Image recognition and deep learning combined modeling-based antique coin appearance state comprehensive evaluation method and system

Abstract

The invention discloses an antique coin appearance state comprehensive evaluation method and system based on image recognition and deep learning joint modeling, and relates to the field of artificial intelligence, and the method comprises the steps: firstly obtaining an antique coin evaluation model containing a plurality of network branches, carrying out the evaluation dimension distribution of each branch, and determining a training direction; configuring an evaluation training framework according to the training direction, wherein the evaluation training framework comprises evaluation units corresponding to the branches; during training, the evaluation unit is called to execute the evaluation dimension training process according to the corresponding training direction, and model training is completed. And importing a to-be-evaluated image into the trained model to obtain an appearance state comprehensive evaluation result. According to the method, the accuracy and efficiency of antique coin appearance state evaluation are improved through scientific configuration of the model and the training process.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for comprehensively evaluating the appearance status of antique coins based on joint modeling of image recognition and deep learning. Background Art

[0002] With the development of the antique market, the demand for accurate evaluation of the appearance of antique coins is growing. Traditional evaluation methods for the appearance of antique coins mostly rely on manual experience, which is highly subjective and inefficient. Although there have been some evaluation attempts based on image recognition or deep learning, there are problems such as a single evaluation dimension and low model training efficiency, making it difficult to comprehensively and accurately evaluate the appearance of antique coins. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for comprehensively evaluating the appearance status of antique coins based on joint modeling of image recognition and deep learning.

[0004] In a first aspect, an embodiment of the present invention provides a method for comprehensively evaluating the appearance status of antique coins based on joint modeling of image recognition and deep learning, comprising:

[0005] Acquire an untrained antique coin evaluation model, wherein the antique coin evaluation model includes a plurality of antique coin evaluation network branches;

[0006] The evaluation dimension allocation process is performed on each antique coin evaluation network branch respectively to obtain the antique coin training direction of each antique coin evaluation network branch; the antique coin training direction of each antique coin evaluation network branch is used to characterize the evaluation dimension of each antique coin evaluation network branch in the training stage and the training process of the evaluation dimension;

[0007] Based on the antique coin training directions of the multiple antique coin evaluation network branches, an evaluation training framework of the antique coin evaluation model is configured, wherein the evaluation training framework includes multiple evaluation units; one evaluation unit corresponds to one antique coin evaluation network branch, and each evaluation unit is configured based on the antique coin training direction of the corresponding antique coin evaluation network branch;

[0008] When executing the training process for the antique coin evaluation model, calling multiple evaluation units in the evaluation training framework to respectively execute the evaluation dimension training process of the corresponding antique coin evaluation network branch based on the training process corresponding to the antique coin training direction of the corresponding antique coin evaluation network branch, so as to obtain the antique coin evaluation model that has completed the training;

[0009] The antique coin image to be evaluated is imported into the antique coin evaluation model that has completed the training, and a comprehensive evaluation result of the appearance state of the antique coin image to be evaluated is obtained.

[0010] 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.

[0011] Compared with the prior art, the beneficial effects provided by the present invention include: adopting a method and system for comprehensive evaluation of the appearance status of antique coins based on joint modeling of image recognition and deep learning disclosed by the present invention, by obtaining an antique coin evaluation model containing multiple network branches, assigning evaluation dimensions to each branch, and determining the training direction. Configure the evaluation training framework according to the training direction, which contains evaluation units corresponding to each branch. During training, call the evaluation unit to execute the evaluation dimension training process according to the corresponding training direction to complete the model training. Import the image to be evaluated into the trained model to obtain the comprehensive evaluation result of the appearance status. This method improves the accuracy and efficiency of the appearance status evaluation of antique coins by scientifically configuring the model and training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] 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.

[0013] Figure 1 A schematic diagram of the steps of a method for comprehensively evaluating the appearance of antique coins based on joint modeling of image recognition and deep learning provided by an embodiment of the present invention;

[0014] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] 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.

[0016] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0017] In order to solve the technical problems in the aforementioned background technology, Figure 1A flow chart of a method for comprehensively evaluating the appearance status of antique coins based on joint modeling of image recognition and deep learning provided in an embodiment of the present disclosure is provided below. The method for comprehensively evaluating the appearance status of antique coins based on joint modeling of image recognition and deep learning is introduced in detail.

[0018] Step S201, obtaining an untrained antique coin evaluation model, wherein the antique coin evaluation model includes a plurality of antique coin evaluation network branches;

[0019] Step S202, respectively performing evaluation dimension allocation processing on each antique coin evaluation network branch to obtain the antique coin training direction of each antique coin evaluation network branch; the antique coin training direction of each antique coin evaluation network branch is used to characterize the evaluation dimension of each antique coin evaluation network branch in the training stage and the training process of the evaluation dimension;

[0020] Step S203, configuring an evaluation training framework of the antique coin evaluation model based on the antique coin training directions of the multiple antique coin evaluation network branches, wherein the evaluation training framework includes multiple evaluation units; one evaluation unit corresponds to one antique coin evaluation network branch, and each evaluation unit is configured based on the antique coin training direction of the corresponding antique coin evaluation network branch;

[0021] Step S204, when executing the training process for the antique coin evaluation model, calling multiple evaluation units in the evaluation training framework to respectively execute the evaluation dimension training process of the corresponding antique coin evaluation network branch based on the training process corresponding to the antique coin training direction of the corresponding antique coin evaluation network branch, so as to obtain a trained antique coin evaluation model;

[0022] Step S205, importing the antique coin image to be evaluated into the trained antique coin evaluation model to obtain a comprehensive evaluation result of the appearance state of the antique coin image to be evaluated.

