the surrounding area. After processing, the server will generate a standardized porcelain image to be identified. The preset image type here is an overall image or a partial image. For this blue and white porcelain vase, the server will generate an overall image of it to fully display information such as the shape and decoration of the bottle body; at the same time, it will also generate partial images of some key parts, such as areas with typical decorations on the bottle body and areas with obvious changes in the glaze surface, etc., for more accurate subsequent analysis of the aging degree. The server loads the generated porcelain image to be identified into a pre-trained target glaze aging degree identification model. Suppose the image of the target porcelain includes an overall image and multiple partial images. The server will sequentially input these images into the model. After receiving the images, the model will analyze and process the images. For the overall image, the model will first identify the general outline of the porcelain and the overall glaze characteristics, and judge the overall color and glossiness of the glaze. Taking the overall image of the blue and white porcelain vase as an example, the model can detect the overall color distribution of the glaze on the bottle body and whether there is uneven color. For partial images, the model will analyze the characteristics of each glaze area more carefully. For example, in a partial image with crackles, the model will detect the density and shape of the crackles and analyze the color change at the crackles; at the same time, it will also observe the distribution characteristics of microscopic bubbles, including the size, quantity, and distribution law of the bubbles. After the analysis and processing of the model, the glaze aging degree target value of the aging degree classification result of each glaze area in the porcelain image is finally output. For example, for a certain glaze area on the main part of the bottle body of the blue and white porcelain vase, the aging degree target value output by the model may indicate that this area is in a moderately aged state; while for a small area at the bottleneck, the output target value shows that its aging degree is relatively light. After receiving the glaze aging degree target value of the aging degree classification result of each glaze area output by the model, the server will comprehensively analyze these values to determine the target aging degree classification result of the target porcelain.For this blue and white porcelain vase, the server will calculate the area proportion of the glaze areas with different aging degree classification results. Suppose the target value of the aging degree in most areas of the vase body shows moderate aging, and only a small part of the areas are slightly aged and severely aged. The server will calculate a comprehensive aging degree score according to the area weights of each area. Then, according to the pre-set aging degree classification criteria, this comprehensive score will be mapped to the corresponding classification result. For example, if the comprehensive score falls within the range of moderate aging, the server will determine that the target aging degree classification result of this blue and white porcelain vase is moderate aging. After the server determines the target aging degree classification result of the target porcelain, it will obtain the trace element data detected by X-ray fluorescence spectroscopy. Suppose this blue and white porcelain vase has been detected by X-ray fluorescence spectroscopy, and the detection data will show the types and contents of various trace elements in the porcelain. The server will perform spatio-temporal alignment and fusion on the target aging degree classification result and the trace element data. Taking the blue and white porcelain vase as an example, the server will analyze the changes in the contents of trace elements corresponding to different aging degree areas. For example, if it is found that the contents of certain trace elements in the moderately aged area are different from those in the slightly aged area, the server will further study the relationship between this change and the age and preservation status of the porcelain. Through the comprehensive analysis of the target aging degree classification result and the trace element data, the server will generate an identification report on the aging degree of the porcelain glaze, including the age inference result and the evaluation of the glaze preservation status. For this Ming Dynasty blue and white porcelain vase, the report will infer the approximate age range of the porcelain and evaluate the preservation status of its glaze, such as the integrity of the glaze, whether there are obvious damage or repair marks, etc. At the same time, the report will also give some suggestions on the protection and collection of the porcelain to help the collector better preserve and manage this porcelain.

[0022] In the embodiment of the present invention, the target glaze aging degree identification model is trained in the following manner and can be implemented through the following examples.

[0023] Obtain an initial glaze aging degree identification model, where the initial glaze aging degree identification model includes a first glaze recognition model and at least one second glaze recognition model. Each first image instance of the first glaze recognition model includes a porcelain image of the overall image of the target porcelain, and each second image instance of the second glaze recognition model includes a porcelain image of the partial image of the target porcelain;

[0024] Obtain the first aging feature mapping parameters of the first glaze recognition model and the second aging feature mapping parameters of each second glaze recognition model when the training in the current training stage is completed;

[0025] For each second glaze recognition model, determine the first directional parameter group for fusion in the first aging feature mapping parameters and the second directional parameter group for fusion in the second aging feature mapping parameters according to the directional consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed; wherein, the first parameter update direction is the model parameter update direction of the first aging feature mapping parameters of the first feature mapping component, and the second parameter update direction is the model parameter update direction of the second aging feature mapping parameters of the second feature mapping component;

[0026] For each second glaze recognition model, fuse the first directional parameter group when the current training stage is completed with the second directional parameter group of this second glaze recognition model to obtain the second aging feature mapping parameters corresponding to the next training stage training, so as to perform one training iteration on this second glaze recognition model based on the obtained second aging feature mapping parameters and each second image instance of this second glaze recognition model, and obtain the second aging feature mapping parameters when the next training stage is completed;

[0027] Fuse the second directional parameter group when each second glaze recognition model is completed in the next training stage and the first directional parameter group when the first glaze recognition model is completed in the current training stage to obtain the first aging feature mapping parameters corresponding to the next training stage training, so as to perform one training iteration on the first glaze recognition model based on the obtained first aging feature mapping parameters and each first image instance, and obtain the first aging feature mapping parameters when the next training stage is completed;

[0028] Among them, determine the first glaze recognition model when the last training stage is completed as the first glaze aging degree identification model, and determine the second glaze recognition model when the last training stage is completed as the second glaze aging degree identification model;

[0029] Among the trained first glaze aging degree identification model or the second glaze recognition model, determine the glaze aging degree identification model including the feature extraction components corresponding to each image type in the preset image type as the target glaze aging degree identification model.

