Oracle bone auxiliary decipherment classification method and system based on classification model

Abstract

The application discloses a kind of oracle bone inscriptions auxiliary decipherment classification methods based on classification model, the method includes: obtaining data source, and based on the data source carries out the division of data set;Training classification model, wherein it includes: training set is input into training model and is trained, in training process, feature coefficient weighting is carried out to different period Chinese character image in each group of image of training set;After each round of training ends, the model performance evaluation parameter obtained from validation set is output, and the classification model obtained according to the advantages and disadvantages of these parameters is screened;After all training ends, the picture in test set is removed label as un-deciphered oracle bone inscriptions and placed in the classification model for prediction, and the result with higher than preset value is output Prediction accuracy;Based on classification model, the classification and auxiliary decipherment of un-deciphered oracle bone inscriptions are carried out.The method conveniently realizes the classification paradigm of oracle bone inscriptions auxiliary decipherment.The application also provides a corresponding oracle bone inscriptions auxiliary decipherment classification system based on classification model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, computer vision and ancient Chinese characters, and more particularly relates to a classification method and system for auxiliary deciphering of oracle bone inscriptions based on a classification model. BACKGROUND

[0002] Characters are a symbol of civilization and an important symbol of a nation. As the earliest systematic characters in China, oracle bone inscriptions are a precious archaeological resource. Deepening the study of oracle bone inscriptions and thoroughly exploring their historical heritage and cultural connotations can not only enhance the cultural self-confidence and historical identity of the Chinese nation, but also promote cultural self-strengthening.

[0003] Since 2016, the China Museum of Writing has twice published a "reward announcement" in Guangming Daily to reward experts and scholars who have made significant contributions to the study of oracle bone inscriptions. Although the state has been promoting the study of oracle bone inscriptions, and more and more experts and scholars are involved in the study of oracle bone inscriptions, and social attention is gradually increasing, the proportion of oracle bone inscriptions deciphered is still small, and follow-up research work is difficult to carry out, and the interpretation of oracle bone inscriptions is still the core of the study of oracle bone inscriptions.

[0004] However, the study of oracle bone inscriptions still relies on ancient Chinese character experts to consult ancient books and use reasoning and ancient book proof to decipher oracle bone inscriptions. Due to the complexity of the interpretation of oracle bone inscriptions, it requires multidimensional in-depth research and a large amount of evidence to be recognized by Chinese character scholars, and the traditional method of oracle bone inscription study is time-consuming and laborious. SUMMARY

[0005] The embodiment of the present application provides a classification method for auxiliary deciphering of oracle bone inscriptions based on a classification model, which comprises:

[0006] Step 1, obtaining a data source and dividing a data set based on the data source, which comprises:

[0007] Collecting images of Chinese character font style evolution in each period, removing variant characters, dead characters and combined characters in the images to obtain a character evolution data set;

[0008] Determining a training set, a test set and a validation set of the data set, which comprises: selecting one-tenth of the oracle bone inscription images in the collected font image of each period as the test set for verifying the generalization ability of the model. When selecting the test set, all oracle bone inscription images corresponding to the same Chinese character but with a difference higher than a preset difference degree in the data set are removed. The remaining picture data of the character evolution data set after removing the test set is the training set. In the determined training set, a part of images in each group of Chinese character images is randomly selected according to a preset proportion to form a validation set, wherein the validation set contains font evolution images of different periods;

[0009] Step two involves training the classification model, which includes: inputting the training set into the training model for training; during the training process, weighting the feature coefficients of Chinese character images from different periods in each group of images in the training set; after each round of training, outputting the model performance evaluation parameters of the validation set; and selecting the classification model based on the quality of these parameters; finally, removing the labels from the images in the test set and placing them as undeciphered oracle bone script characters into the classification model for prediction, and outputting the results with a prediction accuracy higher than a preset value.

[0010] Step 3: Based on the classification model, classify and assist in the deciphering of undeciphered oracle bone script. This includes building an interactive prediction interface, inputting the undeciphered oracle bone script into the interactive prediction interface, outputting the corresponding Chinese characters and their similarity scores, and finally determining the corresponding Chinese characters based on the similarity scores between the output Chinese characters of different categories and the oracle bone script.