[0023] In an embodiment of the present invention, illustratively, the server obtains an antique coin evaluation model from a specific storage location or data source. For example, the model may be stored in a local database of the server, or downloaded from a remote model storage server. This antique coin evaluation model includes multiple antique coin evaluation network branches, each of which may focus on different aspects of the evaluation of the appearance status of antique coins, such as some branches may be responsible for evaluating the degree of wear of coins, and some branches may be responsible for evaluating the clarity of patterns on coins. The server performs evaluation dimension allocation processing on each antique coin evaluation network branch respectively. Take one of the branches (the first antique coin evaluation network branch) as an example: Obtaining the evaluation dimension feature domain: The server obtains the evaluation dimension feature domain of the first antique coin evaluation network branch, and this feature domain contains multiple standard antique coin training directions pre-determined for the branch. For example, the standard antique coin training direction may have multiple directions such as overall coin wear evaluation, pattern detail evaluation, and text clarity evaluation. Determine the evaluation cycle: Determine the evaluation cycle based on the standard evaluation configuration parameters corresponding to each standard antique coin training direction in the evaluation dimension feature domain. Assume that the sample evaluation configuration parameters in the evaluation configuration parameters include evaluation granularity (for example, the antique coin image is divided into different levels of fineness for evaluation, such as fine to a single pattern element or coarse to the entire coin surface area) and sample division dimension (for example, the sample is divided according to different parts of the coin such as the front, back, and edge), and the architecture evaluation configuration parameters include training process segmentation dimension (such as dividing the training into different stages such as initial training and optimization training) and feature analysis dimension (such as analyzing from different feature angles such as color and shape). Taking the first standard antique coin training direction as an example, the server standardizes the first antique coin evaluation network branch into a network entity in the first model training environment, and uses the network entity to process a sample antique coin image to obtain a standard evaluation cycle in the first model training environment. Then, the evaluation granularity in the standard evaluation configuration parameters corresponding to the first standard antique coin training direction is obtained, multiplied by the standard evaluation cycle, and then the multiplication result is divided by the corresponding architecture evaluation configuration parameters to obtain the evaluation cycle under the standard training direction. Select the training direction: The server selects the standard antique coin training direction corresponding to the shortest evaluation cycle in the evaluation dimension feature domain as the antique coin training direction of the first antique coin evaluation network branch. In this way, the antique coin training direction is determined for each antique coin evaluation network branch, and the direction characterizes the evaluation dimension and training process in the training phase. Before determining the evaluation cycle, the server will also perform the following operations: obtain the standard processing load of multiple antique coin evaluation network branches in the same model training environment. For example, some branches have a larger load when processing complex pattern evaluations, while others have a smaller load when processing simple contour evaluations. Determine the target antique coin evaluation network branch with the largest standard processing load from multiple branches.If the target antique coin evaluation network branch is not the first antique coin evaluation network branch being processed, the server will adjust the various standard antique coin training directions in the evaluation dimension feature domain of the first antique coin evaluation network branch according to the antique coin training direction of the target antique coin evaluation network branch. For example, obtain the multiplication result between the evaluation granularity and the sample division dimension corresponding to the antique coin training direction of the target antique coin evaluation network branch, and use this result to divide the sample division dimension corresponding to the first standard antique coin training direction to adjust the evaluation granularity corresponding to the first standard antique coin training direction; or use the multiplication result to divide the evaluation granularity corresponding to the first standard antique coin training direction to adjust the sample division dimension corresponding to the first standard antique coin training direction. Construct an evaluation unit: The server constructs an evaluation unit corresponding to each branch based on the antique coin training direction of each antique coin evaluation network branch. For example, for the first antique coin evaluation network branch, the number of model training environments in the first evaluation unit is determined based on the sample division dimension and architecture evaluation configuration parameters in the evaluation configuration parameters corresponding to its training direction. Then, the environment configuration in the preset model training environment set is obtained, such as different hardware configurations, software versions and other environmental factors, and the standard evaluation cycle of the first antique coin evaluation network branch on the model training environment of each environment configuration is obtained. Then, based on the evaluation granularity corresponding to the training direction, the architecture evaluation configuration parameters and the standard evaluation cycle, the evaluation cycle on the model training environment of each environment configuration is determined, and the target environment configuration corresponding to the shortest evaluation cycle is found. The number of model training environments configured in the target environment is obtained from the model training environment set. When the number of model training environments configured in the target environment is greater than or equal to the number of model training environments in the first evaluation unit, the target environment configuration is used as the environment configuration of the model training environment in the first evaluation unit, and the first antique coin evaluation network branch is standardized in the model training environment set based on the training direction and these configurations, so as to configure the first evaluation unit. For each antique coin evaluation network branch, the corresponding evaluation unit is constructed in a similar manner. Establish a multimodal feature fusion architecture: Assume that the antique coin evaluation model also includes a second antique coin evaluation network branch connected to the first antique coin evaluation network branch, the first branch corresponds to the first evaluation unit, and the second branch corresponds to the second evaluation unit. The server groups the first evaluation unit and the second evaluation unit, obtains the common evaluation cardinality between the sample division dimension corresponding to the antique coin training direction of the first antique coin evaluation network branch and the sample division dimension corresponding to the antique coin training direction of the second antique coin evaluation network branch, and uses it as the number of evaluation unit pools, and groups the two evaluation units based on the number to obtain multiple evaluation unit pools, each evaluation unit pool includes at least one evaluation sub-unit obtained by the first evaluation unit and at least one evaluation sub-unit obtained by the second evaluation unit, and each evaluation sub-unit contains one or more model training environments.When establishing a multimodal feature fusion architecture, dynamic resource allocation is configured for each evaluation unit pool, such as allocating different proportions of computing resources according to the needs of different evaluation tasks, and determining the feature processing flow sequence between the first evaluation unit and the second evaluation unit according to the order of connection between the first antique coin evaluation network branch and the second antique coin evaluation network branch. In each evaluation sub-unit obtained by the evaluation unit with the previous feature processing flow sequence, a model training environment is selected as a feature acquisition node; in each evaluation sub-unit obtained by the evaluation unit with the later feature processing flow sequence, a model training environment is selected as a feature fusion node, and a cross-modal connection path between the feature acquisition node and the feature fusion node is established. In this way, an evaluation training framework for the antique coin evaluation model is generated. When the server executes the training process for the antique coin evaluation model, the training stage is divided into multiple training steps in the training link, and one training step corresponds to the evaluation dimension of an antique coin evaluation network branch. For example, the training link may include initial feature extraction, feature optimization and other steps, wherein the initial feature extraction step may correspond to the preliminary recognition evaluation dimension of a branch for coin patterns. When executing the training link, the server determines the target evaluation unit for executing the current training step. For example, when performing the preliminary recognition and evaluation of the coin pattern, the corresponding evaluation unit is determined to be the evaluation unit responsible for the evaluation dimension (assuming it is the first evaluation unit), and the training process corresponding to the antique coin training direction of the target evaluation unit based on its corresponding antique coin evaluation network branch (the first antique coin evaluation network branch) is called to execute the evaluation dimension training process of the target evaluation unit corresponding to the antique coin evaluation network branch. The evaluation dimension training processes of the corresponding branches are respectively executed by multiple evaluation units, and finally the antique coin evaluation model that has been trained is obtained. The server imports the image of the antique coin to be evaluated into the antique coin evaluation model that has been trained. The model analyzes and processes the image based on the ability obtained by training, and comprehensively evaluates the appearance of the antique coin from multiple evaluation dimensions, and finally obtains the comprehensive evaluation result of the appearance of the antique coin image to be evaluated, such as providing evaluation information such as the wear level of the coin and the pattern clarity score.

[0024] In the embodiment of the present invention, the antique coin training direction of each antique coin evaluation network branch is also used to characterize the evaluation configuration parameters of each antique coin evaluation network branch, and each antique coin evaluation network branch is represented as a first antique coin evaluation network branch;

[0025] The evaluation dimension allocation process is performed on the first antique coin evaluation network branch to obtain the antique coin training direction of the first antique coin evaluation network branch, which can be implemented through the following example.

[0026] Acquire an evaluation dimension feature domain of the first antique coin evaluation network branch, wherein the evaluation dimension feature domain of the first antique coin evaluation network branch includes a plurality of standard antique coin training directions predetermined for the first antique coin evaluation network branch;

[0027] Determining the evaluation cycle of the first antique coin evaluation network branch under each of the standard antique coin training directions based on the standard evaluation configuration parameters corresponding to each of the standard antique coin training directions in the evaluation dimension feature domain of the first antique coin evaluation network branch;

[0028] The standard antique coin training direction corresponding to the shortest evaluation period in the evaluation dimension feature domain of the first antique coin evaluation network branch is selected as the antique coin training direction of the first antique coin evaluation network branch;

[0029] Among them, the evaluation configuration parameters include sample evaluation configuration parameters and architecture evaluation configuration parameters; the sample evaluation configuration parameters include evaluation granularity and sample division dimension; the architecture evaluation configuration parameters include training phase process segmentation dimension and feature analysis dimension.

[0030] In an embodiment of the present invention, exemplarily, the server obtains the evaluation dimension feature domain of the first antique coin evaluation network branch from a pre-set storage area. For example, in a database that specifically stores parameters and data related to antique coin evaluation, there is a specific data table that records the relevant information of each evaluation network branch. For the first antique coin evaluation network branch, its evaluation dimension feature domain contains a plurality of predetermined standard antique coin training directions. Assume that these standard training directions are: evaluation from the perspective of the clarity of the coin's text, evaluation from the perspective of the integrity of the coin's pattern, and evaluation from the perspective of the degree of edge wear of the coin. In terms of sample evaluation configuration parameters: taking the standard antique coin training direction of evaluating from the perspective of the clarity of the coin's text as an example, in the corresponding sample evaluation configuration parameters, the evaluation granularity may be set to finely evaluate the text on the coin as a single character; the sample division dimension is set to use the text on the front and back of the coin as different sample subsets. In terms of architecture evaluation configuration parameters: the training process segmentation dimension may divide the training into data preprocessing, feature extraction, model training and evaluation optimization stages; the feature analysis dimension is set to analyze the stroke thickness, color contrast and other aspects of the text. Calculation evaluation cycle: The server standardizes the first antique coin evaluation network branch into a network entity in a given first model training environment. Next, the server selects a sample antique coin image and lets the network entity process it to obtain a standard evaluation cycle in the first model training environment, assuming it is 10 seconds. Then, the evaluation granularity (such as single character evaluation) and the standard evaluation cycle (10 seconds) are associated with a calculation (such as multiplication) to obtain a preliminary result. This preliminary result is then divided by the architecture evaluation configuration parameters (for example, divided by a certain influence coefficient brought by the training process segmentation dimension and the feature analysis dimension), and finally the evaluation cycle under the standard antique coin training direction is obtained, assuming that it is 8 seconds after calculation. In the same way, the server calculates the evaluation cycle under other standard antique coin training directions such as pattern integrity angle evaluation and edge wear angle evaluation. For example, the evaluation period under the pattern integrity angle evaluation is calculated to be 12 seconds, and the evaluation period under the edge wear angle evaluation is calculated to be 10 seconds. The server compares the evaluation periods calculated for each standard antique coin training direction. In the above example, 8 seconds (text clarity angle evaluation) < 10 seconds (edge ​​wear angle evaluation) < 12 seconds (pattern integrity angle evaluation), so the server selects the text clarity angle evaluation, a standard antique coin training direction, as the antique coin training direction for the first antique coin evaluation network branch. Through this process, the server determines the evaluation dimensions and corresponding training processes for the first antique coin evaluation network branch during the training phase, and the training direction also characterizes the corresponding evaluation configuration parameters.