[0030] In an embodiment of the present invention, exemplarily, the server starts to train the target glaze aging degree identification model. First, it obtains the initial glaze aging degree identification model. This initial model includes a first glaze recognition model and multiple second glaze recognition models. The server extracts each first image instance of the first glaze recognition model from the database, and these instances are all the overall images of the target porcelain. For example, there is a batch of porcelains from different dynasties and different kilns, such as a Song Dynasty Ru kiln porcelain plate, a Yuan Dynasty blue and white porcelain vase, a Qing Dynasty famille rose porcelain bowl, etc. The server collects their complete appearance images as the first image instances. For the Song Dynasty Ru kiln porcelain plate, the overall image clearly shows information such as the shape, size, overall glaze color, and luster of the porcelain plate. At the same time, the server obtains each second image instance of the second glaze recognition model, which are the partial images of the target porcelain. Still taking the Song Dynasty Ru kiln porcelain plate as an example, the server will select areas with crackles on the edge of the porcelain plate, areas with inscriptions on the bottom of the plate, areas with tiny air bubbles on the glaze, etc. to take partial close-up images as the second image instances. These partial images can more precisely present the microscopic features of the porcelain glaze. After completing the training in the current training stage, the server obtains the first aging feature mapping parameters of the first glaze recognition model and the second aging feature mapping parameters of each second glaze recognition model. The first aging feature mapping parameters are the parameters learned by the first feature mapping component during the process of processing the first image instances, and they reflect the mapping rules of the first glaze recognition model for the aging features of the porcelain in the overall image. Taking the overall image of a Yuan Dynasty blue and white porcelain vase as an example, the first aging feature mapping parameters can reflect the mapping relationships for aging features such as the overall color change of the vase body and the degree of blurring of the patterns. The second aging feature mapping parameters of each second glaze recognition model are obtained by the second feature mapping component learning for the second image instances. For example, for the partial image of a Qing Dynasty famille rose porcelain bowl, the second aging feature mapping parameters can reflect the mapping situations of local aging features such as the area where the famille rose has peeled off on the bowl wall and the degree of color fading. For each second glaze recognition model, the server determines the first directional parameter group and the second directional parameter group for fusion according to the direction consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed. The first parameter update direction is the model parameter update direction of the first aging feature mapping parameters of the first feature mapping component, and the second parameter update direction is the model parameter update direction of the second aging feature mapping parameters of the second feature mapping component. The server judges the degree of consistency by comparing the consistency of these two directions, such as calculating the cosine value of the angle between them. Suppose the server has three second glaze recognition models, which respectively process the partial images of different parts of a Song Dynasty Ru kiln porcelain plate. When comparing the first parameter update direction and the second parameter update direction, the server finds that when a certain second glaze recognition model processes the partial image of the crackle area of the porcelain plate, the consistency between its second parameter update direction and the first parameter update direction is relatively low.The server further analyzes each texture analysis component of the second feature mapping component of the second glaze recognition model, determines those texture analysis components with relatively low consistency as the second texture analysis components for fusion, and takes the parameters of these components as the second orientation parameter group. Then, based on these second texture analysis components, the server finds the components other than the corresponding texture analysis components in the first feature mapping component, and takes the parameters of these components as the first orientation parameter group. For each second glaze recognition model, the server fuses the first orientation parameter group when the current training stage is completed with the second orientation parameter group of the second glaze recognition model. The server first obtains the first feature contribution degree corresponding to the first glaze recognition model and the second feature contribution degrees corresponding to each second glaze recognition model. These contribution degrees are preset or calculated based on the performance and feature extraction ability of the model in the previous training stage. For the target second orientation parameter group of the texture analysis components corresponding to the same feature level identifier in the second orientation parameter group of each second feature mapping component, the server obtains the model parameter update direction when the current training stage is completed. For example, for a second feature mapping component that processes a local image of a Qing Dynasty famille rose porcelain bowl, a certain texture analysis component of it corresponds to a specific feature level identifier, and the server obtains the model parameter update direction of this component when the current training stage is completed. Based on these update directions, the server calculates the third feature contribution degree of each second feature mapping component. This contribution degree has a negative correlation with the parameter update direction of the corresponding second aging feature mapping parameter. That is to say, for a component with a relatively large change in the parameter update direction, its third feature contribution degree is relatively low. Then, the server performs a linear superposition process on the target second orientation parameter group of the texture analysis components corresponding to the same feature level identifier with the third feature contribution degree of the corresponding second feature mapping component to obtain the first integration result corresponding to the texture analysis components with the same feature level identifier in the next stage. The server determines the mean value of the first integration results of the texture analysis components of each feature level identifier as the third orientation parameter group, and then, based on the first feature contribution degree and the second feature contribution degree, performs weighted fusion on the first orientation parameter group and the third orientation parameter group when the first glaze recognition model completes training in the current training stage to obtain the second aging feature mapping parameters corresponding to the next training stage. The server performs one training iteration on the second glaze recognition model based on the obtained second aging feature mapping parameters and each second image instance of the second glaze recognition model. For example, for a certain second glaze recognition model that processes a local image of a Song Dynasty Ru kiln porcelain plate, the server uses the new second aging feature mapping parameters to train it, inputs the corresponding local image instance, adjusts the parameters of the model, and finally obtains the second aging feature mapping parameters when the next training stage is completed. The server fuses the second orientation parameter group when each second glaze recognition model completes training in the next training stage and the first orientation parameter group when the first glaze recognition model completes training in the current training stage.The specific fusion process is similar to that of obtaining the second aging feature mapping parameters corresponding to the next training stage through the above-mentioned fusion, and factors such as feature contribution degree also need to be considered. After the server obtains the first aging feature mapping parameters corresponding to the next training stage, based on these parameters and each first image instance, a training iteration is performed on the first glaze recognition model. For example, the server processes the overall image of a Yuan Dynasty blue and white porcelain vase using the new first aging feature mapping parameters, continuously adjusts the model parameters, enabling the model to better learn the aging features in the overall image, and finally obtains the first aging feature mapping parameters when the next training stage is completed. After iterative training through multiple training stages, the server determines the first glaze recognition model at the end of the last training stage as the first glaze aging degree identification model, and determines the second glaze recognition model at the end of the last training stage as the second glaze aging degree identification model. The server will select, according to the preset image type, that is, the overall image or the local image, the glaze aging degree identification model including the feature extraction components corresponding to each image type in the preset image type from the trained first glaze aging degree identification model or the second glaze recognition model as the target glaze aging degree identification model. If the porcelain image to be identified is an overall image, the server will select the first glaze aging degree identification model; if it is a local image, the corresponding second glaze aging degree identification model will be selected. In this way, the trained target glaze aging degree identification model can be used for subsequent intelligent identification of the aging degree of porcelain glazes.

[0031] In the embodiment of the present invention, the component architectures of the feature mapping components of each glaze recognition model in the initial glaze aging degree identification model are the same, and the feature mapping components of each glaze recognition model include a first number of texture analysis components;

[0032] For each second glaze recognition model, determining the first directional parameter group for fusion in the first aging feature mapping parameters and the second directional parameter group for fusion in the second aging feature mapping parameters of this second glaze recognition model according to the direction consistency between the first parameter update direction and the second parameter update direction at the end of each training stage in the latest target training stage can be implemented through the following examples.

[0033] According to the direction consistency between the first parameter update direction and the second parameter update direction at the end of each training stage in the latest target training stage, determine the second texture analysis components for fusion in each texture analysis component of the second feature mapping component of this second glaze recognition model, and use the parameters of each second texture analysis component as the second directional parameter group for fusion in the second aging feature mapping parameters of this second glaze recognition model;

[0034] According to each of the second texture analysis components, determine the first texture analysis components for fusion among the texture analysis components of the first feature mapping component, and use the parameters of each first texture analysis component as the first orientation parameter group for fusion in the first aging feature mapping parameters, where the first texture analysis component is a texture analysis component other than the texture analysis components corresponding to the respective second texture analysis components among the texture analysis components of the first feature mapping component.

[0035] In an embodiment of the present invention, exemplarily, in the initial glaze aging degree identification model obtained by the server, the feature mapping component architectures of each glaze identification model are the same, and each includes a first number of texture analysis components. During the training process, the server needs to determine a first orientation parameter group and a second orientation parameter group for fusion for each second glaze identification model. When each training stage in the latest target training stage is completed, the server compares the consistency between the first parameter update direction and the second parameter update direction. Taking a Yue kiln celadon pot of the Tang Dynasty as an example, the first glaze identification model processes its overall image, and the second glaze identification model processes the image of the locally cracked area of the pot body. The server calculates the parameter update directions of each texture analysis component in the first feature mapping component and the second feature mapping component at the end of each training stage. For the second glaze identification model that processes the image of the locally cracked area of the pot body, the server filters the second texture analysis components for fusion according to the direction consistency. For example, the server finds that for a certain texture analysis component in multiple training stages, the included angle between its parameter update direction and the parameter update direction of the corresponding texture analysis component in the first feature mapping component is large, and the consistency is low. The server then determines this texture analysis component as the second texture analysis component for fusion. The server makes such judgments on all the texture analysis components of the second glaze identification model, selects the qualified components, and takes the parameters of these second texture analysis components as the second orientation parameter group for fusion in the second aging feature mapping parameters. Then, based on the determined second texture analysis components, the server finds the first texture analysis components for fusion in the first feature mapping component. Still taking the Yue kiln celadon pot of the Tang Dynasty as an example, the server will check each texture analysis component of the first feature mapping component, exclude those components corresponding to the second texture analysis components, and the remaining components are the first texture analysis components. For example, if the second texture analysis component corresponds to analyzing the texture features of the cracks on the pot body, then the texture analysis components in the first feature mapping component that analyze other features such as the overall color and overall shape may be determined as the first texture analysis components. The server takes the parameters of these first texture analysis components as the first orientation parameter group for fusion in the first aging feature mapping parameters. In this way, the server accurately determines the first orientation parameter group and the second orientation parameter group of each second glaze identification model, provides a basis for subsequent parameter fusion and model training iteration, helps improve the model's ability to identify the aging degree of porcelain glazes, and thus more accurately completes the identification work of the aging degree of porcelain glazes.

[0036] In an embodiment of the present invention, the fusion of the second orientation parameter group when each second glaze identification model completes training in the next training stage and the first orientation parameter group when the first glaze identification model completes training in the current training stage to obtain the first aging feature mapping parameters corresponding to the training in the next training stage can be implemented through the following example.

[0037] Obtain the first feature contribution degree corresponding to the first glaze recognition model and the second feature contribution degrees corresponding to the respective second glaze recognition models;

[0038] Integrate the second orientation parameter groups of the respective second glaze recognition models when the training in the next training stage is completed to obtain a third orientation parameter group;

[0039] Based on the first feature contribution degree and the second feature contribution degrees, perform weighted fusion on the first orientation parameter group of the first glaze recognition model when the training in the current training stage is completed and the third orientation parameter group to obtain the first aging feature mapping parameter corresponding to the training in the next training stage.