[0011] In one embodiment of the present invention, the glyph evolution images in the character evolution dataset are constructed based on at least one of the following data sources: books associated with oracle bone script, books associated with bronze script, Spring and Autumn script, Warring States script, and clerical script, and Internet data, wherein the bronze script images and the oracle bone script images have the highest similarity.

[0012] In one embodiment of the present invention, in step one, when collecting data from the book, the present invention constructs a chain-like file processing system, including: using a border extraction edge detection method to segment and extract content related to font evolution in the electronic document of the book, and saving the extracted images according to the corresponding Chinese character categories.

[0013] In one embodiment of the present invention, some images in the text evolution dataset contain noise and large areas of black pixels. A denoising network is used to clean the dataset. For noise that cannot be removed by the denoising network, manual removal is performed to obtain a clean text evolution dataset. In particular, the obtained images from the oracle bone script period are rotated by a certain angle (<15°) to expand the dataset, increase the proportion of images from the oracle bone script period in the dataset, and improve classification accuracy.

[0014] In one embodiment of the present invention, the specific steps for training a classification model using a text evolution dataset include: inputting the true label represented as y n The training images, which have different dimensional feature representations {x1, x2, ..., x... i The features are input into the classification model Ψ({x1,x2,……,x) i In the process of generating prediction results, the corresponding prediction results are produced. By and y nThe weighted loss function L, used as the independent variable, measures the error and optimizes the classification model parameters to obtain the feature space corresponding to each category. Based on the training data, the classification algorithm is optimized. During the testing phase, the classification model can output the corresponding predicted label based on the input undeciphered oracle bone inscription image.

[0015] In one embodiment of the present invention, the classification model specifically used is a ResNet deep convolutional neural network based on residual blocks, and the structure of the deep residual convolutional neural network is as follows:

[0016] Input a text evolution image, which is a 128×128 grayscale image Z∈R. 128×128×1 First, the image is processed through an initial convolutional layer and a pooling layer. Specifically, it is downsampled through a 7×7 ConV layer with a stride of 2, resulting in 64 feature maps. Then, the feature maps are passed through a 3×3 Max-Pooling layer to reduce the image size by half.

[0017] The image then enters four residual stages, each consisting of 3, 4, 6, and 3 residual blocks respectively, with each block containing multiple residual units. Each residual block consists of two 3×3 convolutional layers, including batch normalization and the ReLU non-linear activation function. Within each residual block, a skip connection is provided to directly pass the input to the output of the residual block, thus propagating gradients without introducing additional parameters.

[0018] Next, a global average pooling layer is used to perform global average pooling on the output of the last residual block, reducing the feature map size to 1×1. Finally, the output of the global average pooling layer is connected to a fully connected layer to obtain the final classification result.

[0019] In one embodiment of the present invention, when weighting the feature coefficients of the training set images, the data from different periods have their own weights, and the weighting relationship depends on the similarity between the images from each period and the oracle bone script images. Specifically, a weighting factor is introduced into the loss function of ResNet-50, and the value of the weighting factor is modified according to the similarity between the font in each image and the oracle bone script, thereby changing the weight of the sample in the weighted loss function. For the text evolution image classification model, the formula of its sample weighted loss function is as follows:

[0020]

[0021] Where N is the total number of samples, C is the number of categories, and ω i It is the weight of the i-th sample, y i,c It is the true label of the i-th sample belonging to category c. It is the model's predicted probability for the i-th sample with respect to class c.

[0022] This application also provides an oracle bone script-aided deciphering classification system based on a classification model, including a processor and a memory, wherein the memory stores an executable program, and the processor executes the executable program to perform the steps of the method described above.

[0023] In summary, the technical solutions conceived by this invention have the following innovative points compared with the prior art:

[0024] Innovation Point 1: Integrating artificial intelligence technology with the study of ancient Chinese characters for social and cultural heritage, this project is the first to propose the application of classification models in the field of oracle bone script deciphering. By utilizing classification models to assist in the deciphering of oracle bone script, we innovatively apply advanced artificial intelligence technology to the interpretation of ancient cultural heritage, injecting new technological power into this traditional field. This interdisciplinary collaboration not only fully utilizes the advantages of modern technology but also promotes the value of traditional culture, opening up new possibilities for ancient text research.