[0031] In an embodiment of the present invention, the plurality of standard antique coin training directions include a first standard antique coin training direction; the evaluation configuration parameters include sample evaluation configuration parameters and architecture evaluation configuration parameters, and the sample evaluation configuration parameters include evaluation granularity;

[0032] Based on the standard evaluation configuration parameters corresponding to the first standard antique coin training direction, determining the evaluation period of the first antique coin evaluation network branch under the first standard antique coin training direction can be implemented through the following example.

[0033] In a first model training environment, the first antique coin evaluation network branch is standardized into a network entity of the first antique coin evaluation network branch, and the network entity of the first antique coin evaluation network branch is called to process a sample antique coin image to obtain a standard evaluation cycle of the first antique coin evaluation network branch in the first model training environment;

[0034] Obtaining a first multiplication result between an evaluation granularity in a standard evaluation configuration parameter corresponding to the first standard antique coin training direction and a standard evaluation period of the first antique coin evaluation network branch in the first model training environment;

[0035] A first division result between the first multiplication result and the architecture evaluation configuration parameter corresponding to the first standard antique coin training direction is used as an evaluation period of the first antique coin evaluation network branch under the first standard antique coin training direction.

[0036] In an embodiment of the present invention, exemplarily, the server runs in a specific first model training environment, which includes a predetermined hardware configuration (such as a certain model of CPU, GPU, a certain capacity of memory, etc.) and software settings (a specific version of a deep learning framework, an operating system, etc.). The server standardizes the first antique coin evaluation network branch to make it an executable network entity, just like converting a design blueprint into an actual executable machine. Next, the server randomly selects an image from a database storing sample antique coin images. For example, the database stores a large number of antique coin images of different ages and types, and the server selects a front image of a Qing Dynasty copper coin. Then, the server calls the standardized network entity of the first antique coin evaluation network branch to process this sample antique coin image. During the processing, the network entity analyzes the image according to its internal algorithm and structure. After the processing is completed, the server records the time spent on the entire processing process to obtain the standard evaluation cycle of the first antique coin evaluation network branch in the first model training environment, assuming that this cycle is 15 seconds. The server obtains the evaluation granularity in the standard evaluation configuration parameters corresponding to the first standard antique coin training direction. Assuming that the first standard antique coin training direction is to evaluate the detail of the coin pattern, its evaluation granularity is set to divide the pattern into the smallest element unit for evaluation, such as refining the dragon pattern on the coin to each dragon scale. The evaluation granularity can be represented by a quantitative value, assuming that it is 10 here (representing that the pattern is divided into 10 smallest evaluation units). The server multiplies the obtained evaluation granularity 10 with the standard evaluation cycle 15 seconds of the first antique coin evaluation network branch obtained above in the first model training environment, that is, 10×15=150, and obtains the first multiplication result 150. The server obtains the architecture evaluation configuration parameters corresponding to the first standard antique coin training direction. For example, the architecture evaluation configuration parameters include the training process segmentation dimension (dividing the training into four stages: data preparation, feature extraction, model training, and optimization adjustment) and the feature analysis dimension (analyzing from the three aspects of pattern shape, color, and texture). These factors can be combined to represent a coefficient, assuming that it is 5. The server divides the first multiplication result 150 by the coefficient 5 represented by the architecture evaluation configuration parameter corresponding to the first standard antique coin training direction, that is, 150÷5=30, and uses the result 30 as the evaluation period of the first antique coin evaluation network branch under the first standard antique coin training direction. Through such a series of operations, the server accurately determines the evaluation period of the evaluation network branch under a specific standard training direction, providing key data for the subsequent selection of the optimal training direction.

[0037] In an embodiment of the present invention, the standard evaluation configuration parameters corresponding to each of the standard antique coin training directions in the evaluation dimension feature domain of the first antique coin evaluation network branch are used to determine the evaluation cycle of the first antique coin evaluation network branch under each of the standard antique coin training directions, and the following implementation method is also provided.

[0038] Obtaining a standard processing load of the plurality of antique coin evaluation network branches on the same model training environment;

[0039] Determining a target antique coin evaluation network branch with the largest standard processing load from among the plurality of antique coin evaluation network branches;

[0040] If the target antique coin evaluation network branch is not the first antique coin evaluation network branch, then the various standard antique coin training directions in the evaluation dimension feature domain of the first antique coin evaluation network branch are adjusted according to the antique coin training direction of the target antique coin evaluation network branch.

[0041] In an embodiment of the present invention, exemplarily, the server is in a unified model training environment, which has specific hardware resources (such as a multi-core CPU, a high-performance GPU, and a certain amount of memory, etc.) and software conditions (a specific version of a deep learning framework, an operating system, etc.). In this environment, the server runs multiple antique coin evaluation network branches at the same time. For example, assume that there are three antique coin evaluation network branches: the first antique coin evaluation network branch is responsible for evaluating the text-related features of the coin, the second antique coin evaluation network branch focuses on the recognition of coin patterns, and the third antique coin evaluation network branch mainly processes the state evaluation of the coin edge. The server obtains the resource occupancy of each branch in real time during operation through the monitoring system, including indicators such as CPU usage, GPU usage, and memory usage. After a period of monitoring, the server calculates the standard processing load of each branch based on these indicators. Assume that the standard processing load of the first antique coin evaluation network branch is 30% (indicating that 30% of the system resources are occupied), the standard processing load of the second antique coin evaluation network branch is 50%, and the standard processing load of the third antique coin evaluation network branch is 20%. The server compares the standard processing loads of each antique coin evaluation network branch obtained. In the above example, 50% (the second antique coin evaluation network branch)>30% (the first antique coin evaluation network branch)>20% (the third antique coin evaluation network branch), so the server determines that the second antique coin evaluation network branch is the target antique coin evaluation network branch with the largest standard processing load. Since the target antique coin evaluation network branch (the second antique coin evaluation network branch) is not the first antique coin evaluation network branch, the server starts to adjust the operation. Assume that in the sample evaluation configuration parameters corresponding to the antique coin training direction of the second antique coin evaluation network branch, the evaluation granularity is to divide the pattern according to the main elements for evaluation (for example, the floral pattern on the coin is evaluated separately according to flowers and branches and leaves), and the sample division dimension is to evaluate the front and back patterns of the coin separately. The server obtains the multiplication result between the evaluation granularity of the second antique coin evaluation network branch and the sample division dimension. Assuming the evaluation granularity is 5 (representing the division of the pattern into 5 main element categories), and the sample division dimension is 2 (front and back), the multiplication result is 5×2=10. For a standard antique coin training direction in the evaluation dimension feature domain of the first antique coin evaluation network branch (such as the style characteristics of the coin text), the original sample division dimension is 3 (the text is divided into three categories according to the font type: seal script, official script, and regular script). The server divides the multiplication result 10 with the sample division dimension 3 corresponding to the standard antique coin training direction, that is, 10÷3≈3.33.The server can adjust the evaluation granularity corresponding to the standard antique coin training direction according to the calculation result, such as refining the originally coarse evaluation granularity so as to be more coordinated with the training direction of the target antique coin evaluation network branch in terms of resource utilization and evaluation focus, thereby performing similar adjustment operations on each standard antique coin training direction in the evaluation dimension feature domain of the first antique coin evaluation network branch.

[0042] In an embodiment of the present invention, the plurality of standard antique coin training directions include a first standard antique coin training direction; the evaluation configuration parameters include sample evaluation configuration parameters, and the sample evaluation configuration parameters include evaluation granularity and sample division dimension;

[0043] Adjusting the first standard antique coin training direction in the evaluation dimension feature domain of the first antique coin evaluation network branch according to the antique coin training direction of the target antique coin evaluation network branch can be implemented through the following example.

[0044] Obtaining a second multiplication result between the evaluation granularity corresponding to the antique coin training direction of the target antique coin evaluation network branch and the sample division dimension;

[0045] Using the second multiplication result and the second division result between the sample division dimension corresponding to the first standard antique coin training direction, the evaluation granularity corresponding to the first standard antique coin training direction is adjusted; or,

[0046] The sample division dimension corresponding to the first standard antique coin training direction is adjusted by using the third division result between the second multiplication result and the evaluation granularity corresponding to the first standard antique coin training direction.