[0040] In an embodiment of the present invention, exemplarily, when the server trains each second glaze recognition model and the first glaze recognition model, it is necessary to fuse the second orientation parameter group of each second glaze recognition model when the training in the next training stage is completed with the first orientation parameter group of the first glaze recognition model when the training in the current training stage is completed, so as to obtain the first aging feature mapping parameter corresponding to the training in the next training stage. First, the server obtains the first feature contribution degree corresponding to the first glaze recognition model and the second feature contribution degrees corresponding to each second glaze recognition model. Taking the identification of a batch of ancient porcelain as an example, the first glaze recognition model processes the overall image of the porcelain. It can grasp the overall aging characteristics of the porcelain from a macroscopic perspective, such as the overall color and luster, shape changes, etc. The server determines a relatively high first feature contribution degree, such as 0.6, according to its performance in the previous training and the ability to extract overall aging characteristics. Each second glaze recognition model processes the local image of the porcelain and can deeply analyze the local aging characteristics, such as crack details, bubble distribution, etc. The server determines different second feature contribution degrees for each second glaze recognition model according to factors such as the extraction accuracy of different local features. For example, the second feature contribution degrees of three second glaze recognition models are 0.1, 0.15, and 0.15 respectively. Next, the server integrates the second orientation parameter groups of each second glaze recognition model when the training in the next training stage is completed to obtain a third orientation parameter group. Suppose three second glaze recognition models process the local images of the bottle mouth, bottle body, and bottle bottom of the porcelain respectively. After the training in the next training stage is completed, the server integrates the corresponding parameters in the second orientation parameter group of each model. For example, for the parameters of the texture analysis component, the server will calculate the average value or weighted average value of the parameters of this component of each model, and finally obtain the third orientation parameter group. Finally, the server performs weighted fusion on the first orientation parameter group of the first glaze recognition model when the training in the current training stage is completed and the third orientation parameter group based on the first feature contribution degree and the second feature contribution degrees. The server multiplies the first orientation parameter group by the first feature contribution degree 0.6, multiplies the third orientation parameter group by the sum of the three second feature contribution degrees 0.4, and then adds these two results to obtain the first aging feature mapping parameter corresponding to the training in the next training stage. Through such a fusion method, the server synthesizes the grasp of the overall features by the first glaze recognition model and the analysis of the local features by each second glaze recognition model, so that the first aging feature mapping parameter in the next stage can more comprehensively and accurately reflect the aging characteristics of the porcelain, thereby improving the accuracy of the model in identifying the aging degree of the porcelain glaze.

[0041] In an embodiment of the present invention, the integration of the second orientation parameter groups of each second glaze recognition model when the training in the next training stage is completed to obtain a third orientation parameter group can be implemented through the following example.

[0042] For the target second orientation parameter groups of the texture analysis components corresponding to the same feature level identifier in the second orientation parameter groups of each second feature mapping component, integrate the target second orientation parameter groups of each of the second glaze recognition models when the training is completed in the next training stage, to obtain the first integration result corresponding to the texture analysis component with the same feature level identifier in the next training stage;

[0043] Integrate the first integration results of the texture analysis components for each feature level identifier to obtain the third orientation parameter group.

[0044] In an embodiment of the present invention, exemplarily, when the server integrates the second orientation parameter groups of each second glaze recognition model when they are completed in the next training stage to obtain the third orientation parameter group, it will operate according to specific steps. The server faces multiple second glaze recognition models, and each second feature mapping component of each model has its own second orientation parameter group. Taking the identification of a Song Dynasty official kiln porcelain bowl as an example, assume that there are three second glaze recognition models that respectively process the local images of the bowl rim, the bowl body, and the bowl bottom. The server focuses on the target second orientation parameter groups of the texture analysis components corresponding to the same feature level identifier in the second orientation parameter groups of each second feature mapping component. For example, for the feature level identifier for analyzing the glaze crack texture, all three second glaze recognition models have corresponding texture analysis components, and their parameters form the target second orientation parameter group. The server integrates the target second orientation parameter groups of these three models when they are completed in the next training stage. For the target second orientation parameter groups for processing the crack texture of the bowl rim, the bowl body, and the bowl bottom, the server uses the method of calculating the average value. If the corresponding parameter of the rim model is 0.2, the body model is 0.3, and the bottom model is 0.25, then the average value is (0.2 + 0.3 + 0.25) ÷ 3 = 0.25, and this 0.25 is the first integration result corresponding to the texture analysis component of the same feature level identifier (analyzing the glaze crack texture) in the next training stage. The server performs such an operation on the texture analysis components of each feature level identifier. In addition to the glaze crack texture, there are also texture analysis components for analyzing different feature level identifiers such as bubble distribution and color change. The server will calculate their respective first integration results according to the above method respectively. The server integrates the first integration results of the texture analysis components of each feature level identifier to obtain the third orientation parameter group. For the first integration results of analyzing different features, the server integrates them again. For example, the first integration results of analyzing features such as crack texture, bubble distribution, and color change are processed according to certain rules (such as calculating the average value again or weighted average) to finally form the third orientation parameter group. Through such an integration process, the server effectively integrates the information of each second glaze recognition model for different local images and different feature levels, enabling the third orientation parameter group to comprehensively reflect the training results of each second glaze recognition model, laying a foundation for fusing with the first orientation parameter group in the subsequent stage to obtain more accurate first aging feature mapping parameters corresponding to the next training stage, and further improving the performance of the porcelain glaze aging degree identification model.

[0045] In an embodiment of the present invention, the integration of the target second orientation parameter groups of each of the second glaze recognition models when they are completed in the next training stage to obtain the first integration result corresponding to the texture analysis component of the same feature level identifier in the next training stage can be executed through the following example.

[0046] Obtain the model parameter update direction of each second aging feature mapping parameter when the current training stage is completed;

[0047] According to the model parameter update direction of each second aging feature mapping parameter when the current training stage is completed, obtain the third feature contribution degree of each second feature mapping component, and there is a negative correlation between the third feature contribution degree and the parameter update direction of the corresponding second aging feature mapping parameter;

[0048] For the target second orientation parameter group of the texture analysis components corresponding to the same feature level identifier in the second orientation parameter group of each second feature mapping component, perform linear superposition processing with the third feature contribution degree of the corresponding second feature mapping component to obtain the first integration result corresponding to the texture analysis components with the same feature level identifier in the next stage;

[0049] The integration of the first integration results of the texture analysis components of each feature level identifier to obtain the third orientation parameter group can be implemented through the following examples.

[0050] Determine the mean value of the first integration results of the texture analysis components of each feature level identifier as the third orientation parameter group.

[0051] In an embodiment of the present invention, exemplarily, when the server integrates the target second orientation parameter groups of each second glaze recognition model to obtain a first integration result and then obtains a third orientation parameter group, it will operate according to a specific process. The server obtains the model parameter update direction of each second aging feature mapping parameter when the current training stage is completed. Taking the identification of a Ming Dynasty blue and white porcelain vase as an example, there are three second glaze recognition models that respectively process the local images of the bottle mouth, the bottle body, and the bottle bottom. After the current training stage ends, the server analyzes the change situation of the second aging feature mapping parameters of each model to determine its parameter update direction. For example, for the model processing the local image of the bottle mouth, its second aging feature mapping parameter changes from value A to value B in the current stage. The server determines the update direction of this parameter of the model by calculating the difference between the two values. According to the model parameter update direction of each second aging feature mapping parameter when the current training stage is completed, the server obtains the third feature contribution degree of each second feature mapping component. Since there is a negative correlation between the third feature contribution degree and the parameter update direction of the corresponding second aging feature mapping parameter, if the parameter update direction of a certain second aging feature mapping parameter changes greatly, it indicates that the learning of this model fluctuates greatly and has poor stability in this stage, and its third feature contribution degree is relatively low. For example, if the parameter update direction of the model processing the local image of the bottle body changes violently, the server assigns it a lower third feature contribution degree, such as 0.2; while the parameter update of the model processing the local image of the bottle bottom is relatively stable, the server assigns it a higher third feature contribution degree, such as 0.8. For the target second orientation parameter groups of the texture analysis components corresponding to the same feature level identifier in the second orientation parameter group of each second feature mapping component, the server performs a linear superposition process with the third feature contribution degree of the corresponding second feature mapping component. For example, for the texture analysis component analyzing the change of the local glaze color of the blue and white porcelain vase at this feature level identifier, the target second orientation parameter groups of the models of the bottle mouth, the bottle body, and the bottle bottom are 0.3, 0.4, and 0.5 respectively, and the corresponding third feature contribution degrees are 0.3, 0.2, and 0.8. The server performs a linear superposition: 0.3×0.3 + 0.4×0.2 + 0.5×0.8 = 0.53. This 0.53 is the first integration result corresponding to the texture analysis component of this same feature level identifier in the next stage. The server integrates the first integration results of the texture analysis components of each feature level identifier to obtain a third orientation parameter group. The server calculates the average value of the first integration results of the texture analysis components analyzing different feature level identifiers (such as glaze color change, crack density, etc.). Assuming that the first integration results of analyzing three different feature level identifiers are 0.53, 0.6, and 0.55 respectively, the server calculates (0.53 + 0.6 + 0.55)÷3 = 0.56 and determines 0.56 as the third orientation parameter group. Through such operations, the server can comprehensively integrate the information of each second glaze recognition model, provide more accurate data for subsequent fusion with the first orientation parameter group, and improve the performance of the porcelain glaze aging degree identification model.

