Jade quality intelligent evaluation method and system based on deep learning reinforcement feature extraction

Abstract

The invention discloses an intelligent jade quality evaluation method and system based on deep learning enhanced feature extraction, and relates to the field of artificial intelligence, and the method comprises the steps: firstly, obtaining jade appearance data, importing a target feature distiller in a target enhanced feature extraction path corresponding to each jade type to obtain target appearance distillation information, a target jade quality grading model is obtained by constraining and optimizing a preset jade quality association unit through a characteristic difference degree, and finally, the jade quality grade is determined according to the target jade appearance characteristics and a target quality grading path, so that intelligent evaluation of the jade quality is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for intelligently evaluating jade quality based on deep learning and enhanced feature extraction. Background Art

[0002] At present, jade quality assessment mostly relies on manual experience, which is inefficient and easily affected by subjective factors. The assessment results of different appraisers may differ, making it difficult to ensure the accuracy and consistency of the assessment. With the development of deep learning technology, although it has been applied in some fields, in terms of jade quality assessment, how to effectively use deep learning to extract jade features and accurately assess quality is still a difficult problem. Summary of the invention

[0003] The object of the present invention is to provide a method and system for intelligent jade quality assessment based on deep learning enhanced feature extraction.

[0004] In a first aspect, an embodiment of the present invention provides a method for intelligently assessing jade quality based on deep learning enhanced feature extraction, comprising:

[0005] Get jade appearance data;

[0006] Importing the jade appearance data into the target feature distiller in the target enhancement feature extraction path corresponding to each jade type to perform feature distillation operation, so as to obtain the target appearance distillation information corresponding to each jade type;

[0007] The target appearance distillation information corresponding to each jade type is imported into the target feature extractor in the target enhanced feature extraction path corresponding to each jade type for feature extraction, so as to obtain the target jade appearance features corresponding to each jade type; the target jade quality grading model is obtained by tuning the preset jade quality association unit in the preset enhanced feature extraction path according to the feature difference constraint, the feature difference constraint is used to quantify the correlation of the target jade quality association features, and the target jade quality association features are the jade quality association features generated by the preset jade quality association units corresponding to different jade types;

[0008] The target jade quality grade corresponding to the jade appearance data is determined according to the target jade appearance features corresponding to each jade type and the target quality grading path corresponding to each jade type.

[0009] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein 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 a jade quality intelligent assessment method and system based on deep learning enhanced feature extraction disclosed by the present invention, by acquiring jade appearance data, importing the target feature distiller in the target enhanced feature extraction path corresponding to each jade type to obtain target appearance distillation information, and then obtaining the target jade appearance features through the target feature extractor, wherein the target jade quality grading model is obtained by optimizing the preset jade quality association unit by feature difference constraint, and finally determining the jade quality grade according to the target jade appearance features and the target quality grading path, thereby realizing intelligent assessment of jade quality. 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 are briefly introduced below. It should be understood that the following drawings only illustrate certain 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 also be obtained based on these drawings without creative work.

[0012] Figure 1 A schematic diagram of the steps of a jade quality intelligent assessment method based on deep learning enhanced feature extraction provided by an embodiment of the present invention;

[0013] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, 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 drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0015] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0016] In order to solve the technical problems in the aforementioned background technology, Figure 1 A flow chart of a method for intelligently assessing jade quality based on deep learning and enhanced feature extraction provided in an embodiment of the present disclosure is provided below. The method for intelligently assessing jade quality based on deep learning and enhanced feature extraction is introduced in detail.

[0017] Step S201, obtaining jade appearance data;

[0018] Step S202, importing the jade appearance data into the target feature distiller in the target enhancement feature extraction path corresponding to each jade type to perform feature distillation operation, so as to obtain the target appearance distillation information corresponding to each jade type;

[0019] Step S203, importing the target appearance distillation information corresponding to each jade type into the target feature extractor in the target enhanced feature extraction path corresponding to each jade type for feature extraction, and obtaining the target jade appearance features corresponding to each jade type; the target jade quality grading model is obtained by tuning the preset jade quality association unit in the preset enhanced feature extraction path according to the feature difference constraint, the feature difference constraint is used to quantify the correlation of the target jade quality association features, and the target jade quality association features are the jade quality association features generated by the preset jade quality association units corresponding to different jade types;

[0020] Step S204, determining the target jade quality grade corresponding to the jade appearance data according to the target jade appearance features corresponding to each jade type and the target quality grading path corresponding to each jade type.

[0021] In an embodiment of the present invention, illustratively, the server is connected to a high-definition camera device in the market. When a merchant places a piece of jade in a specific shooting area, the camera will shoot the jade in all directions according to preset parameters, such as suitable lighting conditions, shooting angles and resolutions. After the shooting is completed, the image data containing the appearance information of the jade, such as color, texture, shape, etc., are transmitted to the server through the network in the form of digital signals. For example, a piece of Hetian jade is placed in the shooting area, and the camera takes multiple clear pictures from different sides. These picture data are quickly transmitted to the server and become the jade appearance data to be processed. The server pre-stores the target enhancement feature extraction paths corresponding to a variety of jade types. Taking jadeite as an example, the server will accurately import the jade appearance data just obtained into the target feature distiller in the target enhancement feature extraction path corresponding to jadeite. This target feature distiller is like a knowledge refining factory, which will perform complex mathematical operations and feature screening on the input jade appearance data. It will extract key and representative feature information from the jade appearance data, and integrate and refine this information to remove redundant parts. After a series of operations, the target appearance distillation information corresponding to the jadeite type is obtained. For example, the target feature distiller extracts key features related to the color brightness and texture uniqueness of jadeite from the jade appearance data to form the target appearance distillation information. For jadeite, the server then imports its target appearance distillation information into the target feature extractor in the target enhanced feature extraction path corresponding to the jade type. The target feature extractor contains multiple functional units working together. Assuming the first feature focusing unit in the target feature extractor, it will focus on certain key features of the target appearance distillation information like a magnifying glass, perform feature focusing operations, highlight features closely related to the quality of jadeite, and obtain the first focused feature map. Then, the first focused feature map and the target appearance distillation information are imported into the first feature reorganizer, which will recombine these features in a more reasonable way like a puzzle master to obtain the first intermediate feature data. At the same time, the first focused feature map and the target appearance distillation information will also be imported into the preset jade quality association unit for feature calibration operations to obtain the first jade quality association feature. Finally, the first intermediate feature data, the first jade quality-related feature, the first focused feature map, and the target appearance distillation information are all sent to the first feature aggregation node, where feature aggregation operations are performed to aggregate all relevant features together, and finally the target jade appearance features corresponding to the jadeite are obtained. These target jade appearance features contain various key information that can accurately reflect the quality of jadeite. The target jade quality grading model here is obtained by the server previously tuning the preset jade quality-related unit in the preset enhanced feature extraction path according to the feature difference constraint.For example, during the training phase, the server compares the jade quality association features generated by the preset jade quality association units corresponding to the two types of jade, jadeite and Hetian jade, quantifies the correlation between them, and obtains the feature difference constraint. Through this constraint, the preset jade quality association unit is adjusted and optimized to obtain the target jade quality grading model. The server imports the target jade appearance features corresponding to jadeite into the target quality grading path corresponding to jadeite. This target quality grading path is like an experienced appraiser, which will make quality inferences based on the input target jade appearance features. Through the internal preset algorithms and models, these features are analyzed and judged to obtain the first jade quality inference value of jadeite. Similarly, for other types of jade, the server will perform similar operations to obtain their respective first jade quality inference values. Then, the server will combine the first jade quality inference values ​​configured for multiple types of jade to determine the target jade quality level corresponding to the jade appearance data. For example, for a piece of jade appearance data, the server analyzes it under different jade types (such as jadeite, Hetian jade, Xiuyan jade, etc.) and obtains the first jade quality inference value corresponding to each type. Then, the server will determine the target jade quality grade of the jade, such as high quality, medium quality or low quality, based on the comprehensive situation of these inference values. In addition, the server will also import the jade appearance data into the target dynamic weight allocator in the target jade quality grading model for weight optimization and allocation to obtain the target feature contribution parameter. According to this target feature contribution parameter, the server performs adaptive synthesis operations on the first jade quality inference values ​​configured for multiple jade types respectively, and obtains the second jade quality inference value corresponding to the jade appearance data. Finally, based on this second jade quality inference value, the server determines the target jade quality grade more accurately. For example, the target feature contribution parameter indicates that the jade feature is more important for the jade quality assessment. During the synthesis operation, the server will give the first jade quality inference value corresponding to the jade a higher weight, thereby obtaining a more accurate second jade quality inference value, and then determining a more accurate target jade quality grade.