[0047] In an embodiment of the present invention, exemplarily, the server has determined the target antique coin evaluation network branch (assuming that it is a branch that focuses on evaluating the complexity of coin patterns) and obtains the sample evaluation configuration parameters corresponding to its antique coin training direction. Among them, the evaluation granularity is set to divide the pattern according to key components, such as dividing the character pattern on the coin into five parts such as head, body, and clothing, that is, the evaluation granularity is 5; the sample division dimension is to evaluate the front and back patterns of the coin separately, that is, the sample division dimension is 2. The server multiplies the evaluation granularity and sample division dimension of the target antique coin evaluation network branch to obtain the second multiplication result, that is, 5×2=10. Assume that the first standard antique coin training direction in the evaluation dimension feature domain of the first antique coin evaluation network branch is to evaluate the clarity of the coin text, and its original corresponding sample division dimension is 3 (the text is divided into three categories according to the position on the coin: edge text, middle text, and corner text). The server divides the second multiplication result 10 by the sample division dimension 3 corresponding to the first standard antique coin training direction to obtain the second division result, that is, 10÷3≈3.33. The server adjusts the evaluation granularity corresponding to the first standard antique coin training direction according to this second division result. For example, the original evaluation granularity is to evaluate the text according to the overall outline. Now the evaluation granularity is refined according to the calculation results, and the text is evaluated according to the stroke details, making the evaluation more refined to adapt to the training direction characteristics of the target antique coin evaluation network branch, and achieve better coordination in resource utilization and evaluation focus. Taking the evaluation of the clarity of the coin text in the first standard antique coin training direction as an example, it is assumed that the original corresponding evaluation granularity is 4 (the text clarity is divided into four levels for evaluation). The server divides the second multiplication result 10 by the evaluation granularity 4 corresponding to the first standard antique coin training direction to obtain the third division result, that is, 10÷4=2.5. The server adjusts the sample division dimension corresponding to the first standard antique coin training direction according to this third division result. For example, the characters were originally divided into three categories according to their positions on the coins. Now, based on the calculation results, the characters are adjusted to be divided into two categories (seal script and non-seal script) according to their font styles, thereby changing the sample division dimension and making the evaluation method of the first antique coin evaluation network branch in the standard training direction more consistent with the training direction of the target antique coin evaluation network branch, which helps to improve the training efficiency and evaluation accuracy of the entire antique coin evaluation model.

[0048] In the embodiment of the present invention, the antique coin training direction based on the multiple antique coin evaluation network branches configures the evaluation training framework of the antique coin evaluation model, which can be implemented through the following examples.

[0049] Based on the antique coin training direction of each antique coin evaluation network branch, construct an evaluation unit corresponding to each antique coin evaluation network branch;

[0050] A multimodal feature fusion architecture is established between multiple evaluation units to generate an evaluation training framework for the antique coin evaluation model.

[0051] In an embodiment of the present invention, illustratively, it is assumed that the antique coin evaluation model includes three antique coin evaluation network branches: the first branch is responsible for evaluating the degree of wear of the coin, the second branch is responsible for evaluating the color of the coin pattern, and the third branch is responsible for evaluating the integrity of the coin text. For the first antique coin evaluation network branch, in the evaluation configuration parameters corresponding to the antique coin training direction, the sample division dimension is to divide the coin surface into three parts: the front, the back and the edge, and the architecture evaluation configuration parameters include dividing the training link into four stages: data collection, feature extraction, model training and optimization adjustment. Based on these parameters, the server determines the number of model training environments in the first evaluation unit. For example, according to past experience and computing resources, it is set to 3. The server then obtains the environment configuration in the preset model training environment set, which includes a variety of environments with different hardware performance (such as different models of CPU, GPU) and software versions (different deep learning framework versions). The server respectively allows the first antique coin evaluation network branch to run on the model training environment of each environment configuration and obtains its standard evaluation cycle. Assume that the standard evaluation cycle is 10 seconds in environment A, 12 seconds in environment B, and 8 seconds in environment C. The server then determines the evaluation cycle on the model training environment configured in each environment based on the evaluation granularity corresponding to the antique coin training direction of the first antique coin evaluation network branch (such as dividing the degree of wear into three levels of light, medium, and severe), the architecture evaluation configuration parameters, and the above-mentioned standard evaluation cycle. After calculation, the evaluation cycle of environment A is 15 seconds, environment B is 18 seconds, and environment C is 10 seconds. Therefore, the server determines that the target environment configuration corresponding to the shortest evaluation cycle is environment C. When the number of model training environments configured in the target environment is greater than or equal to the number of model training environments in the first evaluation unit, the server uses the target environment configuration as the environment configuration of the model training environment in the first evaluation unit. Finally, based on the antique coin training direction of the first antique coin evaluation network branch, and combined with the number and environment configuration of the model training environment in the first evaluation unit, the server standardizes the first antique coin evaluation network branch in the model training environment set and successfully constructs the first evaluation unit. In the same way, the server constructs the evaluation units corresponding to the second and third antique coin evaluation network branches respectively. Taking the first evaluation unit (wear degree evaluation) and the second evaluation unit (pattern color evaluation) as examples, the server obtains the common evaluation cardinality between the sample division dimensions (front, back, edge) corresponding to the antique coin training direction of the first antique coin evaluation network branch and the sample division dimensions (such as dividing the pattern into the main pattern and the background pattern) corresponding to the antique coin training direction of the second antique coin evaluation network branch, which is assumed to be 2. The server uses the common evaluation cardinality as the number of evaluation unit pools, and groups the first evaluation unit and the second evaluation unit based on this to obtain 2 evaluation unit pools.Each evaluation unit pool contains at least one evaluation subunit from the first evaluation unit (each evaluation subunit contains one or more model training environments) and at least one evaluation subunit from the second evaluation unit. When establishing a multimodal feature fusion architecture, the server configures dynamic resource allocation for each evaluation unit pool. For example, according to the importance and computational amount of wear degree evaluation and pattern color evaluation, 60% of the computing resources are allocated to the first evaluation unit pool and 40% of the computing resources are allocated to the second evaluation unit pool. The server determines the feature processing flow sequence between the first evaluation unit and the second evaluation unit according to the order of connection between the first antique coin evaluation network branch and the second antique coin evaluation network branch. Assume that the wear degree evaluation is performed first, and then the pattern color evaluation is performed. The server selects a model training environment as a feature acquisition node from each evaluation subunit obtained by the first evaluation unit; selects a model training environment as a feature fusion node from each evaluation subunit obtained by the second evaluation unit, and establishes a cross-modal connection path between the two. For the third evaluation unit (text completeness evaluation), the server also establishes a multimodal feature fusion architecture with the first and second evaluation units in a similar manner, and finally generates an evaluation training framework for the antique coin evaluation model.

[0052] In the embodiment of the present invention, each of the antique coin evaluation network branches is represented as a first antique coin evaluation network branch, and the antique coin evaluation model also includes a second antique coin evaluation network branch connected to the first antique coin evaluation network branch, the first antique coin evaluation network branch corresponds to a first evaluation unit, and the second antique coin evaluation network branch corresponds to a second evaluation unit;

[0053] The establishment of a multimodal feature fusion architecture between multiple evaluation units may be implemented through the following examples.

[0054] The first evaluation unit and the second evaluation unit are grouped to obtain a plurality of evaluation unit pools; each of the evaluation unit pools includes at least one evaluation sub-unit obtained by the first evaluation unit and at least one evaluation sub-unit obtained by the second evaluation unit;

[0055] When establishing a multimodal feature fusion architecture between the first evaluation unit and the second evaluation unit, a dynamic resource allocation is configured for each evaluation unit pool, and a cross-modal connection path between evaluation sub-units belonging to different evaluation units is established in each evaluation unit pool.