[0052] In an embodiment of the present invention, for each second glaze recognition model, according to the direction consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed, the first directional parameter group for fusion in the first aging feature mapping parameter and the second directional parameter group for fusion in the second aging feature mapping parameter are determined, and the implementation can be performed through the following examples.

[0053] If the direction consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed does not exceed a preset consistency threshold, the feature selection matrix when this training stage of the second glaze recognition model is completed is updated to obtain the feature selection matrix corresponding to the next training stage; wherein, the feature selection matrix is used to indicate the second directional parameter group for fusion in the second aging feature mapping parameter of the second glaze recognition model, and the feature selection matrix corresponding to the first training stage is a preset feature selection matrix;

[0054] According to the feature selection matrix corresponding to the next training stage, the first directional parameter group for fusion in the first aging feature mapping parameter and the second directional parameter group for fusion in the second aging feature mapping parameter of the second glaze recognition model are determined, wherein the first directional parameter group is the array in the first aging feature mapping parameter except for the parameters corresponding to the second directional parameter group.

[0055] In an embodiment of the present invention, exemplarily, when the server processes each second glaze recognition model, it determines the orientation parameter group for fusion based on the direction consistency between the first parameter update direction and the second parameter update direction. Taking the identification of a Qing Dynasty famille rose porcelain vase as an example, the vase has multiple second glaze recognition models that respectively process local images of different parts such as the bottleneck, the body, and the bottom of the vase. After each training stage in the latest target training stage is completed, the server calculates the direction consistency between the first parameter update direction and the second parameter update direction of each second glaze recognition model. Suppose the preset consistency threshold set by the server is 0.8. For the second glaze recognition model that processes the local image of the bottleneck, when the server checks the direction consistency data at the completion of each training stage in the latest target training stage, it is found that the direction consistency of all stages is lower than 0.8. At this time, the server updates the feature selection matrix at the completion of this training stage of the second glaze recognition model. The feature selection matrix is like a signpost that can clarify which parameters in the second aging feature mapping parameters of the second glaze recognition model are the second orientation parameter groups for fusion. The feature selection matrix trained in the first training stage is preset, for example, initially setting that the parameters corresponding to some elements in the matrix are the parameters for fusion. According to the situation that the direction consistency does not exceed the threshold, the server adjusts the elements in the matrix, for example, changing the element values that originally indicate that the parameters of some texture analysis components are for fusion, so as to obtain the feature selection matrix trained in the next training stage. The server determines the orientation parameter group for fusion according to the updated feature selection matrix trained in the next training stage. For the second glaze recognition model that processes the local image of the bottleneck, the server finds the corresponding parameters in the second aging feature mapping parameters as the second orientation parameter group according to the new feature selection matrix. Then, the server excludes the parameters corresponding to the second orientation parameter group from the first aging feature mapping parameters, and the remaining parameter array is the first orientation parameter group. The same operation is also applied to other second glaze recognition models that process local images such as the body and the bottom of the vase. In this way, the server can dynamically adjust the feature selection matrix according to the direction consistency, and then accurately determine the first orientation parameter group and the second orientation parameter group for fusion, making the subsequent parameter fusion and model training more accurate, improving the performance of the porcelain glaze aging degree identification model, and more accurately identifying the glaze aging degree of the Qing Dynasty famille rose porcelain vase.

[0056] In an embodiment of the present invention, the direction consistency between the first parameter update direction and the second parameter update direction includes the consistency coefficient between the first sub-parameter update direction and the second sub-parameter update direction of each corresponding texture analysis component in the first glaze recognition model and the second glaze recognition model;

[0057] If the direction consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed does not exceed a preset consistency threshold, update the feature selection matrix when the current training stage of the second glaze recognition model is completed to obtain the feature selection matrix corresponding to the next training stage, which can be implemented through the following example.

[0058] For each texture analysis component in each second glaze recognition model, if the consistency coefficient of the texture analysis component when each training stage in the latest target training stage is completed does not exceed the preset consistency threshold, determine the texture analysis component as the target texture analysis component;

[0059] In the feature selection matrix when the current training stage of the second glaze recognition model is completed, release the parameters corresponding to the target texture analysis component for fusion to obtain the feature selection matrix corresponding to the next training stage.

[0060] In an embodiment of the present invention, exemplarily, when the server processes the update of the feature selection matrix of the second glaze recognition model, it will operate according to the consistency coefficient between the first and second parameter update directions. Taking the identification of a Song Dynasty Ru kiln porcelain plate as an example, there are multiple second glaze recognition models that respectively process different local images such as the edge, center, and pattern of the porcelain plate, and at the same time, there is a first glaze recognition model that processes the overall image of the porcelain plate. After each training stage in the latest target training stage is completed, the server will calculate the consistency coefficient between the first sub-parameter update direction and the second sub-parameter update direction of each corresponding texture analysis component in the first glaze recognition model and the second glaze recognition model. For example, for the second glaze recognition model that processes the edge image of the porcelain plate, one of the texture analysis components is responsible for analyzing the crackle characteristics of the edge glaze. The server will compare its second sub-parameter update direction with the first sub-parameter update direction of the texture analysis component in the first glaze recognition model that corresponds to analyzing the overall crackle characteristics, and obtain the consistency coefficient. Assume that the preset consistency threshold is set to 0.7. For each texture analysis component in each second glaze recognition model, the server will check the consistency coefficient when each training stage in the latest target training stage is completed. If in the second glaze recognition model that processes the center image of the porcelain plate, there is a texture analysis component responsible for analyzing the bubble distribution characteristics of the center glaze, and its consistency coefficient at the end of each training stage is lower than 0.7, the server will determine this texture analysis component as the target texture analysis component. The server will perform operations on the feature selection matrix when the current training stage of the second glaze recognition model is completed. The feature selection matrix originally indicates which parameters in the second aging feature mapping parameters of the second glaze recognition model are used for fusion. The server removes the parameters corresponding to the target texture analysis component from the specified ones for fusion. For example, originally, a row element in the feature selection matrix corresponds to the parameters of the texture analysis component that processes the bubble distribution characteristics of the center of the porcelain plate, and the element value indicates that this parameter is used for fusion. The server modifies this element value so that it no longer indicates that this parameter is used for fusion, thereby obtaining the feature selection matrix corresponding to the training in the next training stage. Similarly, the server will perform the same operation on the second glaze recognition models that process other local images of the porcelain plate. In this way, the server dynamically adjusts the feature selection matrix according to the consistency coefficient, accurately determines the parameters for fusion, provides a more accurate basis for subsequent parameter fusion and model training, improves the performance of the porcelain glaze aging degree identification model, and more accurately identifies the glaze aging condition of the Song Dynasty Ru kiln porcelain plate.

[0061] In an embodiment of the present invention, the model parameter update direction of the second aging feature mapping parameter when the current training stage is completed can be implemented through the following examples.

[0062] Fuse the first set of orientation parameters of the first aging feature mapping parameter at the end of the previous training phase with the second set of orientation parameters of the second feature mapping component at the end of the current training phase to obtain a second fusion result corresponding to the current training phase;

[0063] Obtain the model parameter update direction of the second aging feature mapping parameter at the end of the current training phase according to the deviation between the second aging feature mapping parameter at the end of the current training phase and the second fusion result;

[0064] The model parameter update direction of the first aging feature mapping parameter at the end of the current training phase includes:

[0065] Obtain the model parameter update direction of the first aging feature mapping parameter at the end of the current training phase according to the deviation between the first aging feature mapping parameter at the end of the current training phase and the end time of the previous training phase.