[0022] In an embodiment of the present invention, the target jade quality grading model is obtained by training through the following process, and can be implemented through the following examples.

[0023] Obtaining a first preset jade appearance data set;

[0024] Importing the first preset jade appearance data set into the preset enhanced feature extraction paths respectively configured for the multiple jade types in the initial integrated model to perform feature extraction operations, so as to obtain preset jade quality-related features corresponding to each jade type;

[0025] Determine the characteristic difference constraint corresponding to the target jade type combination according to the preset jade quality associated characteristics corresponding to the target jade type combination; the target jade type combination is any two types of the multiple jade types;

[0026] According to the characteristic difference constraints corresponding to the target jade type combination, the preset jade quality association units corresponding to the target jade type combination are trained to obtain the target jade quality grading model.

[0027] In an embodiment of the present invention, illustratively, the server establishes connections with multiple data sources such as jade identification institutions, major jade trading markets and jade processing plants. These data sources collect and organize a large amount of jade appearance data of different types and qualities, covering jade images under various shooting angles and lighting conditions. For example, the data source provides Hetian jade appearance images from place A, jadeite raw stones and finished products from place B, Xiuyan jade pictures from place C, etc. The server integrates these jade appearance data to form a first preset jade appearance data set, providing a rich data basis for subsequent model training. The initial integrated model in the server sets respective preset enhanced feature extraction paths for different types of jade. Taking jadeite and Hetian jade as examples, the server imports the relevant data in the first preset jade appearance data set into the preset enhanced feature extraction paths corresponding to jadeite and Hetian jade, respectively. Assuming that the preset feature distiller in the preset enhanced feature extraction path of jadeite will perform feature extraction on the input jadeite appearance data to obtain preset appearance distillation information. Then, this information enters the first feature focusing unit, focusing on key features to generate a first focused feature map. The graph and the preset appearance distillation information are reorganized in the first feature reorganizer to obtain the first intermediate feature data. At the same time, they are subjected to feature calibration calculation in the preset jade quality association unit to generate the first jade quality association feature. Finally, the preset jade quality association feature corresponding to jadeite is obtained through the first feature aggregation node calculation. Hetian jade also obtains its corresponding preset jade quality association feature according to this process. The server selects any two types from multiple jade types as the target jade type combination, such as jadeite and Hetian jade. It compares the preset jade quality association features corresponding to the two jades, such as the color brightness and crystal structure characteristics of jadeite with the warmth and texture fineness characteristics of Hetian jade. The degree of difference between these features is calculated through a complex algorithm, and their correlation is quantified, so as to determine the feature difference constraint corresponding to the target jade type combination of jadeite and Hetian jade. This constraint value reflects the difference between the two in quality association features. According to the feature difference constraint corresponding to the target jade type combination of jadeite and Hetian jade, the server adjusts and optimizes the preset jade quality association unit in the preset enhanced feature extraction path of each of the two jades. During the training process, the server continuously adjusts the parameters of the preset jade quality association unit so that the model can better distinguish and understand the differences in quality association characteristics of different jade types. After multiple rounds of training, when the model is able to accurately reflect these differences when processing the characteristics of different jade types, the server obtains the target jade quality grading model, which will be used to accurately assess the quality grade of jade in the future.

[0028] In an embodiment of the present invention, the preset enhanced feature extraction path corresponding to each jade type includes a preset feature distiller and a preset feature extractor, and the preset feature extractor includes a first feature focusing unit, a first feature recombiner, a preset jade quality association unit and a first feature aggregation node; the first preset jade appearance data set is imported into the initial integrated model, and the preset enhanced feature extraction paths respectively configured for the multiple jade types perform feature extraction operations to obtain the preset jade quality association features corresponding to each jade type, which can be implemented through the following examples.

[0029] Importing the first preset jade appearance data set into a preset feature distiller corresponding to each jade type to perform a feature distillation operation, to obtain preset appearance distillation information corresponding to each jade type;

[0030] Importing the preset appearance distillation information corresponding to each jade type into the first feature focusing unit corresponding to each jade type to perform feature focusing operation, so as to obtain the first focusing feature map corresponding to each jade type;

[0031] Importing the first focused feature map corresponding to each jade type and the preset appearance distillation information corresponding to each jade type into the first feature recombiner corresponding to each jade type to perform feature recombination operation, so as to obtain the first intermediate feature data corresponding to each jade type;

[0032] Importing the first focused feature map corresponding to each jade type and the preset appearance distillation information corresponding to each jade type into the preset jade quality association unit corresponding to each jade type to perform feature calibration calculation, so as to obtain the first jade quality association feature corresponding to each jade type;

[0033] The first intermediate feature data corresponding to each jade type, the first jade quality-related feature corresponding to each jade type, the first focused feature map corresponding to each jade type and the preset appearance distillation information corresponding to each jade type are imported into the first feature aggregation node corresponding to each jade type to perform feature aggregation operation to obtain the second intermediate feature data corresponding to each jade type.

[0034] In an embodiment of the present invention, exemplarily, the server obtains a first preset jade appearance data set containing multiple jade appearance data, which contains appearance images of jades such as jadeite and Hetian jade. Taking jadeite as an example, the server imports the jadeite appearance data in the data set into the preset feature distiller corresponding to jadeite. This preset feature distiller is like a filter, which screens and refines a large amount of jadeite appearance data. It identifies key features closely related to the quality of jadeite, such as the unique emerald green tone, fiber interweaving structure, etc., removes irrelevant information, such as minor flaws in the shooting background, and finally generates the preset appearance distillation information corresponding to jadeite. Similarly, for the Hetian jade appearance data in the data set, the server will also import it into the preset feature distiller corresponding to Hetian jade, extract the key features of Hetian jade, such as mutton-like whiteness, delicate texture, etc., to form the preset appearance distillation information of Hetian jade. The server imports the preset appearance distillation information of jadeite into the first feature focusing unit corresponding to jadeite. This unit is like a magnifying glass, which further focuses on the key features in the preset appearance distillation information. For example, it focuses on the core features of jadeite, such as the brightness of its color and the change of its transparency, and presents it in a visual atlas form to generate the first focused feature atlas of jadeite, making the key quality features of jadeite clearer and more specific. For Hetian jade, the first feature focusing unit focuses on the subtle differences in its warmth and moisture, the density of its texture, and other features to form the first focused feature atlas of Hetian jade, highlighting the unique quality characteristics of Hetian jade. The server inputs the first focused feature atlas of jadeite and the preset appearance distillation information into the first feature reorganizer corresponding to jadeite. The first feature reorganizer is like a puzzle master. It recombines these two types of information in a new and more logical way according to preset rules and algorithms. For example, the color features and structural features of jadeite are rearranged and combined according to the weights of their impact on quality to obtain the first intermediate feature data of jadeite. This data form is more conducive to the subsequent in-depth analysis of jadeite quality. Similarly, the first focused feature atlas of Hetian jade and the preset appearance distillation information are reorganized in the first feature reorganizer corresponding to Hetian jade to generate the first intermediate feature data of Hetian jade. For jadeite, the server imports its first focused feature atlas and preset appearance distillation information into the preset jade quality association unit. This unit is like a calibrator, which performs calibration operations on these two types of information. It calibrates the characteristics of jadeite based on the existing jadeite quality standards and a large amount of sample data, such as calibrating the correspondence between jadeite color and quality grade, to ensure the accuracy and reliability of the characteristics, thereby obtaining the first jade quality-related characteristics of jadeite, accurately reflecting the quality-related information of jadeite. Hetian jade also calibrates the characteristics in its preset jade quality-related unit through the same operation to generate the first jade quality-related characteristics of Hetian jade. The server inputs all the first intermediate feature data, first jade quality-related features, first focused feature maps, and preset appearance distillation information of jadeite into the first feature aggregation node corresponding to jadeite.This node is like a fusion center, which comprehensively integrates the feature information obtained from these different stages. Through complex aggregation operations, the quality characteristics of various aspects of jade are integrated together to form a comprehensive and more representative second intermediate feature data, laying the foundation for the subsequent accurate assessment of jade quality. Similarly, the relevant feature information of Hetian jade is aggregated in the first feature aggregation node corresponding to Hetian jade to obtain the second intermediate feature data of Hetian jade.