[0025] Innovation Point Two: In constructing a multi-period character evolution dataset, we have compiled a wide range of authoritative books, covering character styles from different historical periods, including oracle bone script, bronze script, Spring and Autumn script, Warring States script, and clerical script. The creation of this dataset is one of the key contributions of our invention. We meticulously collected data from multiple authoritative books, while also supplementing it with internet data to ensure the comprehensiveness and accuracy of the data. This rigorous and comprehensive data integration work provides a strong foundation for our invention, enabling us to more accurately train models and decipher oracle bone script. By integrating data from different sources, we have constructed a comprehensive and diverse training set, providing richer resources for oracle bone script research and laying a solid foundation for the study of ancient characters.

[0026] Innovation Point Three: For processing image data from ancient texts, an innovative chain-like file processing system was introduced. This system is used to process specific content from electronic book files by segmenting it and then categorizing the extracted portions. This system not only played a crucial role in the integration of oracle bone inscription data but also has broad applicability, applicable to the digitization of other book data. By applying this technology to the field of book digitization, we have provided an efficient and flexible solution for the digital conversion of various book contents, making a positive contribution to promoting the digital preservation and dissemination of documentary resources. This versatile application not only enriches the value of the system but also brings new opportunities and prospects to the development of the book digitization field.

[0027] Beneficial Effects: This invention utilizes an integrated oracle bone script evolution classification network. By inputting oracle bone script images into the classification model, it obtains predicted Chinese characters, thus assisting in the deciphering of oracle bone script. This invention can also serve as an oracle bone script interpretation tool. By inputting deciphered oracle bone script images, it can obtain their corresponding real labels, which will also help oracle bone script experts accurately identify some deciphered oracle bone scripts. The relatively complete character evolution dataset constructed by this invention can also be used by subsequent researchers working on oracle bone script / Chinese character evolution. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the logic flow of an embodiment of the present invention;

[0029] Figure 2 This is a flowchart of the classification-based oracle bone script deciphering and classification method according to an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of the text evolution dataset according to an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of the classification model structure according to an embodiment of the present invention;

[0032] Figure 5 This is a schematic diagram of a deep residual convolutional neural network according to an embodiment of the present invention;

[0033] Figure 6 This is a structural block diagram of the oracle bone script deciphering classification system based on a classification model, according to an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0035] The deciphering of oracle bone inscriptions is fraught with challenges, but existing computer science and artificial neural network technologies can help solve some of the problems encountered in manual deciphering. With the increasing sophistication of classification networks, from VGG to ResNet and later Vit, the classification performance has improved significantly. This invention, in connection with the development of paleography, attempts to apply classification networks to the field of ancient character deciphering. Utilizing data from deciphered oracle bone inscriptions, this invention proposes for the first time a multi-classification system for Chinese characters based on the similarity of images depicting the evolution of Chinese characters across multiple periods.

[0036] The application of this invention in the deciphering of oracle bone inscriptions offers some feasible solutions for this work and provides technical support for promoting the creative transformation and innovative development of outstanding traditional Chinese culture. This project adopts a multidisciplinary approach, providing feasible directions for the study of ancient scripts worldwide and offering important references for exploring the remnants of human history.

[0037] To address the problems of existing technologies, this invention specifically proposes an oracle bone script classification-assisted deciphering system based on a character evolution dataset. Our goal is to provide an oracle bone script classification-assisted deciphering tool applicable to various systems, facilitating the retrieval of oracle bone script characters by oracle bone scholars and paleographers. For some undeciphered oracle bone script characters, we hope to use this classification network to help paleographers narrow down their research scope and more accurately find the modern Chinese characters corresponding to some unknown characters. For example... Figure 2 The diagram shows a flowchart of the oracle bone script classification-assisted deciphering process based on a classification model. It illustrates four parts, corresponding to the four steps: dataset, classification network, UI window, and assisted deciphering analysis. The first part, the dataset module, is the core of this invention. We conducted extensive integration work based on data from books and websites, resulting in a relatively complete oracle bone script classification dataset. This dataset forms the basis for our subsequent experiments. Figure 3 The first part shows the main format of the dataset. The second part is the classification model. Figure 4 This is a schematic diagram illustrating the optimization of the classification model. The third part is a window UI designed based on the classification model obtained in the second step, enabling paleographers and other groups with relatively weak computer skills to better use this classification model. Finally, in the fourth part, based on the output of the classification module, we can obtain a label for an unknown oracle bone image (undeciphered oracle bone script). Figure 5 This demonstrates the specific structure of the deep residual convolutional neural network in this embodiment. This invention proposes an oracle bone script classification-assisted deciphering system based on a character evolution dataset, specifically as follows... Figure 1 As shown, it includes the following steps:

[0038] Step 1: Obtain the data source and divide the dataset based on it. First, collect over 9000 sets of images that trace the evolution of Chinese character font styles, excluding variant characters, lost characters, and compound characters, to obtain the character evolution dataset. Next, determine the training set, test set, and validation set of the dataset, including: selecting one-tenth of the font images from each period that contain oracle bone script images as the test set to test the model's generalization ability. When selecting the test set, remove all oracle bone script images in the dataset that correspond to the same Chinese character but have a character shape difference greater than a preset level. The remaining image data after removing the test set from the character evolution dataset is the training set. In the determined training set, randomly select a portion of images from each group of Chinese character images according to a preset ratio to form the validation set, where the validation set contains character shape evolution images from different periods.

[0039] To obtain these images depicting the evolution of Chinese characters, image data from various periods, sourced from books and websites, was integrated. Simultaneously, a chain-like file processing system was established to extract and process the Chinese character images. Using edge detection and border extraction methods, images related to font evolution were extracted from the electronic book files and saved according to the corresponding Chinese character categories. Figure 3 An example of a collected dataset on the evolution of characters is shown.

[0040] Considering that some images in ancient texts may contain significant noise and large areas of black pixels, the dataset needed to be cleaned. Specifically, a denoising network was used for noise removal. For noise that could not be removed by the denoising network, manual processing was performed. Ultimately, a relatively clean dataset of text evolution was obtained. Furthermore, a certain degree of rotation transformation was applied to the oracle bone script dataset to further expand the classification dataset.

[0041] Step 2: Train the classification model, which includes: inputting the training set into the training model for training; during the training process, weighting the feature coefficients of Chinese character images from different periods in each group of images in the training set; after each round of training, outputting the model performance evaluation parameters of the validation set; and selecting the classification model based on the quality of these parameters.

[0042] Finally, the images in the test set are unlabeled and placed as undeciphered oracle bone script characters in the classification model for prediction, resulting in a prediction accuracy higher than a preset value.

[0043] Specifically, the main function of the classification model is to output a predicted label for each input undeciphered oracle bone script image. When inputting training images, the true labels are represented by y. n This indicates that the image has features of different dimensions {x1, x2, ..., x...}. iThese features are input into the classification model algorithm Ψ({x1,x2,……,x}). i In}), the corresponding prediction results are generated. By and y n The loss function L, used as the independent variable, measures the error and can optimize the classifier parameters, thereby obtaining a feature space corresponding to a specific category. Through the prediction algorithm in the classifier, each image can obtain a corresponding predicted label.

[0044] This invention employs a ResNet-based deep convolutional neural network as a classification model to train a text evolution dataset. The structure of this deep residual convolutional neural network is as follows: Figure 5 As shown, the specific structure is as follows:

[0045] a) Input a grayscale image Z∈R with dimensions of 128×128 for the text evolution image. 128×128×1 To ensure a fixed input image size, a size planning module called "Resize module" is designed. For non-standard square images of arbitrary size, the module can scale the image to the target format. For non-square images with a format of a×b×1, the module uses the smallest square to include the border, fills the blanks, and obtains a square with a side length of MAX{a,b}, which is then scaled to the target format.

[0046] b) The image is processed through an initial convolutional layer and a pooling layer. The detailed process is as follows: the input image is downsampled through a 7×7 convolutional layer with a stride of 2, resulting in 64 feature maps. Subsequently, a 3×3 Max-Pooling layer is applied to reduce the image size by half.

[0047] c) The image enters four residual stages, each consisting of 3, 4, 6, and 3 residual blocks respectively. Each block contains multiple residual units. Each residual block consists of two 3×3 convolutional layers, and includes batch normalization and the ReLU non-linear activation function. Within each residual block, there is a skip connection that directly passes the input to the output of the residual block, thus allowing gradient propagation without introducing additional parameters.

[0048] d) The output of the last residual block is globally averaged using a global average pooling layer, reducing the feature map size to 1×1. Finally, the output of the global average pooling layer is connected to the model prediction result through a fully connected layer to obtain the final classification prediction result.