[0066] In an embodiment of the present invention, exemplarily, the server has multiple second glaze recognition models to process local images of the mouth, body, bottom, etc. of the blue and white porcelain pot. For the second glaze recognition model that processes the local image of the pot body, the server first obtains the first orientation parameter group of the first aging feature mapping parameter when the previous training stage is completed, and the second orientation parameter group of the second feature mapping component when the current training stage is completed. For example, the first orientation parameter group includes parameters for analyzing the overall color change, and the second orientation parameter group includes parameters for analyzing the local crackle characteristics of the pot body. The server fuses these two groups of parameters according to a certain rule, for example, by weighted average, to obtain the second fusion result corresponding to the current training stage. The server calculates the deviation between the second aging feature mapping parameter when the current training stage is completed and the second fusion result. If the current second aging feature mapping parameter indicates that the density of the local crackle of the pot body is a certain value, while the crackle density corresponding to the second fusion result is another value, the server obtains the deviation value through calculation methods such as subtracting the two. According to this deviation, the server determines the model parameter update direction of the second aging feature mapping parameter when the current training stage is completed. If the deviation is positive, it means that the current parameter value is larger, and the update direction may be to decrease the parameter value; if the deviation is negative, the update direction may be to increase the parameter value. For the overall image of the blue and white porcelain pot processed by the first glaze recognition model, the server obtains the values of the first aging feature mapping parameter when the current training stage is completed and when the previous training stage is completed. For example, the parameter for analyzing the overall color uniformity in the previous stage is a certain value, and this parameter becomes another value in the current stage. The server calculates the deviation between these two values to obtain the deviation amount by simple numerical subtraction. According to this deviation, the server determines the model parameter update direction of the first aging feature mapping parameter when the current training stage is completed. If the current value is larger than the value in the previous stage, the update direction may be to decrease the parameter; otherwise, increase the parameter. The server also performs the same operation on the second glaze recognition models that process other local images of the blue and white porcelain pot, so as to accurately grasp the update direction of each model parameter, provide a basis for subsequent model training iteration, thereby improving the accuracy and reliability of the porcelain glaze aging degree identification model, and more accurately identifying the glaze aging degree of the Yuan Dynasty blue and white porcelain pot.

[0067] In an embodiment of the present invention, each image instance further includes a glaze aging degree target value of the aging degree grading result of each glaze area in the corresponding porcelain image; the aging degree grading result is determined according to at least two indicators among the crackle density of the glaze, the color change rate, and the microscopic bubble distribution characteristics;

[0068] For each second glaze recognition model, each training stage further includes the following implementation methods.

[0069] For the second feature extraction component of the second glaze recognition model, the parameter of the first feature extraction component of the first feature extraction component of the same image type when the previous training stage is completed is determined as the parameter of the second feature extraction component updated in this training stage;

[0070] The second glaze recognition model after updating the parameters of each second feature extraction component and the second aging feature mapping parameter is determined as the second glaze recognition model in the process of this training stage, and the second glaze recognition model in the process of this training stage is trained based on the image instances according to the second glaze recognition model until it meets the second training termination state, and the second glaze recognition model when this training stage is completed is obtained;

[0071] For the first glaze recognition model, the following implementation manners are also adopted in each training stage.

[0072] For each first feature extraction component, the parameters of the second feature extraction components of the same image type when this training stage is completed are fused to obtain the updated parameter of the first feature extraction component of the same image type of the first feature extraction component;

[0073] The first glaze recognition model after updating the parameters of each first feature extraction component and the first aging feature mapping parameter is determined as the first glaze recognition model in the process of this training stage, and the first glaze recognition model in the process of this training stage is trained based on the image instances according to the first glaze recognition model until it meets the first training termination state, and the first glaze recognition model when this training stage is completed is obtained.

[0074] In an embodiment of the present invention, exemplarily, when the server trains the porcelain glaze aging degree identification model, it iteratively trains the first glaze recognition model and the second glaze recognition model according to the target value of the glaze aging degree in the image instance. Taking the identification of a Ming Dynasty blue and white porcelain vase as an example, the image instance covers the overall image of the vase and local images such as the bottle mouth, bottle body, and bottle bottom, and each image instance contains the target value of the glaze aging degree grading result of each glaze area, and these target values are determined according to at least two of the indexes of glaze crack density, color change rate, and microscopic bubble distribution characteristics. For the second glaze recognition model that processes the local image of the bottle mouth, when its second feature extraction component is updated in this training stage, the server directly determines the parameter of the first feature extraction component when the first feature extraction component of the same image type (local image) was completed in the previous training stage as the parameter of the second feature extraction component after the update of this second feature extraction component. For example, the parameter when the first feature extraction component analyzes the overall bottle mouth shape feature will be applied to the second feature extraction component that processes the local image of the bottle mouth. After the server updates each parameter of the second feature extraction component and the second aging feature mapping parameter of the second glaze recognition model, it determines it as the second glaze recognition model in the process of this training stage. Then, the server trains the model at this stage according to the image instance of this model, that is, the local image of the bottle mouth, until the second training termination state is satisfied, such as the prediction accuracy rate of the model reaches a certain threshold or the number of training times reaches a preset value, so as to obtain the second glaze recognition model when this training stage is completed. The same process will also be applied to other second glaze recognition models that process local images such as the bottle body and bottle bottom. For the overall image of the blue and white porcelain vase processed by the first glaze recognition model, the server updates each first feature extraction component in each training stage. The server fuses the parameters of all second feature extraction components of the same image type (overall image) when this training stage is completed. For example, operations such as weighted averaging are performed on the parameters for color feature extraction of the second feature extraction components that process the local images of the bottle mouth, bottle body, and bottle bottom at this stage to obtain the updated parameter of the first feature extraction component for color feature extraction. After the server updates each parameter of the first feature extraction component and the first aging feature mapping parameter of the first glaze recognition model, it determines it as the first glaze recognition model in the process of this training stage. Then, the server trains the model at this stage according to the image instance of this model, that is, the overall image of the blue and white porcelain vase, until the first training termination state is satisfied, such as reaching a preset loss function value or the upper limit of the number of training rounds, and finally obtains the first glaze recognition model when this training stage is completed. Through such an interactive training method, the server continuously optimizes the first glaze recognition model and the second glaze recognition model, and improves the accuracy of the model in identifying the aging degree of porcelain glaze.

[0075] In the embodiment of the present invention, the first feature mapping component and each second feature mapping component have the same feature mapping component structure, both including at least one aging texture analysis unit, and each second feature extraction component of the second glaze recognition model includes a feature extraction unit corresponding one-to-one to the aging texture analysis unit of the second feature mapping component in the second glaze recognition model;

[0076] Training the second glaze recognition model in the current training phase to meet the second training termination state based on the image instances of the second glaze recognition model can be implemented through the following examples.

[0077] For each aging texture analysis unit of the second glaze recognition model in the current training phase, integrate the feature information output by the corresponding feature extraction unit of each second feature extraction component of the second glaze recognition model to obtain the first integrated feature corresponding to the aging texture analysis unit;

[0078] For each aging texture analysis unit of the second glaze recognition model in the current training phase, according to each reference feature identifier corresponding to the aging texture analysis unit when the previous training phase is completed, optimize the first integrated feature corresponding to the aging texture analysis unit by gated fusion, and integrate the first integrated features before and after optimization of the aging texture analysis unit to obtain the second integrated feature corresponding to the aging texture analysis unit;

[0079] Determine the mapping result of the first aging texture analysis unit as the mapping result of the first aging texture analysis unit. For non-first aging texture analysis units, perform feature mapping based on the second integrated feature corresponding to the non-first aging texture analysis unit and the mapping result of the previous aging texture analysis unit to obtain the mapping result of the non-first aging texture analysis unit;

[0080] According to the mapping result of the last aging texture analysis unit, obtain the predicted aging degree grading result of each glaze area in the corresponding porcelain image of the image instance. According to the deviation between the predicted aging degree grading result of each glaze area and the target value of the glaze aging degree, optimize the parameters of each second feature extraction component and the second aging feature mapping parameters of the second glaze recognition model in the current training phase;

[0081] Among them, each reference feature identifier corresponding to each aging texture analysis unit is obtained by performing feature grouping on the mapping features output by the corresponding aging texture analysis unit for the glaze areas with the same aging degree grading result in each first image instance.