[0035] In the embodiments of the present invention, the following implementation modes are also provided.

[0036] Obtaining a first jade quality target value corresponding to each preset jade appearance data in the first preset jade appearance data set;

[0037] Importing the second intermediate feature data corresponding to each jade type into the preset quality grading path corresponding to each jade type in the initial integrated model to perform quality inference, and obtaining a third jade quality inference value corresponding to each jade type;

[0038] Importing the first preset jade appearance data set into the preset dynamic weight allocator in the initial integrated model for weight optimization allocation, and obtaining a preset feature contribution parameter corresponding to each preset jade appearance data in the first preset jade appearance data set;

[0039] According to the preset feature contribution parameter, an adaptive synthesis operation is performed on the third jade quality inference values ​​respectively configured for the plurality of jade types to obtain a fourth jade quality inference value corresponding to each preset jade appearance data in the first preset jade appearance data set;

[0040] Determining a first fusion consistency error according to the fourth jade quality inference value and the first jade quality target value;

[0041] The method of training the preset jade quality association unit corresponding to the target jade type combination according to the characteristic difference constraint corresponding to the target jade type combination to obtain the target jade quality grading model includes:

[0042] According to the feature difference constraint corresponding to the target jade type combination, the preset jade quality association unit corresponding to the target jade type combination is trained, and according to the first fusion consistency error, the preset dynamic weight allocator is trained to obtain the target jade quality grading model.

[0043] In an embodiment of the present invention, illustratively, the server obtains the accurate quality assessment result corresponding to each piece of jade in the first preset jade appearance data set from the jade appraisal authority, that is, the first jade quality target value. For example, there is a piece of jade appearance data in the data set, and the appraisal agency assesses the jade as "high-grade quality" according to the industry standard. This "high-grade quality" is the first jade quality target value corresponding to the appearance data of this piece of jade. For other jade appearance data such as Hetian jade and Xiuyan jade in the data set, there are also corresponding authoritatively assessed quality target values, which provide accurate reference basis for subsequent model training. The server inputs the second intermediate feature data of jade obtained by the previous steps into the preset quality grading path corresponding to jade in the initial integrated model. This preset quality grading path is like a "simulated appraiser". According to the internally set algorithm and rules, the second intermediate feature data of jade is analyzed and judged, so as to infer the quality grade of this piece of jade, and obtain the third jade quality inference value, such as inferring that the jade is "medium-high grade". Similarly, the server imports the second intermediate feature data of Hetian jade into the preset quality grading path corresponding to Hetian jade, and obtains the third jade quality inference value of Hetian jade, such as "medium quality". The server inputs the entire first preset jade appearance data set into the preset dynamic weight allocator in the initial integrated model. This allocator analyzes the importance of different jade types and different features in the data set to the quality assessment. For example, for jadeite, the color feature has a greater impact on the quality, followed by the transparency feature, and then the texture feature. The preset dynamic weight allocator assigns different weights to each feature of each piece of jade according to the difference in importance, and obtains the preset feature contribution parameters corresponding to each preset jade appearance data. For a specific piece of jadeite appearance data, the color feature may be assigned a weight of 0.5, the transparency feature is assigned a weight of 0.3, and the texture feature is assigned a weight of 0.2. These weight values ​​are part of the preset feature contribution parameters of the jadeite. The server comprehensively considers the third jade quality inference values ​​of different jade types based on the preset feature contribution parameters obtained above. For example, for a piece of jade appearance data involving both jadeite and Hetian jade features, the third jade quality inference value of jadeite is "medium-high grade", and that of Hetian jade is "medium quality". Combined with their respective preset feature contribution parameters, such as the jade feature with a higher weight, the server performs adaptive synthesis operations through a specific algorithm to fuse the inferred values ​​of the two to obtain a more comprehensive fourth jade quality inference value, such as "mid-to-high-end leaning towards high-end", which more comprehensively reflects the quality of the jade. The server compares the fourth jade quality inference value of each preset jade appearance data with the first jade quality target value obtained from the appraisal agency. For example, the fourth jade quality inference value of a piece of jade is "mid-to-high-end leaning towards high-end", while its first jade quality target value is "high-end quality". The server determines the first fusion consistency error by calculating the degree of difference between the two.This error reflects the deviation between the current model inference result and the actual quality target value. Taking the target jade type combination of jadeite and Hetian jade as an example, the server adjusts the parameters and optimizes the training of the preset jade quality association units in the preset enhanced feature extraction paths of jadeite and Hetian jade according to the previously determined corresponding feature difference constraints, so that the model can better capture the difference in quality characteristics of the two. At the same time, according to the first fusion consistency error, the preset dynamic weight allocator is also trained. For example, if it is found that the weight distribution of a certain feature is unreasonable and causes a large error, the weight of the feature is adjusted. After repeated training, when the error between the inference result of the model and the actual quality target value is within an acceptable range, the server obtains the target jade quality grading model, which is more accurate and reliable in jade quality assessment.

[0044] In an embodiment of the present invention, the preset enhanced feature extraction path corresponding to each jade type includes a preset feature distiller and a preset feature extractor, the preset feature extractor includes a second feature focusing unit, a second feature recombiner, a preset jade quality association unit and a second feature aggregation node, and the preset jade quality association unit includes a first preset quality association subunit and a second preset quality association subunit; the first preset jade appearance data set is imported into the initial integrated model, and the preset enhanced feature extraction paths respectively configured for the multiple jade types perform feature extraction operations to obtain the preset jade quality association features corresponding to each jade type, which can be implemented through the following examples.

[0045] Importing the first preset jade appearance data set into a preset feature distiller corresponding to each jade type to perform a feature distillation operation, to obtain preset appearance distillation information corresponding to each jade type;

[0046] Importing the preset appearance distillation information corresponding to each jade type into the second feature focusing unit corresponding to each jade type to perform feature focusing operation, so as to obtain the second focusing feature map corresponding to each jade type;

[0047] Importing the second focused feature map corresponding to each jade type into the first preset quality association subunit corresponding to each jade type to perform feature calibration calculation, so as to obtain the second jade quality association feature corresponding to each jade type;

[0048] Importing the second jade quality-related feature corresponding to each jade type and the preset appearance distillation information corresponding to each jade type into the second feature recombiner corresponding to each jade type to perform feature recombining operation, so as to obtain third intermediate feature data corresponding to each jade type;

[0049] Importing the third intermediate feature data corresponding to each jade type into the second preset quality association subunit corresponding to each jade type to perform feature calibration calculation, so as to obtain the third jade quality association feature corresponding to each jade type;

[0050] The third jade quality-related feature corresponding to each jade type, the second jade quality-related feature corresponding to each jade type, and the preset appearance distillation information corresponding to each jade type are imported into the second feature aggregation node corresponding to each jade type to perform feature aggregation operation to obtain the fourth intermediate feature data corresponding to each jade type.