[0056] In an embodiment of the present invention, illustratively, it is assumed that the first antique coin evaluation network branch is responsible for evaluating the surface material condition of antique coins, and the corresponding first evaluation unit has been constructed, including multiple evaluation subunits, each evaluation subunit corresponds to a different model training environment, for example, evaluation subunit A runs in a model training environment equipped with a high-end GPU, and evaluation subunit B runs in a model training environment of an ordinary CPU. The second antique coin evaluation network branch is responsible for evaluating the pattern style of antique coins, and the corresponding second evaluation unit also includes multiple evaluation subunits, such as evaluation subunit C runs in a model training environment of a specific version of a deep learning framework, and evaluation subunit D runs in a model training environment of another version of a deep learning framework. The server obtains the public evaluation cardinality between the sample division dimension corresponding to the antique coin training direction of the first antique coin evaluation network branch (for example, dividing the coin surface into a central area and an edge area) and the sample division dimension corresponding to the antique coin training direction of the second antique coin evaluation network branch (for example, dividing the pattern into a traditional pattern and an innovative pattern), assuming that the calculated public evaluation cardinality is 2. The server uses this common evaluation cardinality 2 as the number of evaluation unit pools, and groups the first evaluation unit and the second evaluation unit based on this. The server selects evaluation subunit A from the evaluation subunits A and B of the first evaluation unit, and selects evaluation subunit C from the evaluation subunits C and D of the second evaluation unit to form the first evaluation unit pool; then selects evaluation subunit B of the first evaluation unit and evaluation subunit D of the second evaluation unit to form the second evaluation unit pool. In this way, each evaluation unit pool includes at least one evaluation subunit obtained by the first evaluation unit and at least one evaluation subunit obtained by the second evaluation unit. When establishing a multimodal feature fusion architecture between the first evaluation unit and the second evaluation unit, the server configures dynamic resource allocation for each evaluation unit pool according to the requirements of the antique coin evaluation task and the computational complexity of each evaluation unit. For example, since the evaluation of the surface material condition of the coin is relatively more important in the current evaluation task and the amount of calculation is large, the server allocates 70% of the system computing resources to the first evaluation unit pool (including evaluation subunits A and C) and 30% of the system computing resources to the second evaluation unit pool (including evaluation subunits B and D). The server determines the feature processing flow sequence between the first evaluation unit and the second evaluation unit according to the front-to-back connection order between the first antique coin evaluation network branch and the second antique coin evaluation network branch (assuming that the surface material condition evaluation is performed first, and then the pattern style evaluation is performed). In the evaluation subunit A obtained by the first evaluation unit with the first feature processing flow sequence, the server selects the model training environment as the feature acquisition node; in the evaluation subunit C obtained by the second evaluation unit with the last feature processing flow sequence, the server selects the model training environment as the feature fusion node.Then, the server establishes a cross-modal connection path between the feature collection node (model training environment of evaluation subunit A) and the feature fusion node (model training environment of evaluation subunit C), so that the feature information collected by the first evaluation unit in the process of evaluating the surface material condition of the coin can be smoothly transmitted to the second evaluation unit for feature fusion during pattern style evaluation. For the second evaluation unit pool, the server also establishes a cross-modal connection path between evaluation subunit B and evaluation subunit D in a similar manner, thereby completing the establishment of a multimodal feature fusion architecture between the first evaluation unit and the second evaluation unit to support more efficient evaluation training of the antique coin evaluation model.

[0057] In an embodiment of the present invention, the antique coin training direction of each antique coin evaluation network branch is also used to characterize the evaluation configuration parameters of each antique coin evaluation network branch, and the evaluation configuration parameters include sample evaluation configuration parameters, and the sample evaluation configuration parameters include sample division dimensions;

[0058] The grouping of the first evaluation units and the second evaluation units to obtain a plurality of evaluation unit pools may be implemented through the following example.

[0059] Obtaining a common evaluation cardinality between a sample division dimension corresponding to the antique coin training direction of the first antique coin evaluation network branch and a sample division dimension corresponding to the antique coin training direction of the second antique coin evaluation network branch;

[0060] Taking the public assessment base as the number of assessment unit pools;

[0061] The first evaluation units and the second evaluation units are grouped based on the number of the evaluation unit pools to obtain a plurality of evaluation unit pools.

[0062] In an embodiment of the present invention, illustratively, the first antique coin evaluation network branch is responsible for evaluating the text features of antique coins, and in the sample evaluation configuration parameters corresponding to the antique coin training direction, the sample division dimension is set to divide the text on the coin into three categories according to the font type: seal script, official script, and regular script. The second antique coin evaluation network branch is responsible for evaluating the pattern features of the coin, and the sample division dimension corresponding to the antique coin training direction is to divide the pattern into three categories according to the theme: human pattern, animal pattern, and plant pattern. The server analyzes these two sample division dimensions and determines the common evaluation cardinality between them through a specific algorithm or rule. For example, the server can find the greatest common divisor of the number of categories in the two sample division dimensions. In this example, the greatest common divisor of 3 and 3 is 3, so the common evaluation cardinality is 3. The server directly uses the obtained common evaluation cardinality 3 as the number of evaluation unit pools. This is because the common evaluation cardinality reflects a certain degree of association or commonality between the two evaluation network branches in the sample division dimension, and using this as the number of evaluation unit pools helps to better integrate the resources and feature processing of the two evaluation units in the subsequent grouping process. The first evaluation unit includes multiple evaluation subunits, which correspond to different model training environments. For example, evaluation subunit E runs in a model training environment with high memory capacity and is specifically used to process a large amount of text image data; evaluation subunit F runs in a model training environment equipped with a high-speed CPU and can quickly extract text features. The second evaluation unit also includes multiple evaluation subunits. For example, evaluation subunit G runs in a model training environment with a specific image recognition library installed, which is conducive to pattern recognition; evaluation subunit H runs in an optimized deep learning framework environment, which can efficiently perform pattern feature analysis. The server groups the first evaluation unit and the second evaluation unit based on the determined number of evaluation unit pools 3. The server selects evaluation subunit E from evaluation subunits E and F of the first evaluation unit, and selects evaluation subunit G from evaluation subunits G and H of the second evaluation unit to form the first evaluation unit pool; then selects evaluation subunit F of the first evaluation unit and evaluation subunit H of the second evaluation unit to form the second evaluation unit pool; then the remaining part of the resources of the first evaluation unit or an auxiliary evaluation sub-part and another auxiliary evaluation sub-part of the second evaluation unit form the third evaluation unit pool. Through the above steps, the server successfully grouped the first evaluation unit and the second evaluation unit to obtain multiple evaluation unit pools, each of which contained evaluation sub-units from two different evaluation units, laying the foundation for the subsequent establishment of a multimodal feature fusion architecture, so that in the antique coin evaluation process, the evaluation processing related to text features and pattern features can be better coordinated.

[0063] In an embodiment of the present invention, each of the evaluation subunits includes one or more model training environments;

[0064] The establishment of cross-modal connection paths between evaluation sub-units belonging to different evaluation units in each evaluation unit pool may be implemented through the following example.

[0065] Determine a feature processing flow sequence between the first evaluation unit and the second evaluation unit according to a front-to-back connection order between the first antique coin evaluation network branch and the second antique coin evaluation network branch;

[0066] Selecting a model training environment as a feature collection node from among the evaluation sub-units acquired by the previous evaluation unit in the feature processing flow sequence;

[0067] In each of the evaluation sub-units acquired by the evaluation unit following the feature processing flow sequence, a model training environment is selected as a feature fusion node;

[0068] A cross-modal connection path is established between the feature collection node and the feature fusion node.

[0069] In an embodiment of the present invention, illustratively, it is assumed that the server is performing a cross-modal connection path establishment operation on the first evaluation unit and the second evaluation unit of the antique coin evaluation model, wherein the first antique coin evaluation network branch is responsible for evaluating the texture features of the antique coins, and the second antique coin evaluation network branch is responsible for evaluating the color features of the antique coins. The server clarifies the feature processing flow sequence between the first evaluation unit and the second evaluation unit according to the front-to-back connection order between the first antique coin evaluation network branch and the second antique coin evaluation network branch. In the antique coin evaluation process, the evaluation result of the texture feature will affect the analysis of the color feature, so the texture feature evaluation is performed first, and then the color feature evaluation is performed. That is, the feature processing flow corresponding to the first evaluation unit is in the front row, and the feature processing flow corresponding to the second evaluation unit is in the back row, and such a sequence is determined. The first evaluation unit includes multiple evaluation subunits, each of which corresponds to a different model training environment. The evaluation subunit I runs in a model training environment equipped with professional image texture analysis software, and can extract the texture details of the antique coin image with high precision; the evaluation subunit J runs in a hardware environment with powerful parallel computing capabilities, and can quickly process a large amount of texture data. Since the first evaluation unit is in the front of the feature processing flow sequence, the server selects the model training environment of evaluation subunit I from each evaluation subunit obtained as the feature acquisition node. Because the environment has professional texture analysis software, it is more suitable for collecting data information related to texture features. The second evaluation unit also includes multiple evaluation subunits. Evaluation subunit K runs in a model training environment with an advanced color recognition algorithm library installed, and can accurately identify the color information of antique coin images; evaluation subunit L runs in an optimized deep learning framework environment, and can efficiently perform color feature fusion and analysis. Since the second evaluation unit is in the back of the feature processing flow sequence, the server selects the model training environment of evaluation subunit K from each evaluation subunit obtained as the feature fusion node. Because the environment has an advanced color recognition algorithm library, it can better fuse the texture features collected by the first evaluation unit with its own color features. After determining the feature acquisition node (model training environment of evaluation subunit I) and the feature fusion node (model training environment of evaluation subunit K), the server establishes a cross-modal connection path between the two. The server configures specific network protocols and data transmission interfaces to ensure that the texture feature data of antique coins collected by the first evaluation unit can be smoothly transmitted to the feature fusion node of the second evaluation unit. For example, the server sets up a data transmission channel based on the TCP / IP protocol, and encrypts and compresses the data to ensure the security and efficiency of the data during transmission. By establishing this cross-modal connection path, the texture features and color features of antique coins can effectively interact and fuse between the two evaluation units, thereby improving the comprehensive evaluation capabilities of the antique coin evaluation model.