[0082] In an embodiment of the present invention, exemplarily, when the server trains the second glaze recognition model, it needs to train the model in this training phase to meet the second training termination state based on image instances. Taking the identification of a Tang Dynasty secret-color porcelain bowl as an example, the first feature mapping component and the second feature mapping component have the same structure, both having an aging texture analysis unit, and the second feature extraction component has a corresponding feature extraction unit. The server starts to integrate feature information for each aging texture analysis unit of the second glaze recognition model in the current training phase. For example, for the second glaze recognition model that processes the local image of the bowl mouth, there are three aging texture analysis units, which respectively analyze the crack texture, color and luster texture, and bubble distribution texture. There are corresponding feature extraction units in its second feature extraction component, which respectively extract the feature information of the crack, color and luster, and bubble distribution of the bowl mouth. The server integrates the feature information output by each feature extraction unit corresponding to the same aging texture analysis unit. For the aging texture analysis unit that analyzes the crack texture, the server integrates the information output by the feature extraction units that capture the crack features of the bowl mouth from different angles to obtain the first integrated feature corresponding to this aging texture analysis unit. The server optimizes the first integrated feature of each aging texture analysis unit. Each aging texture analysis unit has a corresponding reference feature identifier at the end of the previous training phase. These identifiers are obtained by feature grouping of the mapped features output by the glaze areas with the same aging degree classification result in the first image instance for the corresponding aging texture analysis unit. For example, for the aging texture analysis unit that analyzes the color and luster texture, the server optimizes the first integrated feature of this unit using gated fusion according to the reference feature identifier obtained in the previous phase. Gated fusion is like an intelligent switch that selectively enhances or weakens some parts of the first integrated feature according to the reference feature identifier. After optimization, the server integrates the first integrated feature before and after optimization again to obtain the second integrated feature corresponding to this aging texture analysis unit. The server determines the mapping result of the first aging texture analysis unit. Taking the aging texture analysis unit that analyzes the crack texture as an example, the second integrated feature corresponding to it is directly determined as the mapping result of this unit. For non-first aging texture analysis units, such as the unit that analyzes the color and luster texture, the server performs feature mapping based on its corresponding second integrated feature and the mapping result of the previous aging texture analysis unit (the crack texture analysis unit). Through a specific mapping function, the features of the two are associated and transformed to obtain the mapping result of this non-first aging texture analysis unit. By analogy, the feature mapping of all aging texture analysis units is completed. The server obtains the predicted aging degree classification results of each glaze area in the image instance (the local image of the bowl mouth) according to the mapping result of the last aging texture analysis unit. The server compares these predicted results with the preset glaze aging degree target values in the image instance and calculates the deviation.For example, if a certain area of the bowl mouth is predicted to be mildly aged while the target value indicates moderate aging, the server optimizes the parameters of each second feature extraction component and the second aging feature mapping parameter of the second glaze recognition model during the current training phase based on this deviation. By adjusting the parameters, the model can more accurately predict the aging degree in subsequent training. The server continuously repeats the above process to train the second glaze recognition model for processing other local images such as the bowl body and the bowl bottom until the second training termination state is met, such as the deviation between the prediction result and the target value being within an acceptable range or reaching the preset number of training rounds, thereby improving the identification accuracy of the second glaze recognition model for the local glaze aging degree of the Tang Dynasty secret-color porcelain bowl.

[0083] In the embodiment of the present invention, according to the reference feature identifier at the end of the previous training phase, gated fusion is used to optimize the first integrated feature corresponding to the aging texture analysis unit, and the implementation can be carried out through the following examples.

[0084] According to the reference feature identifier at the end of the previous training phase, the first integrated feature corresponding to the aging texture analysis unit, and the number of texture analysis components in the aging texture analysis unit, gated fusion is used to perform a normalization operation to obtain the fourth feature contribution degree;

[0085] The first integrated feature corresponding to the aging texture analysis unit is weighted by the fourth feature contribution degree to obtain the optimized first integrated feature.

[0086] In an embodiment of the present invention, exemplarily, when the server trains the second glaze recognition model, it will optimize the first integrated feature corresponding to the aging texture analysis unit. Taking the identification of a Song Dynasty official kiln porcelain plate as an example, when the server processes the second glaze recognition model of the local image of the porcelain plate, it performs an optimization operation on the first integrated feature of each aging texture analysis unit. The server starts to optimize a certain aging texture analysis unit of the second glaze recognition model during the current training phase. For example, the aging texture analysis unit that analyzes the local crackle texture of the porcelain plate. This unit had reference feature identifiers at the end of the previous training phase. These identifiers were obtained by grouping the mapped features output from the glaze regions with the same aging degree classification results of the corresponding aging texture analysis unit in the first image instance. At the same time, the server has obtained the first integrated feature corresponding to this aging texture analysis unit and knows the number of texture analysis components in this unit. The server uses gated fusion to perform a normalization operation to obtain the fourth feature contribution degree. Gated fusion is like an intelligent filter that calculates based on the reference feature identifiers, the first integrated feature, and the number of texture analysis components. Suppose the aging texture analysis unit that analyzes the crackle texture has 3 texture analysis components. The reference feature identifiers of the previous phase indicate the correlation degree between a certain crackle pattern and the aging degree. The first integrated feature contains various feature information of the current local crackle of the porcelain plate. The server integrates and normalizes this information through a specific algorithm to obtain the fourth feature contribution degree. This contribution degree reflects the importance of each feature in the current optimization process of this aging texture analysis unit. The server weights the first integrated feature corresponding to this aging texture analysis unit using the obtained fourth feature contribution degree. If the fourth feature contribution degree shows that a certain crackle feature is more important in identifying the aging degree, the server will increase the weight of this feature in the first integrated feature; otherwise, it will decrease the weight. For example, if the fourth feature contribution degree indicates that the density feature of the crackle has a greater impact on the aging degree identification, the server will increase the proportion of this feature in the first integrated feature. Through such a weighting operation, the server obtains the optimized first integrated feature. The server performs the same operation on the aging texture analysis units that process other texture features of the porcelain plate (such as color texture, bubble distribution texture). In this way, the server continuously optimizes the features of each aging texture analysis unit in the second glaze recognition model, enabling the model to more accurately identify the glaze aging degree based on the local image of the porcelain plate, improving the performance and accuracy of the entire identification model.

[0087] In an embodiment of the present invention, the reference feature identifiers of each aging texture analysis unit are updated in the following manner:

[0088] For each first image instance, according to the target value of the glaze aging degree of the first image instance, determine the aging degree classification results corresponding to the mapped features of each glaze region output by the aging texture analysis unit at the end of the previous training phase;

[0089] For each first image instance, map features of each glaze area with the same aging degree classification result output by the aging texture analysis unit are integrated to obtain recognition processing features of the corresponding aging degree classification result of the aging texture analysis unit for the first image instance;

[0090] For the same aging degree classification result, recognition processing features of the aging degree classification result of the aging texture analysis unit for each first image instance are grouped according to a preset number of preset categories to determine features of the cluster center of each feature group in the current training stage;

[0091] For each feature group, features of the cluster center of the feature group in the current training stage and the previous training stage are integrated to obtain a reference feature identifier corresponding to the feature group when the aging texture analysis unit completes the current training stage.

[0092] In an embodiment of the present invention, exemplarily, when the server trains the porcelain glaze aging degree identification model, it will update the reference feature identifiers of each aging texture analysis unit. Taking the identification of a batch of ancient porcelains as an example, the first image instances of these porcelains include the overall image, and each instance has a corresponding target value of the glaze aging degree. For each first image instance, the server analyzes the mapping features of each glaze area output by the aging texture analysis unit at the end of the previous training stage according to its target value of the glaze aging degree. For example, for the first image instance of a Song Dynasty Ru kiln porcelain plate, its target value of the glaze aging degree shows that some areas are slightly aged and some are moderately aged. The server checks the mapping features output by the aging texture analysis unit that analyzed this porcelain plate in the previous stage, determines the aging degree classification results corresponding to the mapping features of each glaze area, and clarifies which mapping features correspond to the slightly aged areas and which correspond to the moderately aged areas. For each first image instance, the server integrates the mapping features of each glaze area with the same aging degree classification result output by the aging texture analysis unit. Still taking the Song Dynasty Ru kiln porcelain plate as an example, the server integrates the mapping features of all glaze areas corresponding to the slightly aged state to obtain the recognition processing features of the aging texture analysis unit for the slightly aged degree classification result of this porcelain plate. Similarly, the same operation is performed on the moderately aged areas to obtain the corresponding recognition processing features. For the same aging degree classification result, the server groups the recognition processing features of the aging texture analysis unit for each first image instance of this aging degree classification result according to the preset number of preset categories. Suppose the preset number of categories is 3. For the recognition processing features corresponding to the slightly aged state in all first image instances, the server uses a clustering algorithm to divide them into 3 groups. Then it determines the features of the cluster center of each group in the current training stage. The cluster center features represent the typical situation of the features of this group. For each feature group, the server integrates the features of the cluster center of this feature group in the current training stage and the previous training stage. For example, if the cluster center features of a certain feature group are A in the current stage and B in the previous stage, the server integrates A and B. Through such an integration operation, the server obtains a reference feature identifier corresponding to this feature group when the aging texture analysis unit completes the current training stage. The server performs such an operation on all feature groups to complete the update of the reference feature identifiers of each aging texture analysis unit. By updating the reference feature identifiers, the server enables the aging texture analysis unit to more accurately optimize the first integrated features based on these identifiers in subsequent training, thereby improving the performance and accuracy of the porcelain glaze aging degree identification model.