[0051] In an embodiment of the present invention, exemplarily, the server has a first preset jade appearance data set containing multiple jade appearance data, which includes various types of jade such as jadeite and Xiuyan jade. Taking Xiuyan jade as an example, the server imports the appearance data of Xiuyan jade in the data set into the preset feature distiller corresponding to Xiuyan jade. The preset feature distiller is like an information screening refinery, which carefully combs the appearance data of Xiuyan jade. It identifies and extracts the unique color features of Xiuyan jade, such as common green and yellow tones, and key information such as unique texture features, while removing some interference information generated during the shooting process, such as tiny light spots, visual deviations caused by shooting angles, etc., and finally generates the preset appearance distillation information corresponding to Xiuyan jade. Similarly, for the jadeite appearance data in the data set, the server also imports it into the preset feature distiller corresponding to jadeite, extracts the key features of jadeite's unique jadeite and glass luster, and forms the preset appearance distillation information of jadeite. The server transmits the preset appearance distillation information of Xiuyan jade to the second feature focusing unit corresponding to Xiuyan jade. This unit is like a precise magnifying glass, which further focuses and strengthens the key features in the preset appearance distillation information. For example, it emphasizes the core features of Xiuyu jade, such as the uniformity of color and the direction of texture, and presents them in the form of a map, generating the second focused feature map of Xiuyu jade, making the key features related to quality of Xiuyu jade more intuitive and prominent. For jadeite, the second feature focusing unit focuses on the hierarchical changes in the transparency of jadeite, the distribution characteristics of crystal particles, etc., forming the second focused feature map of jadeite, and clearly showing the key characteristics of jadeite that affect quality. The server inputs the second focused feature map of Xiuyu jade into the first preset quality association subunit corresponding to Xiuyu jade. This subunit is similar to a calibration instrument, which calibrates the characteristics of Xiuyu jade according to the relevant standards of Xiuyu jade quality and a large amount of sample data. For example, it calibrates the correspondence between the color of Xiuyu jade and the quality grade, ensures the accuracy of the association between the feature and the actual quality, and thus obtains the second jade quality association feature of Xiuyu jade, which can accurately reflect the characteristics of Xiuyu jade related to quality. Similarly, for jadeite, the first preset quality association subunit calibrates the jadeite features according to the jadeite quality standard, and generates the second jadeite quality association feature of jadeite. The server sends the second jade quality-related features and preset appearance distillation information of Xiuyu jade to the second feature recombiner corresponding to Xiuyu jade. The second feature recombiner is like an experienced jigsaw puzzle master, recombining the two types of information according to specific rules and algorithms. It may rearrange and integrate the color features, texture features and calibrated quality-related features of Xiuyu jade according to the weights of the factors affecting the quality of Xiuyu jade to obtain the third intermediate feature data of Xiuyu jade. This data form is more conducive to the subsequent in-depth analysis of the quality of Xiuyu jade. Similarly, the relevant information of jadeite is reorganized in the second feature recombiner corresponding to jadeite to generate the third intermediate feature data of jadeite. The server inputs the third intermediate feature data of Xiuyu jade into the second preset quality-related subunit corresponding to Xiuyu jade.This subunit calibrates the features again, and calibrates the features of Xiuyu again from another angle or a more detailed level according to the Xiuyu quality standard. For example, it further calibrates the subtle relationship between the texture and quality of Xiuyu jade, so that the correlation between the obtained features and the actual quality is more accurate, thereby obtaining the third jade quality-related features of Xiuyu jade. Similarly, the third intermediate feature data of jadeite is calibrated in the second preset quality-related subunit corresponding to jadeite to generate the third jade quality-related features of jadeite. The server inputs the third jade quality-related features, the second jade quality-related features, and the preset appearance distillation information of Xiuyu jade into the second feature aggregation node corresponding to Xiuyu jade. This node is like a fusion center, which comprehensively integrates various quality-related features obtained from the different stages of Xiuyu jade through complex aggregation operations. It combines various features of Xiuyu jade, such as color, texture, and calibrated quality associations, to form a more comprehensive and representative fourth intermediate feature data, providing more powerful data support for the accurate assessment of Xiuyu jade quality. Similarly, the relevant features of jadeite are aggregated at the second feature aggregation node corresponding to jadeite to obtain the fourth intermediate feature data of jadeite.

[0052] In the embodiments of the present invention, the following implementation modes are also provided.

[0053] Obtaining a first jade quality target value corresponding to each preset jade appearance data in the first preset jade appearance data set;

[0054] Importing the fourth intermediate feature data corresponding to each jade type into the preset quality grading path corresponding to each jade type in the initial integrated model to perform quality inference, and obtaining a fifth jade quality inference value corresponding to each jade type;

[0055] Importing the first preset jade appearance data set into the preset dynamic weight allocator in the initial integrated model for weight optimization allocation, and obtaining a preset feature contribution parameter corresponding to each preset jade appearance data in the first preset jade appearance data set;

[0056] According to the preset feature contribution parameter, an adaptive synthesis operation is performed on the fifth jade quality inference values ​​respectively configured for the plurality of jade types to obtain a sixth jade quality inference value corresponding to each preset jade appearance data in the first preset jade appearance data set;

[0057] Determining a second fusion consistency error according to the sixth jade quality inference value and the first jade quality target value;

[0058] The method of training the preset jade quality association unit corresponding to the target jade type combination according to the characteristic difference constraint corresponding to the target jade type combination to obtain the target jade quality grading model includes:

[0059] According to the feature difference constraint corresponding to the target jade type combination, the preset jade quality association unit corresponding to the target jade type combination is trained, and according to the second fusion consistency error, the preset dynamic weight allocator is trained to obtain the target jade quality grading model.

[0060] In an embodiment of the present invention, illustratively, the server cooperates with a professional jade identification laboratory to obtain the authoritative quality assessment result of each piece of jade in the first preset jade appearance data set, that is, the first jade quality target value. For example, there is a piece of jadeite rough stone appearance data from place B in the data set. The identification laboratory evaluates the jadeite rough stone as "special quality" based on the industry's unified quality identification standards, from the perspective of color, transparency, texture and other aspects. This "special quality" is the first jade quality target value corresponding to the appearance data of this piece of jadeite rough stone. For the appearance data of many different types of jade in the data set, such as Hetian jade, Xiuyan jade, etc., there are also corresponding accurate quality assessments as the first jade quality target value, providing a reliable reference benchmark for subsequent model training. The server inputs the fourth intermediate feature data of jadeite obtained through the previous complex process into the preset quality grading path corresponding to jadeite in the initial integrated model. This preset quality grading path is like an experienced "virtual appraiser" who conducts in-depth analysis of the fourth intermediate feature data of jadeite based on the internally set algorithms and rules. It will consider the color intensity, density of crystal structure and other characteristics in the fourth intermediate feature data of jadeite, and then infer the quality grade of this jadeite, and obtain the fifth jade quality inference value, for example, infer that the jadeite is "first-class quality". Similarly, the server imports the fourth intermediate feature data of Xiuyu jade into the preset quality grading path corresponding to Xiuyu jade, and combines the unique characteristics of Xiuyu jade, such as the softness of color and the fineness of texture, to obtain the fifth jade quality inference value of Xiuyu jade, such as "second-class quality". The server transmits the entire first preset jade appearance data set to the preset dynamic weight allocator in the initial integration model. This allocator will conduct a detailed analysis of the importance of various features of different jade types in the data set to quality assessment. For example, for jadeite, color occupies a large proportion in quality assessment, followed by transparency, and then impurity content. According to this difference in importance, the preset dynamic weight allocator assigns corresponding weights to each feature of each piece of jadeite preset jade appearance data, and obtains the preset feature contribution parameter corresponding to each piece of jadeite. Assume that for a piece of jade appearance data, the color feature is assigned a weight of 0.6, the transparency feature is assigned a weight of 0.3, and the impurity content feature is assigned a weight of 0.1. These weight values ​​constitute the preset feature contribution parameters of the jade. For other types of jade, such as Hetian jade, Xiuyan jade, etc., the preset dynamic weight allocator will also assign corresponding preset feature contribution parameters according to the degree of influence of their respective features on the quality. The server comprehensively processes the fifth jade quality inference values ​​of different types of jade based on the preset feature contribution parameters obtained previously. For example, for a piece of jade appearance data that combines some features of jade and Xiuyan jade, the fifth jade quality inference value of jade is "first-level quality" and that of Xiuyan jade is "second-level quality".Combined with their respective preset feature contribution parameters, since the jade feature has a higher weight in the quality assessment of this piece of jade, the server performs adaptive synthesis operations through a specific algorithm, organically integrates the inferred values ​​of the two, and obtains a sixth jade quality inference value that is more in line with the actual situation, such as "first-level quality tends to special quality", which more comprehensively and accurately reflects the quality of the jade. The server carefully compares the sixth jade quality inference value of each preset jade appearance data with the first jade quality target value obtained from the appraisal laboratory. For example, the sixth jade quality inference value of a piece of jade is "first-level quality tends to special quality", and its first jade quality target value is "special quality". The server measures the degree of difference between the two through a special calculation method, thereby determining the second fusion consistency error. This error intuitively reflects the deviation between the current model inference result and the actual quality target value, providing a key basis for subsequent model optimization. Taking the target jade type combination of jadeite and Xiuyan jade as an example, the server conducts targeted parameter adjustment and optimization training on the preset jade quality association units in the preset enhanced feature extraction paths of jadeite and Xiuyan jade, based on the previously determined corresponding feature difference constraints. This enables the model to more keenly capture the differences in quality characteristics between jadeite and Xiuyan jade. At the same time, according to the second fusion consistency error, the preset dynamic weight allocator is also trained. For example, if it is found that the weight distribution of a certain feature is unreasonable, resulting in a large error between the inferred value and the target value, the server will adjust the weight of the feature. After repeated training and optimization, when the error between the model's inference result and the actual quality target value is reduced to an acceptable range, the server successfully obtains the target jade quality grading model, which will be more accurate and reliable in jade quality assessment.