[0070] In an embodiment of the present invention, each of the antique coin evaluation network branches is represented as a first antique coin evaluation network branch, and the first antique coin evaluation network branch corresponds to a first evaluation unit; the antique coin training direction of each of the antique coin evaluation network branches is also used to characterize the evaluation configuration parameters of each of the antique coin evaluation network branches, and the evaluation configuration parameters include sample evaluation configuration parameters and architecture evaluation configuration parameters, and the sample evaluation configuration parameters include sample division dimensions;

[0071] Based on the antique coin training direction of the first antique coin evaluation network branch, the first evaluation unit is configured, which can be implemented through the following examples.

[0072] Determine the number of model training environments in the first evaluation unit based on the sample partitioning dimension and the architecture evaluation configuration parameter in the evaluation configuration parameter corresponding to the antique coin training direction of the first antique coin evaluation network branch;

[0073] Obtaining an environment configuration in a preset model training environment set, and determining an evaluation cycle of the first antique coin evaluation network branch on the model training environment of each environment configuration based on the antique coin training direction of the first antique coin evaluation network branch;

[0074] Determining the environment configuration of the model training environment in the first evaluation unit based on the evaluation cycle of the first antique coin evaluation network branch on the model training environment of each environment configuration;

[0075] Based on the antique coin training direction of the first antique coin evaluation network branch and based on the number and environment configuration of the model training environments in the first evaluation unit, the first antique coin evaluation network branch is standardized in the model training environment set to configure the first evaluation unit.

[0076] In an embodiment of the present invention, illustratively, it is assumed that the server is processing the configuration work of the first antique coin evaluation network branch and its corresponding first evaluation unit in the antique coin evaluation model, and the first antique coin evaluation network branch is responsible for evaluating the rust condition of antique coins. In the sample evaluation configuration parameters corresponding to the antique coin training direction of the first antique coin evaluation network branch, the sample division dimension is to divide the surface of the antique coin into three categories according to the rust distribution area: the central rust area, the edge rust area and the mixed rust area; the architecture evaluation configuration parameters include dividing the training link into four stages: data preprocessing, feature extraction, model training, and accuracy verification. Based on these evaluation configuration parameters, the server comprehensively considers the complexity of the evaluation task and the expected processing efficiency, and determines the number of model training environments in the first evaluation unit. Since the rust condition evaluation needs to process image data in different areas, and the training link is relatively complex, the server determines that the number of model training environments is 4. The server obtains a preset model training environment set, which contains a variety of different environment configurations. For example, environment M is configured with a high-performance GPU and large-capacity memory, which is suitable for processing complex image data; environment N is configured with an ordinary CPU and medium-capacity memory; environment O is installed with a specific version of the deep learning framework, which is optimized for image feature extraction; and environment P is specially set up in terms of network transmission. Based on the antique coin training direction of the first antique coin evaluation network branch, the server allows the branch to run test tasks on the model training environments configured in the above-mentioned environments. The server selects a group of antique coin image samples containing different rust conditions, and allows the first antique coin evaluation network branch to process them, records the time from the beginning to the completion of the evaluation, and obtains the evaluation cycle on the model training environment configured in each environment. Assume that the evaluation cycle on environment M is 12 seconds, 20 seconds on environment N, 15 seconds on environment O, and 18 seconds on environment P. The server compares the evaluation cycles of the first antique coin evaluation network branch on the model training environments configured in each environment. Since the evaluation cycle of environment M is the shortest, which is 12 seconds, it can complete the rust condition evaluation task more efficiently, so the server determines to use the configuration of environment M as the environment configuration of the model training environment in the first evaluation unit. The server performs standardized operations on the first antique coin evaluation network branch in the model training environment set based on the antique coin training direction of the first antique coin evaluation network branch, as well as the number of model training environments (4) and environment configuration (environment M) in the first evaluation unit. The server deploys the same hardware and software conditions for the four model training environments in the first evaluation unit according to the configuration requirements of environment M, including installing the corresponding GPU driver, deep learning framework version, etc.At the same time, according to the training direction of the first antique coin evaluation network branch, the parameters in the model training environment are set, such as adjusting the algorithm parameters related to rust feature extraction, etc., so as to complete the configuration of the first evaluation unit so that it can efficiently perform the evaluation task of the rust condition of antique coins.

[0077] In the embodiment of the present invention, the environmental configuration of the model training environment in the first evaluation unit is determined based on the evaluation cycle of the first antique coin evaluation network branch on the model training environment of each environmental configuration, which can be implemented through the following example.

[0078] Determine the target environment configuration corresponding to the shortest evaluation cycle based on the evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environment configuration;

[0079] Acquire the number of model training environments configured in the target environment from the set of model training environments;

[0080] When the number of model training environments of the target environment configuration is greater than or equal to the number of model training environments in the first evaluation unit, the target environment configuration is used as the environment configuration of the model training environment in the first evaluation unit.

[0081] In an embodiment of the present invention, illustratively, it is assumed that the server is determining the environment configuration of the model training environment for the first evaluation unit in the antique coin evaluation model, and the first antique coin evaluation network branch is responsible for evaluating the pattern fineness on the antique coin. The server has obtained the evaluation cycle of the first antique coin evaluation network branch in the model training environment with different environment configurations. There are three different environment configurations in the model training environment set: environment X is equipped with a high-end GPU and optimized deep learning software, which is specially used for fast calculation of complex images; environment Y uses an ordinary CPU and a common deep learning framework version; environment Z is specially optimized in terms of storage and data reading. The first antique coin evaluation network branch evaluates a set of antique coin image samples containing various pattern fineness in environment X, and the time taken to complete the evaluation is 8 seconds; the same sample is evaluated in environment Y, and it takes 15 seconds; when evaluating in environment Z, it takes 12 seconds. The server compares these three evaluation cycles, 8 seconds <12 seconds <15 seconds, thereby determining that environment X is the target environment configuration corresponding to the shortest evaluation cycle, because it can complete the evaluation task of the pattern fineness of antique coins at the fastest speed. The server queries the number of model training environments owned by the target environment configuration (environment X) from the database or resource management system that stores the model training environment information. Assume that it is known through the query that there are currently 6 available model training environments in environment X within the resource range of the server. The server previously determined that the number of model training environments in the first evaluation unit is 5 based on factors such as the evaluation task requirements of the first antique coin evaluation network branch. Since the number of model training environments of the target environment configuration (environment X) is 6, which is greater than the number of model training environments in the first evaluation unit, 5, it meets the "greater than or equal to" condition. Therefore, the server uses the target environment configuration (environment X) as the environment configuration of the model training environment in the first evaluation unit. Next, the server will deploy and set up the 5 model training environments in the first evaluation unit according to the hardware and software configuration requirements of environment X, such as installing the same high-end GPU driver as environment X, configuring the corresponding deep learning software parameters, etc., to ensure that the first evaluation unit can run in an efficient environment and accurately evaluate the pattern fineness of antique coins.

[0082] In an embodiment of the present invention, the sample evaluation configuration parameters further include evaluation granularity;

[0083] The antique coin training direction based on the first antique coin evaluation network branch determines the evaluation cycle of the first antique coin evaluation network branch in the model training environment configured in each environment, including:

[0084] Obtaining a standard evaluation cycle of the first antique coin evaluation network branch in the model training environment configured in each environment;

[0085] Based on the evaluation granularity and architecture evaluation configuration parameters corresponding to the antique coin training direction of the first antique coin evaluation network branch, and the standard evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environmental configuration, the evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environmental configuration is determined.