[0093] In an embodiment of the present invention, the method is applied to a distributed collaborative training system. The first glaze recognition model is deployed in a global aggregator, and one second glaze recognition model is deployed in one local training node. The method is executed by the global aggregator. Among them, obtaining the second aging feature mapping parameters of each second glaze recognition model when the training in the current training stage is completed can be implemented through the following examples.

[0094] Receive the model update parameters of each second glaze recognition model sent by each local training node, where the model update parameters include the second aging feature mapping parameters of the second glaze recognition model when the training in the current training stage is completed;

[0095] For each second glaze recognition model, after obtaining the second aging feature mapping parameters corresponding to the training in the next training stage, it further includes:

[0096] Transmit the model update parameters corresponding to the training of the second glaze recognition model in the next training stage to the corresponding local training node, so that the local training node performs one training iteration on the second glaze recognition model based on the model update parameters and the corresponding second image instance, and obtains the model update parameters of the second glaze recognition model when the training in the next stage is completed.

[0097] In an embodiment of the present invention, exemplarily, in a distributed collaborative training system, the server, as a global aggregator, executes a training method for a porcelain glaze aging degree identification model. Taking the identification of a batch of ancient porcelain as an example, the first glaze recognition model is deployed in the global aggregator, and multiple second glaze recognition models are respectively deployed in different local training nodes. The global aggregator waits for each local training node to complete the current training phase. Each local training node is responsible for the training of a second glaze recognition model. For example, there are three local training nodes, which process the local images of the bottle mouth, bottle body and bottle bottom of the porcelain respectively. When the training is completed, each local training node will send the model update parameters of its own second glaze recognition model to the global aggregator. These model update parameters include the second aging feature mapping parameters of the second glaze recognition model when the training is completed in the current training phase. For example, the local training node that processes the local image of the bottle mouth will send the mapping parameters for analyzing the aging features such as cracking and color of the bottle mouth glaze. The global aggregator receives these model update parameters from different local training nodes, thereby obtaining the second aging feature mapping parameters of each second glaze recognition model. The global aggregator performs parameter fusion and other operations on each second glaze recognition model to obtain the second aging feature mapping parameters corresponding to the next training stage. For the second glaze recognition model that processes the local image of the bottle mouth, the global aggregator transmits the model update parameters containing the new second aging feature mapping parameters back to the corresponding local training node. After receiving the parameters, the local training node will perform a training iteration on the second glaze recognition model based on this and the corresponding second image instance (local image of the bottle mouth). During the iteration process, the model will relearn and analyze the glaze aging features in the local image of the bottle mouth according to the new parameters and adjust its own parameters. After the training is completed, the local training node obtains the model update parameters of the second glaze recognition model when the training is completed in the next stage. Similarly, the global aggregator will perform the same operation on the second glaze recognition model that processes the local images of the bottle body, bottle bottom, etc., and transmit the updated parameters to the corresponding local training nodes, allowing them to continue training iterations. Through this distributed collaborative training method, the global aggregator and each local training node cooperate with each other to continuously optimize the second glaze recognition model and improve the accuracy of the identification of the aging degree of the porcelain glaze.

[0098] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent identification method for the aging degree of porcelain glaze. Figure 2 As shown, Figure 2It is a block diagram of a computer device 100 provided by an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To achieve data transmission or interaction, the elements of the memory 111, the processor 112, and the communication unit 113 are electrically connected to each other directly or indirectly. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines.

[0099] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. The embodiments were chosen and described in order to best illustrate the principles of the disclosure and its practical application, to thereby enable those skilled in the art to best utilize the disclosure and to utilize various embodiments with various modifications as are suited to the particular application contemplated.

Claims

1. An intelligent identification method for the aging degree of porcelain glaze, characterized in that: include: The reflectance characteristics of the target porcelain in the visible light and near-infrared bands are collected by a multispectral imaging device, and the glaze reflection suppression process is performed on the collected original image to generate a standardized porcelain image to be identified; the porcelain image to be identified is a porcelain image of a preset image type of the target porcelain, and the preset image type is an overall image or a partial image; Loading the porcelain image to be identified into a pre-trained target glaze aging degree identification model, and obtaining a glaze aging degree target value of the aging degree classification result of each glaze area in the porcelain image output by the target glaze aging degree identification model; Determining a target aging degree grading result of the target porcelain according to a target aging degree grading result of each glaze area in the porcelain image; According to the target aging degree classification result and the trace element data detected by X-ray fluorescence spectroscopy, a spatiotemporal alignment fusion is performed to generate a porcelain glaze aging degree identification report including the age inference result and the glaze preservation state assessment; The target glaze aging degree identification model is trained by the following methods, including: Acquire an initial glaze aging degree identification model, wherein the initial glaze aging degree identification model includes a first glaze recognition model and at least one second glaze recognition model, wherein each first image instance of the first glaze recognition model includes a porcelain image of an overall image of a target porcelain, and each second image instance of the second glaze recognition model includes a porcelain image of a partial image of the target porcelain; Acquire a first aging feature mapping parameter of the first glaze recognition model and a second aging feature mapping parameter of each of the second glaze recognition models when training is completed in the current training phase; For each second glaze recognition model, a first orientation parameter group for fusion in the first aging feature mapping parameter and a second orientation parameter group for fusion in the second aging feature mapping parameter are determined according to the directional consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed; wherein the first parameter update direction is the model parameter update direction of the first aging feature mapping parameter of the first feature mapping component, and the second parameter update direction is the model parameter update direction of the second aging feature mapping parameter of the second feature mapping component; For each second glaze recognition model, the first orientation parameter group when the training is completed in the current training stage is merged with the second orientation parameter group of the second glaze recognition model to obtain the second aging feature mapping parameters corresponding to the training in the next training stage, and based on the obtained second aging feature mapping parameters and each second image instance of the second glaze recognition model, the second glaze recognition model is trained once to obtain the second aging feature mapping parameters when the training is completed in the next training stage; The second orientation parameter group of each second glaze recognition model when training is completed in the next training stage and the first orientation parameter group of the first glaze recognition model when training is completed in the current training stage are merged to obtain the first aging feature mapping parameters corresponding to the training in the next training stage, and the first glaze recognition model is trained once based on the obtained first aging feature mapping parameters and each first image instance to obtain the first aging feature mapping parameters when training is completed in the next training stage; The first glaze recognition model at the completion of the last training phase is determined as the first glaze aging degree identification model, and the second glaze recognition model at the completion of the last training phase is determined as the second glaze aging degree identification model; The glaze aging degree identification model including the feature extraction components corresponding to each image type in the preset image types in the trained first glaze aging degree identification model or the second glaze recognition model is determined as the target glaze aging degree identification model.

2. The method according to claim 1, characterized in that The component architecture of the feature mapping components of each glaze recognition model in the initial glaze aging degree identification model is consistent, and the feature mapping components of each glaze recognition model include a first number of texture analysis components; For each second glaze recognition model, the method of determining a first directional parameter group for fusion in the first aging feature mapping parameter and a second directional parameter group for fusion in the second aging feature mapping parameter of the second glaze recognition model according to the directional consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed, comprises: Determine, according to the directional consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed, a second texture analysis component for fusion among the texture analysis components of the second feature mapping component of the second glaze recognition model, and use the parameters of each second texture analysis component as a second directional parameter group for fusion in the second aging feature mapping parameters of the second glaze recognition model; According to each of the second texture analysis components, a first texture analysis component for fusion among each of the texture analysis components of the first feature mapping component is determined, and the parameters of each of the first texture analysis components are used as the first orientation parameter group for fusion in the first aging feature mapping parameters, wherein the first texture analysis component is a texture analysis component other than the texture analysis components corresponding to each of the second texture analysis components among the texture analysis components of the first feature mapping component.