[0049] like Figure 4As shown, in the training model, feature coefficients are weighted for Chinese character images from different periods in each group of images in the dataset. Considering that characters from periods such as bronze inscriptions, Spring and Autumn period characters, and Warring States period characters have a higher similarity to oracle bone inscriptions in the evolution images from different periods, these ancient characters are given higher weights. In specific implementation, a weight factor is introduced into the loss function of ResNet-50, and the value of the weight factor is modified according to the similarity between the character in each image and the oracle bone inscription, thereby changing the weight of the sample in the weighted loss function. For the multi-classification problem of character evolution images, the formula of its sample weighted cross-entropy loss function is as follows:

[0050]

[0051] Where N is the total number of samples, C is the number of categories, and ω i It is the weight of the i-th sample, y i,c It is the true label of the i-th sample corresponding to category c (1 indicates belonging to this category, 0 indicates not belonging to this category). This represents the model's predicted probability for the i-th sample in class c. In the experiments, fonts closer to the oracle bone script period were given higher weights. By adjusting the weights for different periods, the classification network can function in various application scenarios. For example, by increasing the weights of fonts like regular script, the classification network can be considered a Chinese character OCR recognizer. Furthermore, the optimal Chinese character classification model was selected and saved based on parameters such as training loss and validation loss on the validation set.

[0052] After training, the images in the test set were used as "undeciphered oracle bone script images" and fed into the Chinese character classification model for prediction. The final output was the top 1, top 5, and top 20 accuracy rates for the test set. On a relatively small test set, a 50% Top 1 accuracy rate, a 70% Top 5 accuracy rate, and an 85% Top 20 accuracy rate were achieved.

[0053] Step 3: Based on a classification model, classify and assist in the deciphering of undeciphered oracle bone script. This includes building an interactive prediction interface, inputting the undeciphered oracle bone script into the interface, and then outputting the possible corresponding Chinese characters and their probabilities. Finally, based on the similarity between the output Chinese characters and the oracle bone script, the corresponding Chinese character is determined. If the top-1 probability in the prediction results is greater than 0.3, then the unknown oracle bone script is very likely to correspond to that Chinese character.

[0054] This invention is the first to propose using a classification model to assist in deciphering oracle bone inscriptions, combining artificial intelligence technology with social and cultural heritage. For thousands of undeciphered oracle bone inscriptions, several possible results can be obtained in the character evolution dataset classification model. This will help oracle bone experts narrow down the scope of ancient texts they study and help complete the deciphering of oracle bone inscriptions more quickly.

[0055] Based on the same inventive concept, this invention also provides a classification system for oracle bone script deciphering based on a classification model, such as... Figure 6 As shown, it includes a processor and a memory, the memory storing an executable program, and the processor executing the executable program to perform the steps of the method described above.

[0056] The foregoing has presented and described the basic principles, main features, and advantages of this invention. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this invention, and various changes and modifications can be made without departing from the core ideas and scope of this invention. These changes and modifications all fall within the scope of protection of this invention. The specific scope of protection is determined by the appended claims and their equivalents.

Claims

1. A classification-based method for oracle bone script deciphering, characterized in that, Includes the following steps: Step one, obtain the data source, and partition the dataset based on the data source, including: Collect images of Chinese characters from various periods of style evolution, remove variant characters, dead characters, and compound characters to obtain a character evolution dataset; The determination of the training set, test set, and validation set of the dataset includes: selecting one-tenth of the oracle bone script images from the categories containing oracle bone script images in the collected font images from each period as the test set to test the model's generalization ability; when selecting the test set, removing all oracle bone script images in the dataset that correspond to the same Chinese character but have character shape differences exceeding a preset difference level; the remaining image data after removing the test set from the character evolution dataset is the training set; in the determined training set, randomly selecting a portion of images from each group of Chinese character images according to a preset ratio to form the validation set, wherein the validation set contains character shape evolution images from different periods; Step two involves training a classification model using the character evolution dataset. This includes: inputting the training set into the training model; weighting the feature coefficients of Chinese character images from different periods in each group of images in the training set during training; outputting model performance evaluation parameters for the validation set after each round of training; selecting the classification model based on the quality of these parameters; and finally, removing the labels from the images in the test set and placing them as undeciphered oracle bone script characters into the classification model for prediction, outputting a number of results with a prediction accuracy higher than a preset value. Step 3: Based on the classification model, classify and assist in the deciphering of the undeciphered oracle bone script, including: constructing an interactive prediction interface, inputting the undeciphered oracle bone script into the interactive prediction interface, outputting the possible Chinese characters corresponding to the undeciphered oracle bone script and the corresponding similarity; finally, determining the Chinese characters corresponding to the undeciphered oracle bone script based on the similarity between the output Chinese characters and the oracle bone script.