[0061] In the embodiments of the present invention, the following implementation modes are also provided.

[0062] Acquire a second preset jade appearance data set and a second jade quality target value corresponding to each preset jade appearance data in the second preset jade appearance data set;

[0063] Importing the second preset jade appearance data set into the preset enhanced feature extraction paths respectively configured for the multiple jade types in the initial integrated model to perform feature extraction operations, and obtaining fifth intermediate feature data corresponding to each jade type;

[0064] Importing the fifth intermediate characteristic data corresponding to each jade type into the preset quality grading path corresponding to each jade type to perform quality inference, and obtaining the seventh jade quality inference value corresponding to each jade type;

[0065] Determine the sub-item quality error corresponding to each jade type according to the seventh jade quality inference value corresponding to each jade type and the second jade quality target value;

[0066] The method of training the preset jade quality association unit corresponding to the target jade type combination according to the characteristic difference constraint corresponding to the target jade type combination to obtain the target jade quality grading model includes:

[0067] According to the feature difference constraints corresponding to the target jade type combination, the preset jade quality association units corresponding to the target jade type combination are trained, and according to the sub-item quality errors corresponding to each jade type, the preset model core components corresponding to each jade type in the initial integrated model are trained to obtain the target jade quality grading model; the preset model core components corresponding to each jade type are the feature processing architectures in the preset enhanced feature extraction paths corresponding to each jade type except the preset jade quality association units corresponding to each jade type.

[0068] In an embodiment of the present invention, illustratively, the server collects data from the detection agencies of multiple jade origins to form a second preset jade appearance data set. These data contain the appearance information of jades of different origins and different qualities, such as Huanglong jade from D and jasper from E. At the same time, the detection agency will professionally evaluate each piece of jade and give the corresponding second jade quality target value. For example, if a piece of Huanglong jade is evaluated as "high quality", this "high quality" is the second jade quality target value corresponding to the preset jade appearance data of the Huanglong jade. Each piece of jade has such a quality target value evaluated based on professional standards, which provides a reference for subsequent model training. Taking Huanglong jade as an example, the server imports the relevant data of Huanglong jade in the second preset jade appearance data set into the preset enhanced feature extraction path corresponding to Huanglong jade. The preset feature distiller first screens and refines the appearance data of Huanglong jade, extracts key information such as unique yellow hue and stone pattern distribution, and obtains the preset appearance distillation information. Then, this information enters the second feature focusing unit to further highlight the characteristics of Huanglong jade such as color saturation and texture clarity, and generate a second focused feature map. The atlas enters the first preset quality association subunit to calibrate the feature and obtain the second jade quality association feature. The feature is recombined with the preset appearance distillation information in the second feature recombiner to form the third intermediate feature data. The third intermediate feature data is then calibrated by the second preset quality association subunit to obtain the third jade quality association feature. Finally, the third jade quality association feature, the second jade quality association feature and the preset appearance distillation information are aggregated at the second feature aggregation node to obtain the fifth intermediate feature data corresponding to Huanglongyu. Other jade types, such as E-land jasper, also obtain their respective fifth intermediate feature data according to a similar process. The server inputs the fifth intermediate feature data of Huanglongyu into the preset quality grading path corresponding to Huanglongyu. Based on the internal algorithm, the path comprehensively analyzes the color, texture, transparency and other characteristics in the fifth intermediate feature data of Huanglongyu, infers the quality grade of Huanglongyu, and obtains the seventh jade quality inference value, such as "middle-upper quality". Similarly, for E-land jasper, the server imports its fifth intermediate feature data into the corresponding preset quality grading path, and combines the characteristics of jasper, such as color uniformity, black spot distribution, etc., to obtain the seventh jade quality inference value of E-land jasper, such as "medium quality". For Huanglong Jade, the server compares its seventh jade quality inference value "medium-high quality" with the second jade quality target value "high quality". The difference between the two is calculated through a specific algorithm to determine the sub-item quality error of Huanglong Jade, which reflects the degree of deviation between the model's inference of Huanglong Jade quality and the actual quality. Similarly, for E-site Jade, the seventh jade quality inference value "medium quality" is compared with the corresponding second jade quality target value to obtain the sub-item quality error of E-site Jade, which is used to measure the accuracy of the model's inference of different jade qualities.Taking the target jade type combination of Huanglong Jade and E-land Jasper as an example, the server trains the preset jade quality association units of Huanglong Jade and E-land Jasper according to their corresponding feature difference constraints. At the same time, referring to the sub-item quality errors of Huanglong Jade and E-land Jasper, the preset model core components corresponding to Huanglong Jade and E-land Jasper in the initial integrated model (i.e., other parts in the preset enhanced feature extraction path except the preset jade quality association unit, such as the preset feature distiller, the second feature focusing unit, etc.) are trained. For example, if it is found that the color feature extraction of Huanglong Jade is inaccurate, resulting in a large error in the sub-item quality, the components such as the preset feature distiller and the second feature focusing unit of Huanglong Jade are adjusted to optimize the extraction of color features. After multiple training and adjustments, when the accuracy of the model's inference of various jade qualities reaches a satisfactory level, the server obtains the target jade quality grading model.

[0069] In the embodiment of the present invention, the target jade quality grade corresponding to the jade appearance data is determined according to the target jade appearance features corresponding to each jade type and the target quality grading path corresponding to each jade type, which can be implemented through the following examples.

[0070] Importing the target jade appearance features corresponding to each jade type into the target quality grading path corresponding to each jade type to perform quality inference, and obtaining a first jade quality inference value corresponding to each jade type;

[0071] According to the first jade quality inference values ​​respectively configured for the plurality of jade types, a target jade quality grade corresponding to the jade appearance data is determined.

[0072] In an embodiment of the present invention, illustratively, the server stores target enhanced feature extraction paths and target quality grading paths corresponding to a variety of jade types. For the jade type of jadeite, the corresponding target jade appearance features have been extracted from the jade appearance data. These features include key information such as the brightness of the jade color, transparency, and fineness of the texture. The server imports these target jade appearance features of jadeite into the target quality grading path corresponding to jadeite. The target quality grading path is like an expert system specifically for jadeite quality assessment, with a series of complex algorithms and rules preset inside. It will conduct a comprehensive and in-depth analysis of the input target jade appearance features. For example, according to the quality assessment standards of the jadeite industry, jadeite with bright and uniform color, high transparency, and fine and flawless texture is usually of higher quality. The target quality grading path quantitatively evaluates and comprehensively considers the various features of jade based on these standards, and finally obtains the first jade quality inference value corresponding to jadeite, such as inferring that if the jade is jadeite, its quality is "medium and high-end". Similarly, for Hetian jade, the server imports the corresponding target jade appearance features, such as the whiteness, warmth, and density of the jade, into the target quality grading path corresponding to Hetian jade. This path gives a higher quality evaluation to Hetian jade with whiteness reaching mutton-fat white, excellent warmth, and dense texture according to the quality assessment rules of Hetian jade, thereby obtaining the first jade quality inference value corresponding to Hetian jade, which is assumed to be "superior quality". For other types of jade such as Xiuyan jade and Dushan jade, the server also follows this process to import their respective target jade appearance features into the corresponding target quality grading path, and obtain the corresponding first jade quality inference value. The server has collected multiple first jade quality inference values ​​for this piece of jade under different types of jade, such as "medium-high-end" for jadeite, "superior quality" for Hetian jade, and "medium quality" for Xiuyan jade. Next, the server needs to combine these inference values ​​to determine the target jade quality grade that the appearance data of this piece of jade ultimately corresponds to. The server will first consider factors such as the general value and recognition of different jade types in the market, and assign a certain weight to the inferred value of each jade type. For example, jadeite and Hetian jade have a higher status in the jade market, and their inferred value weights are relatively large; while Xiuyan jade, Dushan jade, etc. have relatively small weights. Then, the server multiplies the first jade quality inference value of each jade type by its corresponding weight through a specific algorithm and adds them together to obtain a comprehensive value. Finally, the server maps this comprehensive value to the corresponding target jade quality grade according to the pre-set quality grade classification standard. For example, if the comprehensive value falls within a certain range and the corresponding quality grade is "high quality", then the server determines that the target jade quality grade corresponding to the appearance data of this piece of jade is "high quality". In this way, by comprehensively considering the inference results of multiple jade types, the server can more comprehensively and accurately assess the quality grade of jade.