[0086] In the embodiment of the present invention, illustratively, it is assumed that the server is processing the determination of the evaluation cycle of the first antique coin evaluation network branch in the antique coin evaluation model under different model training environments, and the first antique coin evaluation network branch is responsible for evaluating the wear condition of antique coins. There are three different environment configurations in the model training environment set managed by the server: environment A has a high-performance CPU and GPU combination with a large memory capacity; environment B uses a CPU with ordinary performance and a relatively small memory; environment C has an advantage in data transmission bandwidth. The server standardizes the first antique coin evaluation network branch on the model training environments of these three environment configurations to make it an executable network entity. Then, a group of representative samples are selected from the database storing antique coin image samples, and these samples cover antique coin images with different degrees of wear. The server allows the network entity of the first antique coin evaluation network branch to process this group of samples on the model training environment of environment A, record the time from start to end, and obtain a standard evaluation cycle of 10 seconds on environment A. Similarly, the standard evaluation cycle obtained on environment B is 18 seconds, and the standard evaluation cycle obtained on environment C is 12 seconds. In the sample evaluation configuration parameters corresponding to the antique coin training direction of the first antique coin evaluation network branch, the evaluation granularity is set to divide the wear condition of antique coins into five levels according to the wear area ratio (mild: wear area ratio <10%; light: 10%≤wear area ratio <25%; moderate: 25%≤wear area ratio <50%; heavy: 50%≤wear area ratio <75%; heavy: wear area ratio ≥75%). The architecture evaluation configuration parameters include dividing the training phase into four stages: data screening, feature extraction, model training, and result verification. For environment A, the server is based on the evaluation granularity, considering that the wear condition needs to be subdivided into five levels. At the same time, combined with the resource requirements of the four training stages in the architecture evaluation configuration parameters, the previously obtained standard evaluation cycle of 10 seconds is adjusted. Assume that according to specific calculation rules (such as considering the calculation amount of the increase in evaluation granularity and the resource occupation of the architecture evaluation configuration parameters), the evaluation cycle on environment A is finally determined to be 13 seconds. For environment B, due to its relatively weak hardware performance, it is less efficient in processing subdivided wear level assessment tasks and meeting the requirements of architecture assessment configuration parameters. Based on the assessment granularity and architecture assessment configuration parameters, the standard assessment cycle of 18 seconds is adjusted to obtain an assessment cycle of 22 seconds on environment B. For environment C, although data transmission has advantages, it still requires a certain amount of additional computing time when processing subdivided wear level assessment tasks in combination with architecture assessment configuration parameters. Based on the assessment granularity and architecture assessment configuration parameters, the standard assessment cycle of 12 seconds is adjusted to obtain an assessment cycle of 15 seconds on environment C.Through the above steps, the server comprehensively considers the evaluation granularity, architecture evaluation configuration parameters and standard evaluation cycle, and determines the evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environment configuration, providing an accurate basis for the subsequent selection of a suitable model training environment to configure the first evaluation unit.

[0087] In an embodiment of the present invention, when executing the training process for the antique coin evaluation model, multiple evaluation units in the evaluation training framework are called to execute the evaluation dimensions of the corresponding antique coin evaluation network branches based on the training processes corresponding to the antique coin training directions of the corresponding antique coin evaluation network branches, which can be implemented through the following examples.

[0088] When executing the training process for the antique coin evaluation model, the training phase of the antique coin evaluation model is divided into a plurality of training steps in the training link, and one training step corresponds to an evaluation dimension of one antique coin evaluation network branch;

[0089] Execute the training link, and when executing the training link, determine the target evaluation unit used to execute the current training step, call the training process corresponding to the antique coin training direction of the target evaluation unit based on the antique coin evaluation network branch corresponding to the target evaluation unit, and execute the evaluation dimension training process of the antique coin evaluation network branch corresponding to the target evaluation unit.

[0090] In an embodiment of the present invention, illustratively, it is assumed that the server is executing a training process for an antique coin evaluation model, and the model includes three antique coin evaluation network branches: branch A is responsible for evaluating the material of antique coins, branch B is responsible for evaluating the age characteristics of the pattern on the coin, and branch C is responsible for evaluating the clarity of the text. When the server executes the training process for the antique coin evaluation model, the training stage is divided into multiple training links, and each training link is further subdivided into multiple training steps. For example, the training link includes a data preprocessing link, a feature extraction link, and a model training link. In the data preprocessing link, it is further divided into training steps such as sample screening and data annotation. Among them, the sample screening step corresponds to the evaluation dimension of branch A, that is, to select samples suitable for evaluating materials from a large number of antique coin image samples; the data annotation step corresponds to the evaluation dimension of branch B, and the information such as the age of the coin pattern is annotated. In the feature extraction link, it is divided into a material feature extraction step, a pattern age feature extraction step, and a text clarity feature extraction step, which correspond to the evaluation dimensions of branch A, branch B, and branch C respectively. In the model training phase, training steps for each branch evaluation dimension are also set, such as material evaluation model training steps, pattern age evaluation model training steps, and text clarity evaluation model training steps. In this way, one training step corresponds to the evaluation dimension of an antique coin evaluation network branch. When the server executes the material feature extraction step of the feature extraction phase, it is determined that the target evaluation unit for executing the current training step is the first evaluation unit corresponding to branch A. The training process corresponding to the antique coin training direction of the first evaluation unit includes: first, the samples are divided according to a specific evaluation granularity (the materials are divided into categories such as metal materials and non-metal materials); then, according to the architecture evaluation configuration parameters (the training phase process segmentation dimensions are data import, preliminary feature extraction, fine feature extraction and feature verification, and the feature analysis dimensions include color, density, etc.), the samples are processed in the selected model training environment. The server calls the first evaluation unit, which extracts material features from the antique coin image samples based on the training process corresponding to the antique coin training direction of branch A. Specifically, the samples are first classified according to the evaluation granularity, and then segmented into dimensions according to the training process, and data is imported in sequence. The preliminary feature extraction algorithm is used to extract preliminary material features such as color and density, and then fine feature extraction is performed to obtain more accurate material information. Finally, feature verification is performed to ensure that the extracted features are accurate and reliable, and the training process of the corresponding evaluation dimension of branch A is completed. Similarly, when executing other training steps, the server will also determine the corresponding target evaluation unit and call it to execute the evaluation dimension training process according to the training direction of the corresponding antique coin evaluation network branch, and gradually complete the training of the entire antique coin evaluation model.

[0091] 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 method for comprehensive evaluation of the appearance state of antique coins based on joint modeling of image recognition and deep learning. Figure 2 As shown, Figure 2 The block diagram of the computer device 100 provided in the 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. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0092] 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. A comprehensive evaluation method for the appearance of antique coins based on joint modeling of image recognition and deep learning, characterized in that: include: Acquire an untrained antique coin evaluation model, wherein the antique coin evaluation model includes a plurality of antique coin evaluation network branches; The evaluation dimension allocation process is performed on each antique coin evaluation network branch respectively to obtain the antique coin training direction of each antique coin evaluation network branch; the antique coin training direction of each antique coin evaluation network branch is used to characterize the evaluation dimension of each antique coin evaluation network branch in the training stage and the training process of the evaluation dimension; Based on the antique coin training directions of the multiple antique coin evaluation network branches, an evaluation training framework of the antique coin evaluation model is configured, wherein the evaluation training framework includes multiple evaluation units; one evaluation unit corresponds to one antique coin evaluation network branch, and each evaluation unit is configured based on the antique coin training direction of the corresponding antique coin evaluation network branch; When executing the training process for the antique coin evaluation model, calling multiple evaluation units in the evaluation training framework to respectively execute the evaluation dimension training process of the corresponding antique coin evaluation network branch based on the training process corresponding to the antique coin training direction of the corresponding antique coin evaluation network branch, so as to obtain the antique coin evaluation model that has completed the training; The antique coin image to be evaluated is imported into the antique coin evaluation model that has completed the training, and a comprehensive evaluation result of the appearance state of the antique coin image to be evaluated is obtained.

2. The method according to claim 1, characterized in that The antique coin training direction of each antique coin evaluation network branch is also used to characterize the evaluation configuration parameters of each antique coin evaluation network branch, and each antique coin evaluation network branch is represented as a first antique coin evaluation network branch; Performing evaluation dimension allocation processing on the first antique coin evaluation network branch to obtain the antique coin training direction of the first antique coin evaluation network branch includes: Acquire an evaluation dimension feature domain of the first antique coin evaluation network branch, wherein the evaluation dimension feature domain of the first antique coin evaluation network branch includes a plurality of standard antique coin training directions predetermined for the first antique coin evaluation network branch; Determining the evaluation cycle of the first antique coin evaluation network branch under each of the standard antique coin training directions based on the standard evaluation configuration parameters corresponding to each of the standard antique coin training directions in the evaluation dimension feature domain of the first antique coin evaluation network branch; The standard antique coin training direction corresponding to the shortest evaluation period in the evaluation dimension feature domain of the first antique coin evaluation network branch is selected as the antique coin training direction of the first antique coin evaluation network branch; Among them, the evaluation configuration parameters include sample evaluation configuration parameters and architecture evaluation configuration parameters; the sample evaluation configuration parameters include evaluation granularity and sample division dimension; the architecture evaluation configuration parameters include training phase process segmentation dimension and feature analysis dimension.