3. The method according to claim 1, characterized in that The second orientation parameter group of each second glaze recognition model when training is completed in the next training stage and the first orientation parameter group of the first glaze recognition model when training is completed in the current training stage are fused to obtain the first aging feature mapping parameter corresponding to the training in the next training stage, including: Acquire a first feature contribution corresponding to the first glaze surface recognition model, and a second feature contribution corresponding to each of the second glaze surface recognition models; For the target second orientation parameter group of the texture analysis component corresponding to the same feature level identifier in the second orientation parameter group of each second feature mapping component, obtain the model parameter update direction of each second aging feature mapping parameter when the current training phase is completed; Obtaining a third feature contribution of each second feature mapping component according to a model parameter update direction of each second aging feature mapping parameter when the current training phase is completed, wherein the third feature contribution is negatively correlated with a parameter update direction of the corresponding second aging feature mapping parameter; For the target second orientation parameter group of the texture analysis component corresponding to the same feature level identifier in the second orientation parameter group of each second feature mapping component, a linear superposition process is performed with the third feature contribution of the corresponding second feature mapping component to obtain the first integration result corresponding to the texture analysis component with the same feature level identifier in the next stage; determining a mean value of the first integrated result of the texture analysis component identified by each feature level as a third orientation parameter group; Based on the first feature contribution and the second feature contribution, the first orientation parameter group and the third orientation parameter group of the first glaze recognition model when the training is completed in the current training stage are weighted fused to obtain the first aging feature mapping parameters corresponding to the next training stage.

4. The method according to claim 1, characterized in that: The directional consistency between the first parameter update direction and the second parameter update direction includes a consistency coefficient between the first sub-parameter update direction and the second sub-parameter update direction of each corresponding texture analysis component in the first glaze recognition model and the second glaze recognition model; For each second glaze recognition model, according to the directional consistency between the first parameter update direction and the second parameter update direction when each training stage in the latest target training stage is completed, a first directional parameter group for fusion in the first aging feature mapping parameter and a second directional parameter group for fusion in the second aging feature mapping parameter are determined, including: For each texture analysis component in each second glaze recognition model, if the consistency coefficient of the texture analysis component at the completion of each training stage in the latest target training stage does not exceed the preset consistency threshold, the texture analysis component is determined as a target texture analysis component; In the feature selection matrix of the second glaze recognition model at the completion of the current training phase, the parameters corresponding to the target texture analysis component are unassigned for fusion, and the feature selection matrix corresponding to the training in the next training phase is obtained; wherein the feature selection matrix is ​​used to indicate the second directional parameter group used for fusion in the second aging feature mapping parameters of the second glaze recognition model, and the feature selection matrix corresponding to the training in the first training phase is a pre-set feature selection matrix; According to the feature selection matrix corresponding to the training in the next training stage, a first orientation parameter group for fusion in the first aging feature mapping parameters and a second orientation parameter group for fusion in the second aging feature mapping parameters of the second glaze recognition model are determined, wherein the first orientation parameter group is an array of the first aging feature mapping parameters excluding the parameters corresponding to the second orientation parameter group.

5. The method according to claim 1, characterized in that The model parameter update direction of the second aging feature mapping parameter when the current training phase is completed includes: Fusing a first orientation parameter group of the first aging feature mapping parameter when the previous training stage is completed with a second orientation parameter group of the second feature mapping component when the current training stage is completed, to obtain a second fusion result corresponding to the current training stage; Obtaining a model parameter update direction of the second aging feature mapping parameter when the current training stage is completed according to a deviation between the second aging feature mapping parameter and the second fusion result when the current training stage is completed; The model parameter update direction of the first aging feature mapping parameter when the current training phase is completed includes: According to the deviation between the completion time of the current training phase and the completion time of the previous training phase when the first aging characteristic mapping parameter is completed, the model parameter update direction of the first aging characteristic mapping parameter when the current training phase is completed is obtained.

6. The method according to claim 1, characterized in that Each image instance also includes a glaze aging degree target value of an aging degree grading result of each glaze area in the corresponding porcelain image; the aging degree grading result is determined based on at least two indicators of glaze crack density, color change rate, and microscopic bubble distribution characteristics; the first feature mapping component and each second feature mapping component have the same feature mapping component structure, both including at least one aging texture analysis unit, and each second feature extraction component of the second glaze recognition model includes a feature extraction unit corresponding to the aging texture analysis unit of the second feature mapping component in the second glaze recognition model; For each second glaze recognition model, each training stage also includes: For the second feature extraction component of the second glaze recognition model, the first feature extraction component parameters of the first feature extraction component of the same image type when the previous training phase is completed are determined as the second feature extraction component parameters of the second feature extraction component after being updated in the current training phase; Determining the second glaze recognition model after updating each second feature extraction component parameter and the second aging feature mapping parameter as the second glaze recognition model in the present training phase; For each aging texture analysis unit of the second glaze recognition model in the current training phase, integrating feature information output by a feature extraction unit corresponding to each second feature extraction component of the second glaze recognition model, to obtain a first integrated feature corresponding to the aging texture analysis unit; For each aging texture analysis unit of the second glaze recognition model in the current training phase, a standardization operation is performed using gated fusion according to the reference feature identifier at the completion of the previous training phase, the first integrated feature corresponding to the aging texture analysis unit, and the number of texture analysis components in the aging texture analysis unit to obtain a fourth feature contribution; The first integrated feature corresponding to the aging texture analysis unit is weighted by the fourth feature contribution to obtain an optimized first integrated feature, and the first integrated feature before and after the optimization of the aging texture analysis unit is integrated to obtain a second integrated feature corresponding to the aging texture analysis unit; Determine the second integrated feature corresponding to the first aging texture analysis unit as the mapping result of the first aging texture analysis unit, and for non-first aging texture analysis units, perform feature mapping according to the second integrated feature corresponding to the non-first aging texture analysis unit and the mapping result of the previous aging texture analysis unit to obtain the mapping result of the non-first aging texture analysis unit; According to the mapping result of the last aging texture analysis unit, the predicted aging degree grading result of each glaze area in the corresponding porcelain image of the image instance is obtained, and according to the deviation between the predicted aging degree grading result of each glaze area and the glaze aging degree target value, each second feature extraction component parameter and the second aging feature mapping parameter of the second glaze recognition model in the current training stage are optimized to obtain the second glaze recognition model when the training stage is completed; wherein each reference feature identifier corresponding to each aging texture analysis unit is obtained by feature grouping the mapping features output by the corresponding aging texture analysis unit for the glaze area of ​​the same aging degree grading result in each first image instance; For the first glaze recognition model, each training stage further includes: For each first feature extraction component, the second feature extraction component parameters of all second feature extraction components of the same image type at the completion of this training phase are merged to obtain updated first feature extraction component parameters of the same image type of the first feature extraction component; The first glaze recognition model after updating each first feature extraction component parameter and the first aging feature mapping parameter is determined as the first glaze recognition model in this training stage, and the first glaze recognition model in this training stage is trained according to the image instance of the first glaze recognition model until the first training termination state is met, so as to obtain the first glaze recognition model when this training stage is completed.

7. The method according to claim 6, characterized in that The reference feature identifiers of each aging texture analysis unit are updated in the following way: For each first image instance, according to the glaze aging degree target value of the first image instance, determine the aging degree classification result corresponding to the mapping feature of each glaze area output by the aging texture analysis unit when the previous training stage is completed; For each first image instance, integrating the mapping features of each glazed area of ​​the same aging degree grading result output by the aging texture analysis unit, to obtain the recognition processing features of the corresponding aging degree grading result of the first image instance by the aging texture analysis unit; For the same aging degree grading result, the recognition processing features of the aging degree grading result of each first image instance by the aging texture analysis unit are grouped according to a preset number of categories, and the features of the cluster center of each feature group in the current training stage are determined; For each feature group, the features of the cluster centers of the feature group in the current training stage and the previous training stage are integrated to obtain a reference feature identifier corresponding to the feature group when the aging texture analysis unit completes the current training stage.

8. The method according to claim 1, characterized in that The method is applied in a distributed collaborative training system, wherein the first glaze recognition model is deployed in a global aggregator, and one of the second glaze recognition models is deployed in a local training node, and the method is executed by the global aggregator, wherein obtaining the second aging feature mapping parameters of each of the second glaze recognition models when training is completed in the current training phase includes: Receiving model update parameters of each second glaze recognition model sent by each local training node, wherein the model update parameters include second aging feature mapping parameters of the second glaze recognition model when training is completed in the current training stage; For each second glaze recognition model, after obtaining the second aging feature mapping parameters corresponding to the next training stage, the method further includes: The model update parameters corresponding to the next training stage of the second glaze recognition model are transmitted to the corresponding local training node, so that the local training node performs a training iteration on the second glaze recognition model based on the model update parameters and the corresponding second image instance, and obtains the model update parameters of the second glaze recognition model when the training is completed in the next stage.

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

Images