2. The oracle bone script-aided deciphering classification method based on a classification model according to claim 1, characterized in that, The font images in the character evolution dataset are all constructed based on at least one of the following data sources: books associated with oracle bone script, books associated with bronze script, Spring and Autumn script, Warring States script, and clerical script, and Internet data, wherein the bronze script images are similar to the oracle bone script images.

3. The classification-based oracle bone script deciphering method according to claim 1 or 2, characterized in that, Some images in the text evolution dataset contain noise and large areas of black pixels. A noise reduction network is used to clean the dataset. For noise that cannot be removed by the noise reduction network, manual removal is chosen to obtain a clean text evolution dataset.

4. The classification-based oracle bone script deciphering method according to claim 2, wherein, The character evolution dataset contains at least 9,000 sets of evolution images corresponding to Chinese characters. Among these 9,000 sets of images, there are 5,000 evolution images of commonly used Chinese characters, as well as a certain number of rare Chinese characters. The 9,000 sets of Chinese character images are used to train an oracle bone script deciphering classification network.

5. The classification-based oracle bone script deciphering method according to claim 1 or 2, characterized in that, The specific steps for training a classification model using the aforementioned character evolution dataset are as follows: The true label of the input training image is represented by y. n Features with different dimensions {x1,x2,……,x i These features are input into the classification model Ψ({x1,x2,……,x}). i In the process of generating prediction results, the corresponding prediction results are produced. By and y n The weighted loss function L, used as the independent variable, measures the error; the error is reduced by optimizing the classification model parameters, thereby obtaining the feature space corresponding to each category; during the testing phase, the classification model can output the corresponding predicted label based on the input undeciphered oracle bone image.

6. The oracle bone script-aided deciphering classification method based on a classification model according to claim 5, characterized in that, The classification model used is ResNet, a deep convolutional neural network based on residual blocks. The structure of this deep convolutional neural network is as follows: The initial training image Z∈R is a grayscale image with a size of 128×128. 128×128×1 First, the image is processed through an initial convolutional layer and a pooling layer. That is, the image is first downsampled through a 7×7 ConV layer with a stride of 2 to obtain 64 feature maps. Then, the feature maps are passed through a 3×3 Max-Pooling layer to reduce the image size by half. The image then enters four residual stages, each stage consisting of 3, 4, 6, and 3 residual blocks respectively. Each block contains multiple residual units. Each residual block consists of two 3×3 convolutional layers and includes batch normalization and ReLU nonlinear activation functions. Within each residual block, there is a skip connection that directly passes the input to the output of the residual block. Next, the output of the last residual block is subjected to global average pooling through a global average pooling layer, reducing the feature map size to 1×1; Finally, the output of the global average pooling layer is connected to the prediction result through a fully connected layer to obtain the final prediction result.

7. The oracle bone script-aided deciphering classification method based on a classification model as described in claim 1, characterized in that, In step two, when weighting the feature coefficients of the training set images, different data have their own weights. The magnitude of the weight depends on the similarity between the image and the oracle bone script at each time period. Specifically, a weight factor is introduced into the loss function of ResNet-50. The value of the weight factor is modified according to the similarity between the font and the oracle bone script in each image, thereby changing the weight of the sample in the weighted loss function. For the image classification model of text evolution, the formula of its sample weighted loss function is as follows: Where N is the total number of samples, C is the number of categories, and ω i It is the weight of the i-th sample, y i,c It is the true label of the i-th sample belonging to category c. It is the model's predicted probability for the i-th sample with respect to class c.

8. A classification system for oracle bone script deciphering based on a classification model, characterized in that, It includes a processor and a memory, wherein the memory stores an executable program, and the processor executes the executable program to perform the steps of the classification model-based oracle bone script deciphering classification method as described in any one of claims 1 to 7.

Images