3. The method according to claim 2, characterized in that The plurality of standard antique coin training directions include a first standard antique coin training direction; the evaluation configuration parameters include sample evaluation configuration parameters and architecture evaluation configuration parameters, and the sample evaluation configuration parameters include evaluation granularity; Determining an evaluation cycle of the first antique coin evaluation network branch under the first standard antique coin training direction based on the standard evaluation configuration parameters corresponding to the first standard antique coin training direction includes: In a first model training environment, the first antique coin evaluation network branch is standardized into a network entity of the first antique coin evaluation network branch, and the network entity of the first antique coin evaluation network branch is called to process a sample antique coin image to obtain a standard evaluation cycle of the first antique coin evaluation network branch in the first model training environment; Obtaining a first multiplication result between an evaluation granularity in a standard evaluation configuration parameter corresponding to the first standard antique coin training direction and a standard evaluation period of the first antique coin evaluation network branch in the first model training environment; A first division result between the first multiplication result and the architecture evaluation configuration parameter corresponding to the first standard antique coin training direction is used as an evaluation period of the first antique coin evaluation network branch under the first standard antique coin training direction.

4. The method according to claim 2, characterized in that: Before determining the evaluation cycle of the first antique coin evaluation network branch under each of the standard antique coin training directions based on the standard evaluation configuration parameters corresponding to each of the standard antique coin training directions in the evaluation dimension feature domain of the first antique coin evaluation network branch, the method further includes: Obtaining a standard processing load of the plurality of antique coin evaluation network branches on the same model training environment; Determining a target antique coin evaluation network branch with the largest standard processing load from among the plurality of antique coin evaluation network branches; If the target antique coin evaluation network branch is not the first antique coin evaluation network branch, then the various standard antique coin training directions in the evaluation dimension feature domain of the first antique coin evaluation network branch are adjusted according to the antique coin training direction of the target antique coin evaluation network branch.

5. The method according to claim 4, characterized in that The plurality of standard antique coin training directions include a first standard antique coin training direction; the evaluation configuration parameters include sample evaluation configuration parameters, and the sample evaluation configuration parameters include evaluation granularity and sample division dimension; Adjusting the first standard antique coin training direction in the evaluation dimension feature domain of the first antique coin evaluation network branch according to the antique coin training direction of the target antique coin evaluation network branch includes: Obtaining a second multiplication result between the evaluation granularity corresponding to the antique coin training direction of the target antique coin evaluation network branch and the sample division dimension; Using the second multiplication result and the second division result between the sample division dimension corresponding to the first standard antique coin training direction, the evaluation granularity corresponding to the first standard antique coin training direction is adjusted; or, The sample division dimension corresponding to the first standard antique coin training direction is adjusted by using the third division result between the second multiplication result and the evaluation granularity corresponding to the first standard antique coin training direction.

6. The method according to claim 1, characterized in that The antique coin training direction based on the multiple antique coin evaluation network branches configures the evaluation training framework of the antique coin evaluation model, including: Based on the antique coin training direction of each antique coin evaluation network branch, construct an evaluation unit corresponding to each antique coin evaluation network branch; A multimodal feature fusion architecture is established between multiple evaluation units to generate an evaluation training framework for the antique coin evaluation model.

7. The method according to claim 6, characterized in that Each of the antique coin evaluation network branches is represented as a first antique coin evaluation network branch, and the antique coin evaluation model also includes a second antique coin evaluation network branch connected to the first antique coin evaluation network branch, the first antique coin evaluation network branch corresponds to a first evaluation unit, and the second antique coin evaluation network branch corresponds to a second evaluation unit; The multimodal feature fusion architecture between multiple evaluation units is established, including: The first evaluation unit and the second evaluation unit are grouped to obtain a plurality of evaluation unit pools; each of the evaluation unit pools includes at least one evaluation sub-unit obtained by the first evaluation unit and at least one evaluation sub-unit obtained by the second evaluation unit; When establishing a multimodal feature fusion architecture between the first evaluation unit and the second evaluation unit, dynamic resource allocation is configured for each evaluation unit pool, and a cross-modal connection path between evaluation sub-units belonging to different evaluation units is established in each evaluation unit pool; The antique coin training direction of each antique coin evaluation network branch is also used to characterize the evaluation configuration parameters of each antique coin evaluation network branch, wherein the evaluation configuration parameters include sample evaluation configuration parameters, and the sample evaluation configuration parameters include sample division dimensions; The first evaluation unit and the second evaluation unit are grouped to obtain a plurality of evaluation unit pools, including: Obtaining a common evaluation cardinality between a sample division dimension corresponding to the antique coin training direction of the first antique coin evaluation network branch and a sample division dimension corresponding to the antique coin training direction of the second antique coin evaluation network branch; Taking the public assessment base as the number of assessment unit pools; Based on the number of the evaluation unit pools, grouping the first evaluation unit and the second evaluation unit to obtain a plurality of evaluation unit pools; Each of the evaluation subunits includes one or more model training environments; The step of establishing a cross-modal connection path between evaluation sub-units belonging to different evaluation units in each evaluation unit pool includes: Determine a feature processing flow sequence between the first evaluation unit and the second evaluation unit according to a front-to-back connection order between the first antique coin evaluation network branch and the second antique coin evaluation network branch; Selecting a model training environment as a feature collection node from among the evaluation sub-units acquired by the previous evaluation unit in the feature processing flow sequence; In each of the evaluation sub-units acquired by the evaluation unit following the feature processing flow sequence, a model training environment is selected as a feature fusion node; A cross-modal connection path is established between the feature collection node and the feature fusion node.

8. The method according to claim 6, characterized in that Each of the antique coin evaluation network branches is represented as a first antique coin evaluation network branch, and the first antique coin evaluation network branch corresponds to a first evaluation unit; the antique coin training direction of each of the antique coin evaluation network branches is also used to characterize the evaluation configuration parameters of each of the antique coin evaluation network branches, and the evaluation configuration parameters include sample evaluation configuration parameters and architecture evaluation configuration parameters, and the sample evaluation configuration parameters include sample division dimensions and evaluation granularity; Based on the antique coin training direction of the first antique coin evaluation network branch, configuring the first evaluation unit includes: Determine the number of model training environments in the first evaluation unit based on the sample partitioning dimension and the architecture evaluation configuration parameter in the evaluation configuration parameter corresponding to the antique coin training direction of the first antique coin evaluation network branch; Obtaining the environment configuration in the preset model training environment set, and obtaining the standard evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environment configuration; Determine the evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environment configuration based on the evaluation granularity and architecture evaluation configuration parameters corresponding to the antique coin training direction of the first antique coin evaluation network branch and the standard evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environment configuration; Determine the target environment configuration corresponding to the shortest evaluation cycle based on the evaluation cycle of the first antique coin evaluation network branch in the model training environment of each environment configuration; Acquire the number of model training environments configured in the target environment from the set of model training environments; When the number of model training environments of the target environment configuration is greater than or equal to the number of model training environments in the first evaluation unit, using the target environment configuration as the environment configuration of the model training environment in the first evaluation unit; Based on the antique coin training direction of the first antique coin evaluation network branch and based on the number and environment configuration of the model training environments in the first evaluation unit, the first antique coin evaluation network branch is standardized in the model training environment set to configure the first evaluation unit.

9. The method according to claim 1, characterized in that: When executing the training process for the antique coin evaluation model, calling multiple evaluation units in the evaluation training framework to respectively perform the evaluation dimensions of the corresponding antique coin evaluation network branches based on the training processes corresponding to the antique coin training directions of the corresponding antique coin evaluation network branches, including: When executing the training process for the antique coin evaluation model, the training phase of the antique coin evaluation model is divided into a plurality of training steps in the training link, and one training step corresponds to an evaluation dimension of one antique coin evaluation network branch; Execute the training link, and when executing the training link, determine the target evaluation unit used to execute the current training step, call the training process corresponding to the antique coin training direction of the target evaluation unit based on the antique coin evaluation network branch corresponding to the target evaluation unit, and execute the evaluation dimension training process of the antique coin evaluation network branch corresponding to the target evaluation unit.

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