[0073] In the embodiments of the present invention, the following implementation modes are also provided.

[0074] Importing the jade appearance data into the target dynamic weight allocator in the target jade quality grading model for weight optimization allocation to obtain a target feature contribution parameter;

[0075] The step of determining the target jade quality grade corresponding to the jade appearance data according to the first jade quality inference values ​​respectively configured for the plurality of jade types comprises:

[0076] According to the target feature contribution parameter, an adaptive synthesis operation is performed on the first jade quality inference values ​​respectively configured for the multiple jade types to obtain a second jade quality inference value corresponding to the jade appearance data;

[0077] The target jade quality grade is determined according to the second jade quality inference value.

[0078] In an embodiment of the present invention, exemplarily, the server inputs the appearance data of the jade into the target dynamic weight allocator in the target jade quality grading model. This allocator is like an intelligent analysis master, which will deeply analyze the degree of correlation between each feature in the jade appearance data and the quality of different jade types. For example, the unique green color presented by the appearance of jade is highly similar to the emerald green feature of jadeite after analysis, which has a great influence on whether it is high-quality jadeite; and the fineness of the texture of jade not only plays an important role in judging whether it is high-quality Hetian jade, but also has a certain influence on the quality judgment of other types of jade, but the influence is relatively small. Based on this analysis of the correlation between features and the quality of jade types, the target dynamic weight allocator will assign different weights to each feature for different types of jade, so as to obtain the target feature contribution parameter. For example, for jadeite, the color feature weight may be set to 0.6, the texture feature weight to 0.3, and the transparency weight to 0.1; for Hetian jade, the whiteness weight is 0.5, the warmth weight is 0.3, and the texture fineness weight is 0.2, etc. These weight values ​​are the specific embodiment of the target feature contribution parameters, which reflect the relative importance of each feature in the quality assessment of different jade types. The server obtains the first jade quality inference value for this jade under multiple jade types, such as jadeite is inferred as "medium-high", Hetian jade is inferred as "superior quality", Xiuyan jade is inferred as "medium quality", etc. At the same time, the server has obtained the target feature contribution parameters. Next, the server performs adaptive synthesis operations on the first jade quality inference value based on these parameters. Taking jadeite and Hetian jade as examples, since the color feature weight of jadeite is relatively high, it is assumed that the value corresponding to its first jade quality inference value "medium-high" is 80 (setting a value is convenient for calculation and explanation), and the whiteness weight of Hetian jade is relatively high, and the value corresponding to its first jade quality inference value "superior quality" is 90. The server calculates according to the target feature contribution parameter. The weighted value of jadeite is \(80×0.6=48\), the weighted value of Hetian jade is \(90×0.5=45\), and then the weighted values ​​of all jade types are added together. At the same time, the weighted values ​​of other jade types such as Xiuyan jade are considered to obtain a comprehensive weighted value. Then, this comprehensive weighted value is converted into a new value through a specific algorithm. This new value is the second jade quality inference value. For example, after a series of calculations and conversions, the value corresponding to the obtained second jade quality inference value is 85 (this value is only an example). The server pre-sets a set of clear quality grade classification standards. According to this standard, different value ranges correspond to different quality grades. For example, a value of 90-100 corresponds to "special quality", 80-89 corresponds to "high quality", 70-79 corresponds to "intermediate quality", etc.The server compares the second jade quality inference value (e.g., value 85) with these standards and finds that the value falls within the range of 80-89, thus determining that the target jade quality level corresponding to the appearance data of this jade is "high quality". In this way, the server uses the target feature contribution parameter to synthesize the inference values ​​of multiple jade types, thereby more accurately determining the quality level of the jade.

[0079] In an embodiment of the present invention, the target feature extractor corresponding to each jade type includes a third feature focusing unit, a third feature recombiner, a target jade quality association unit and a third feature aggregation node; the target feature extractor that imports the target appearance distillation information corresponding to each jade type into the target enhanced feature extraction path corresponding to each jade type performs feature extraction to obtain the target jade appearance features corresponding to each jade type, which can be implemented through the following examples.

[0080] Importing the target appearance distillation information corresponding to each jade type into the third feature focusing unit corresponding to each jade type to perform feature focusing operation, and obtaining the third focusing feature map corresponding to each jade type;

[0081] Importing the target appearance distillation information corresponding to each jade type and the third focused feature map corresponding to each jade type into the third feature recombiner corresponding to each jade type to perform feature recombination operation, so as to obtain the first target quality key feature corresponding to each jade type;

[0082] Importing the target appearance distillation information corresponding to each jade type and the third focus feature map corresponding to each jade type into the target jade quality association unit corresponding to each jade type to perform feature calibration calculation, so as to obtain the fourth jade quality association feature corresponding to each jade type;

[0083] The fourth jade quality-related feature corresponding to each jade type, the first target quality key feature corresponding to each jade type, the target appearance distillation information corresponding to each jade type and the third focused feature map corresponding to each jade type are imported into the third feature aggregation node corresponding to each jade type to perform feature aggregation operation to obtain the target jade appearance feature corresponding to each jade type.

[0084] In the embodiment of the present invention, illustratively, taking jade as an example, the server has obtained the target appearance distillation information corresponding to jade, which is the key features after being refined by the target feature distiller, such as the key information of the color, transparency, crystal structure, etc. of jade. The server inputs the target appearance distillation information of jade into the third feature focusing unit corresponding to jade. The third feature focusing unit is like a high-precision microscope, which will further focus and refine the input target appearance distillation information. It will focus on the core features such as the hue, saturation and distribution uniformity of the color of jade, the specific value and level change of transparency, and the arrangement of the crystal structure. By focusing and strengthening these features, they are presented in the form of a spectrum to generate the third focused feature spectrum of jade. This spectrum can more clearly and intuitively show the key features closely related to the quality of jade. Similarly, for Hetian jade, the server inputs its target appearance distillation information into the third feature focusing unit corresponding to Hetian jade, highlighting the whiteness, warm texture, and the direction and density of the texture of Hetian jade, forming the third focused feature spectrum of Hetian jade. The server inputs the target appearance distillation information of jade and the third focused feature spectrum into the third feature recombiner corresponding to jade. The third feature reorganizer is like a skilled jigsaw puzzler. It will reorganize the two sets of information according to preset rules and algorithms. It will take into account the relationship and importance of each feature in the jade quality assessment, such as color has a greater impact on quality, followed by crystal structure. The first target quality key feature of jade is obtained by rearranging and combining the features of jade such as color, transparency, and crystal structure according to certain weights and logic. For example, it may combine the bright and uniform color features of jade with the tightly ordered crystal structure features to form a key feature combination that better represents high-quality jade. For Hetian jade, the third feature reorganizer will combine the characteristics of Hetian jade to reasonably reorganize the whiteness, warmth and texture features to obtain the first target quality key feature of Hetian jade. The server inputs the target appearance distillation information of jade and the third focus feature map into the target jade quality association unit corresponding to jade. This unit is like a strict calibration expert. It performs calibration operations on the input features based on the industry standards of jade quality and a large amount of sample data. It checks whether the color characteristics of jade are consistent with the known high-quality jade color standards, whether the transparency meets the requirements of the corresponding quality grade, and whether the crystal structure meets the normal physical properties. By calibrating and adjusting these features, more accurate and reliable features directly related to jade quality are obtained, namely the fourth jade quality-related features. For example, if the color feature of jade is found to deviate slightly from the standard in some aspects, the target jade quality-related unit will fine-tune it to make it more consistent with the color feature of high-quality jade.Similarly, for Hetian jade, the target jade quality association unit will calibrate its whiteness, warmth and texture characteristics according to the quality standard of Hetian jade, and obtain the fourth jade quality association feature of Hetian jade. The server inputs the fourth jade quality association feature of jadeite, the first target quality key feature, the target appearance distillation information and the third focus feature map into the third feature aggregation node corresponding to jadeite. The third feature aggregation node is like a powerful fusion center, which will perform comprehensive and in-depth aggregation operations on the feature information obtained at different stages. It will comprehensively consider the characteristics of various aspects of jadeite, and fuse the quality-related features accurately calibrated in the fourth jade quality association feature, the key feature combination that has been reasonably reorganized in the first target quality key feature, and the target appearance distillation information and the basic key features contained in the third focus feature map. Through complex algorithms, these features are organically integrated to form a comprehensive, comprehensive and accurate target jade appearance feature that can accurately reflect the quality of jadeite. Similarly, for Hetian jade, the third feature aggregation node will also aggregate the corresponding information to obtain the target jade appearance features of Hetian jade. These target jade appearance features will provide a key basis for the subsequent accurate assessment of jade quality grades.

[0085] In an embodiment of the present invention, the target feature extractor corresponding to each jade type includes a fourth feature focusing unit, a fourth feature recombiner, a target jade quality association unit and a fourth feature aggregation node, and the target jade quality association unit corresponding to each jade type includes a first target quality association subunit and a second target quality association subunit; the target feature extractor that imports the target appearance distillation information corresponding to each jade type into the target enhanced feature extraction path corresponding to each jade type performs feature extraction to obtain the target jade appearance features corresponding to each jade type, which can be implemented through the following examples.

[0086] Importing the target appearance distillation information corresponding to each jade type into the fourth feature focusing unit corresponding to each jade type to perform feature focusing operation, so as to obtain the fourth focusing feature map corresponding to each jade type;

[0087] Importing the fourth focused feature map corresponding to each jade type into the first target quality association subunit corresponding to each jade type to perform feature calibration calculation, so as to obtain the fifth jade quality association feature corresponding to each jade type;

[0088] Importing the fifth jade quality-related feature corresponding to each jade type and the target appearance distillation information corresponding to each jade type into the fourth feature recombiner corresponding to each jade type to perform feature recombining operation, so as to obtain the second target quality key feature corresponding to each jade type;

[0089] Importing the second target quality key feature corresponding to each jade type into the second target quality association subunit corresponding to each jade type to perform feature calibration operation, so as to obtain the sixth jade quality association feature corresponding to each jade type;

[0090] The sixth jade quality-related feature corresponding to each jade type, the fifth jade quality-related feature corresponding to each jade type, and the target appearance distillation information corresponding to each jade type are imported into the fourth feature aggregation node corresponding to each jade type to perform feature aggregation operation to obtain the target jade appearance feature corresponding to each jade type.

[0091] In the embodiment of the present invention, for example, taking Xiuyu as an example, the server has obtained the target appearance distillation information of Xiuyu after being refined by the target feature distiller, which includes the key features of Xiuyu such as color, texture, and texture. The server inputs the target appearance distillation information of these Xiuyu jades into the fourth feature focusing unit corresponding to Xiuyu jade. This unit is like a precise magnifying glass, which deeply analyzes the target appearance distillation information of Xiuyu jade, focusing on the core features closely related to quality such as the unique color tone of Xiuyu jade, the fineness and distribution law of texture, and the tightness of texture. By strengthening and sorting these features, they are presented in the form of a map to generate the fourth focused feature map of Xiuyu jade. This map can more clearly and intuitively highlight the key information related to the quality of Xiuyu jade. Similarly, for Dushan jade, the server inputs its target appearance distillation information into the corresponding fourth feature focusing unit, focusing on the richness of the color of Dushan jade, the characteristics of the block structure, and other features to form the fourth focused feature map of Dushan jade. The server inputs the fourth focused feature map of Xiuyu jade into the first target quality association subunit corresponding to Xiuyu jade. This subunit is similar to a professional calibration instrument. It calibrates the characteristics of Xiuyu jade according to the industry standards of Xiuyu jade quality and a large amount of Xiuyu jade sample data. For example, referring to the standard of high-quality Xiuyu jade with bright and uniform color and fine and coherent texture, the color and texture characteristics in the fourth focused feature map are calibrated to remove possible deviations or inaccurate information, and obtain more accurate characteristics associated with the quality of Xiuyu jade, that is, the fifth jade quality-related characteristics. For Dushan jade, the first target quality-related subunit calibrates the color, structure and other characteristics in its fourth focused feature map according to the quality standard of Dushan jade, and obtains the fifth jade quality-related characteristics of Dushan jade. The server inputs the fifth jade quality-related characteristics of Xiuyu jade and the target appearance distillation information into the fourth feature recombiner corresponding to Xiuyu jade. This recombiner is like a skilled jigsaw puzzle master. According to the preset rules and algorithms, combined with the importance and mutual relationship of each feature in the quality assessment of Xiuyu jade, this information is recombined. For example, the calibrated color and texture features are arranged and combined with the texture and other features in the target appearance distillation information with specific weights and logic to obtain the second target quality key feature of Xiuyu jade, so that it can more accurately reflect the quality of Xiuyu jade. Similarly, for Dushan jade, the fourth feature reorganizer will reorganize its fifth jade quality-related feature and the target appearance distillation information according to the characteristics of Dushan jade to obtain the second target quality key feature of Dushan jade. The server inputs the second target quality key feature of Xiuyu jade into the second target quality-related subunit corresponding to Xiuyu jade. This subunit performs a calibration operation on the feature again, and performs a second calibration on the Xiuyu jade feature from another angle or a more refined level according to the Xiuyu jade quality standard. For example, the coordination and matching degree between the texture of Xiuyu jade and the color and texture are further calibrated, so that the correlation between the obtained feature and the actual quality is more accurate, thereby obtaining the sixth jade quality-related feature of Xiuyu jade.Similarly, for Dushan jade, the second target quality association subunit performs a secondary calibration on its second target quality key feature to obtain the sixth jade quality association feature of Dushan jade. The server inputs the sixth jade quality association feature, the fifth jade quality association feature and the target appearance distillation information of Xiuyu jade into the fourth feature aggregation node corresponding to Xiuyu jade. This node is like an efficient fusion center. Through complex aggregation operations, it comprehensively integrates various quality-related features obtained from different stages of Xiuyu jade. It comprehensively considers various aspects of Xiuyu jade, such as color, texture, and texture, and organically integrates the features that have been calibrated twice and reorganized once to form a more comprehensive, representative and accurate target jade appearance feature that can accurately reflect the quality of Xiuyu jade. Similarly, for Dushan jade, the fourth feature aggregation node aggregates the corresponding information to obtain the target jade appearance feature of Dushan jade. These target jade appearance features will provide a key basis for the subsequent accurate assessment of jade quality grades.

[0092] 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 jade quality intelligent assessment method based on deep learning enhanced feature extraction. Figure 2 As shown, Figure 2 The computer device 100 is a block diagram of a structure of a computer device 100 provided in 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 memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other.

[0093] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. An intelligent jade quality assessment method based on deep learning enhanced feature extraction, characterized in that: include: Get jade appearance data; Importing the jade appearance data into the target feature distiller in the target enhancement feature extraction path corresponding to each jade type to perform feature distillation operation, so as to obtain the target appearance distillation information corresponding to each jade type; The target appearance distillation information corresponding to each jade type is imported into the target feature extractor in the target enhanced feature extraction path corresponding to each jade type for feature extraction, so as to obtain the target jade appearance features corresponding to each jade type; the target jade quality grading model is obtained by tuning the preset jade quality association unit in the preset enhanced feature extraction path according to the feature difference constraint, the feature difference constraint is used to quantify the correlation of the target jade quality association features, and the target jade quality association features are the jade quality association features generated by the preset jade quality association units corresponding to different jade types; The target jade quality grade corresponding to the jade appearance data is determined according to the target jade appearance features corresponding to each jade type and the target quality grading path corresponding to each jade type.

2. The method according to claim 1, characterized in that: The target jade quality grading model is trained through the following process, including: Obtaining a first preset jade appearance data set; Importing the first preset jade appearance data set into the preset enhanced feature extraction paths respectively configured for the multiple jade types in the initial integrated model to perform feature extraction operations, so as to obtain preset jade quality-related features corresponding to each jade type; Determine the characteristic difference constraint corresponding to the target jade type combination according to the preset jade quality associated characteristics corresponding to the target jade type combination; the target jade type combination is any two types of the multiple jade types; According to the characteristic difference constraints corresponding to the target jade type combination, the preset jade quality association units corresponding to the target jade type combination are trained to obtain the target jade quality grading model.

3. The method according to claim 2, characterized in that The preset enhanced feature extraction path corresponding to each jade type includes a preset feature distiller and a preset feature extractor, and the preset feature extractor includes a first feature focusing unit, a first feature recombiner, a preset jade quality association unit and a first feature aggregation node; The first preset jade appearance data set is imported into the preset enhanced feature extraction paths respectively configured for the multiple jade types in the initial integrated model to perform feature extraction operations to obtain the preset jade quality associated features corresponding to each jade type, including: Importing the first preset jade appearance data set into a preset feature distiller corresponding to each jade type to perform a feature distillation operation, to obtain preset appearance distillation information corresponding to each jade type; Importing the preset appearance distillation information corresponding to each jade type into the first feature focusing unit corresponding to each jade type to perform feature focusing operation, so as to obtain the first focusing feature map corresponding to each jade type; Importing the first focused feature map corresponding to each jade type and the preset appearance distillation information corresponding to each jade type into the first feature recombiner corresponding to each jade type to perform feature recombination operation, so as to obtain the first intermediate feature data corresponding to each jade type; Importing the first focused feature map corresponding to each jade type and the preset appearance distillation information corresponding to each jade type into the preset jade quality association unit corresponding to each jade type to perform feature calibration calculation, so as to obtain the first jade quality association feature corresponding to each jade type; The first intermediate feature data corresponding to each jade type, the first jade quality-related feature corresponding to each jade type, the first focused feature map corresponding to each jade type and the preset appearance distillation information corresponding to each jade type are imported into the first feature aggregation node corresponding to each jade type to perform feature aggregation operation to obtain the second intermediate feature data corresponding to each jade type.

4. The method according to claim 3, characterized in that The method further comprises: Obtaining a first jade quality target value corresponding to each preset jade appearance data in the first preset jade appearance data set; Importing the second intermediate feature data corresponding to each jade type into the preset quality grading path corresponding to each jade type in the initial integrated model to perform quality inference, and obtaining a third jade quality inference value corresponding to each jade type; Importing the first preset jade appearance data set into the preset dynamic weight allocator in the initial integrated model for weight optimization allocation, and obtaining a preset feature contribution parameter corresponding to each preset jade appearance data in the first preset jade appearance data set; According to the preset feature contribution parameter, an adaptive synthesis operation is performed on the third jade quality inference values ​​respectively configured for the plurality of jade types to obtain a fourth jade quality inference value corresponding to each preset jade appearance data in the first preset jade appearance data set; Determining a first fusion consistency error according to the fourth jade quality inference value and the first jade quality target value; The method of training the preset jade quality association unit corresponding to the target jade type combination according to the characteristic difference constraint corresponding to the target jade type combination to obtain the target jade quality grading model includes: According to the feature difference constraint corresponding to the target jade type combination, the preset jade quality association unit corresponding to the target jade type combination is trained, and according to the first fusion consistency error, the preset dynamic weight allocator is trained to obtain the target jade quality grading model.

5. The method according to claim 2, characterized in that: The preset enhanced feature extraction path corresponding to each jade type includes a preset feature distiller and a preset feature extractor, the preset feature extractor includes a second feature focusing unit, a second feature recombiner, a preset jade quality association unit and a second feature aggregation node, and the preset jade quality association unit includes a first preset quality association subunit and a second preset quality association subunit; the first preset jade appearance data set is imported into the preset enhanced feature extraction path respectively configured for the multiple jade types in the initial integrated model to perform feature extraction operations, and the preset jade quality association features corresponding to each jade type are obtained, including: Importing the first preset jade appearance data set into a preset feature distiller corresponding to each jade type to perform a feature distillation operation, to obtain preset appearance distillation information corresponding to each jade type; Importing the preset appearance distillation information corresponding to each jade type into the second feature focusing unit corresponding to each jade type to perform feature focusing operation, so as to obtain the second focusing feature map corresponding to each jade type; Importing the second focused feature map corresponding to each jade type into the first preset quality association subunit corresponding to each jade type to perform feature calibration calculation, so as to obtain the second jade quality association feature corresponding to each jade type; Importing the second jade quality-related feature corresponding to each jade type and the preset appearance distillation information corresponding to each jade type into the second feature recombiner corresponding to each jade type to perform feature recombining operation, so as to obtain third intermediate feature data corresponding to each jade type; Importing the third intermediate feature data corresponding to each jade type into the second preset quality association subunit corresponding to each jade type to perform feature calibration calculation, so as to obtain the third jade quality association feature corresponding to each jade type; The third jade quality-related feature corresponding to each jade type, the second jade quality-related feature corresponding to each jade type, and the preset appearance distillation information corresponding to each jade type are imported into the second feature aggregation node corresponding to each jade type to perform feature aggregation operation to obtain the fourth intermediate feature data corresponding to each jade type.

6. The method according to claim 5, characterized in that The method further comprises: Obtaining a first jade quality target value corresponding to each preset jade appearance data in the first preset jade appearance data set; Importing the fourth intermediate feature data corresponding to each jade type into the preset quality grading path corresponding to each jade type in the initial integrated model to perform quality inference, and obtaining a fifth jade quality inference value corresponding to each jade type; Importing the first preset jade appearance data set into the preset dynamic weight allocator in the initial integrated model for weight optimization allocation, and obtaining a preset feature contribution parameter corresponding to each preset jade appearance data in the first preset jade appearance data set; According to the preset feature contribution parameter, an adaptive synthesis operation is performed on the fifth jade quality inference values ​​respectively configured for the plurality of jade types to obtain a sixth jade quality inference value corresponding to each preset jade appearance data in the first preset jade appearance data set; Determining a second fusion consistency error according to the sixth jade quality inference value and the first jade quality target value; The method of training the preset jade quality association unit corresponding to the target jade type combination according to the characteristic difference constraint corresponding to the target jade type combination to obtain the target jade quality grading model includes: According to the feature difference constraint corresponding to the target jade type combination, the preset jade quality association unit corresponding to the target jade type combination is trained, and according to the second fusion consistency error, the preset dynamic weight allocator is trained to obtain the target jade quality grading model.

7. The method according to claim 2, characterized in that: The method further comprises: Acquire a second preset jade appearance data set and a second jade quality target value corresponding to each preset jade appearance data in the second preset jade appearance data set; Importing the second preset jade appearance data set into the preset enhanced feature extraction paths respectively configured for the multiple jade types in the initial integrated model to perform feature extraction operations, and obtaining fifth intermediate feature data corresponding to each jade type; Importing the fifth intermediate characteristic data corresponding to each jade type into the preset quality grading path corresponding to each jade type to perform quality inference, and obtaining the seventh jade quality inference value corresponding to each jade type; Determine the sub-item quality error corresponding to each jade type according to the seventh jade quality inference value corresponding to each jade type and the second jade quality target value; The method of training the preset jade quality association unit corresponding to the target jade type combination according to the characteristic difference constraint corresponding to the target jade type combination to obtain the target jade quality grading model includes: According to the feature difference constraints corresponding to the target jade type combination, the preset jade quality association units corresponding to the target jade type combination are trained, and according to the sub-item quality errors corresponding to each jade type, the preset model core components corresponding to each jade type in the initial integrated model are trained to obtain the target jade quality grading model; the preset model core components corresponding to each jade type are the feature processing architectures in the preset enhanced feature extraction paths corresponding to each jade type except the preset jade quality association units corresponding to each jade type.

8. The method according to claim 1, characterized in that Determining the target jade quality grade corresponding to the jade appearance data according to the target jade appearance features corresponding to each jade type and the target quality grading path corresponding to each jade type includes: Importing the target jade appearance features corresponding to each jade type into the target quality grading path corresponding to each jade type to perform quality inference, and obtaining a first jade quality inference value corresponding to each jade type; According to the first jade quality inference values ​​respectively configured for the plurality of jade types, a target jade quality grade corresponding to the jade appearance data is determined.

9. The method according to claim 8, characterized in that The method further comprises: Importing the jade appearance data into the target dynamic weight allocator in the target jade quality grading model for weight optimization allocation to obtain a target feature contribution parameter; The step of determining the target jade quality grade corresponding to the jade appearance data according to the first jade quality inference values ​​respectively configured for the plurality of jade types comprises: According to the target feature contribution parameter, an adaptive synthesis operation is performed on the first jade quality inference values ​​respectively configured for the plurality of jade types to obtain a second jade quality inference value corresponding to the jade appearance data; The target jade quality grade is determined according to the second jade quality inference value.

10. 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 9.

Images