A method for recognizing Qin dynasty bamboo slip text

Abstract

The application provides a Qin Dynasty bamboo slip character text recognition method, which recognizes the predicted text ideograph of unknown Qin Dynasty bamboo slip character text through a trained text recognition model, and the text recognition model is trained based on a training data set; a simulated Qin Dynasty bamboo slip character text data set is used in the training process of the text recognition model, and sample Qin Dynasty bamboo slip character texts in the simulated Qin Dynasty bamboo slip character text data set are simulated based on a simulated Qin Dynasty bamboo slip character text generation model; the simulated Qin Dynasty bamboo slip character text generation model comprises a word / sentence generation module and an image splicing module, the word / sentence generation module is used for generating simulated words / sentences conforming to the corresponding Qin Dynasty bamboo slip character text order, and the image splicing module is used for generating sample Qin Dynasty bamboo slip character texts according to the simulated words / sentences. Under the training of a large amount of simulated data with high authenticity, the text recognition model has good Qin Dynasty bamboo slip character text recognition effect, and the Qin Dynasty bamboo slip character text recognition method has good recognition accuracy in practical application.

Description

Technical Field

[0001] This invention relates to the field of ancient Chinese translation, specifically to a method for recognizing Qin Dynasty bamboo slips. Background Technology

[0002] Ancient scripts from past dynasties are typically recorded on physical carriers such as books, steles, and vessels. These carriers often deteriorate over time, resulting in damage to the ancient scripts as well. In the field of ancient script translation, the process generally involves first restoring the damaged scripts to their original, complete form through rubbings, and then identifying the meaning of the characters. This process encompasses two steps: script restoration and meaning recognition.

[0003] In practical applications, most ancient script usage scenarios require the direct identification of the meaning of damaged ancient characters. Since ancient characters are generally found within written texts, the real need is to directly identify the meaning of multiple consecutive damaged ancient characters. Therefore, based on these practical requirements, when constructing relevant text recognition models, the input data format is typically limited to Qin Dynasty bamboo slip text, and the model's output is the corresponding meaning of the Qin Dynasty bamboo slip text.

[0004] In practical applications, once the structure of a text recognition model is determined, it needs to be trained with a sufficient amount of training data to achieve good text meaning recognition results. However, the number of ancient script samples from many dynasties is relatively small, insufficient to support the training of the text recognition model. Therefore, appropriate methods are needed to expand the number of samples to meet training requirements and ensure that the trained text recognition model can achieve good text recognition results for Qin bamboo slips. Summary of the Invention

[0005] This invention provides a method for recognizing Qin bamboo slip text. The method uses a trained text recognition model to identify the predicted meaning of unknown Qin bamboo slip text. Most of the training data used in the training process of the text recognition model comes from a dataset of simulated Qin bamboo slip text. With sufficient and highly authentic simulated data for training, the text recognition model has a good effect on recognizing Qin bamboo slip text.

[0006] Accordingly, this invention provides a method for recognizing Qin Dynasty bamboo slip text, which uses a trained text recognition model to identify the predicted meaning of unknown Qin Dynasty bamboo slip text. The text recognition model is based on a training dataset. Trained;

[0007] The training dataset ,in, This is a dataset of authentic Qin Dynasty bamboo slip texts. To replicate the Qin Dynasty bamboo slips text dataset, For the number The sample Qin bamboo slip text data, each of the sample Qin bamboo slip text data includes the sample Qin bamboo slip text and the corresponding text meaning, and the data form of the sample Qin bamboo slip text is a grayscale image with the same height.

[0008] In the real Qin bamboo slip text dataset, the sample Qin bamboo slip text in the sample Qin bamboo slip text data is obtained by extracting real Qin bamboo slip text;

[0009] In the simulated Qin bamboo slip text dataset, the sample Qin bamboo slip texts in the sample Qin bamboo slip text data are based on a simulated Qin bamboo slip text generation model;

[0010] The simulated Qin bamboo slip text generation model includes a word and sentence creation module and an image stitching module. The word and sentence creation module is used to generate simulated words and sentences that conform to the word order of Qin bamboo slip text. The image stitching module is used to generate sample Qin bamboo slip text that meets the format of the sample Qin bamboo slip text data based on the simulated words and sentences.

[0011] In an optional implementation, the training process of the word-creation module includes:

[0012] S301: Based on the extractor, extract several sub-texts of real Qin bamboo slips and their corresponding meanings from the real Qin bamboo slip text and store them in a real Qin bamboo slip text sub-text database.

[0013] S302: Extract the meanings of individual characters and words from the actual Qin bamboo slip text, and store them in a word list in sequence;

[0014] S303: Randomly extract a subtext of the real Qin bamboo and wooden characters from the real Qin bamboo and wooden characters subtext database, and convert the meaning of the real Qin bamboo and wooden characters subtext into an index sequence according to the number in the word list;

[0015] S304: Map the index sequence into a low-dimensional dense vector based on the Embedding layer;

[0016] S305: Perform time-by-time processing on the low-dimensional dense vector based on the LSTM layer to obtain the hidden states of all time parts in the low-dimensional dense vector;

[0017] S306: Based on a fully connected layer, map the hidden states of all time parts corresponding to the low-dimensional dense vector to the probability distribution of the word list;

[0018] S307: Compare the loss between the probability distribution and the true label, and adjust the LSTM layer, Embedding layer and fully connected layer through backpropagation;

[0019] S308: Repeat steps S303 to S307 until the word-creation module is trained.

[0020] In an optional implementation, the process of generating the simulated phrases from the simulated Qin bamboo slips text includes:

[0021] S201: Construct short sentences within the range of single characters and words included in the vocabulary list;

[0022] S202: Input the short sentence as the first text into the Embedding layer and select the corresponding single character or word as the second text from the word list according to the probability distribution output by the fully connected layer;

[0023] S203: Combine the first text and the second text to form a new short sentence;

[0024] S204: Repeat steps S202 and S203 until the length of the short sentence in step S203 reaches the preset word count, and the final generated short sentence is the required imitation sentence.

[0025] In an optional implementation, the image stitching module is used to generate sample Qin bamboo text that satisfies the format of the sample Qin bamboo text data based on the imitated phrases, including:

[0026] Extract individual Qin bamboo slip characters from the authentic Qin bamboo slip text and store them along with their meanings in the corresponding Qin bamboo slip character database.

[0027] Based on the fabricated phrases, several corresponding Qin bamboo slip characters are extracted from the Qin bamboo slip character database and spliced ​​together to form a spliced ​​Qin bamboo slip text.

[0028] The spliced ​​Qin bamboo slip text is processed according to the corresponding format to generate sample Qin bamboo slip text that meets the data format of the sample Qin bamboo slip text.

[0029] In optional implementations, the stitching methods include vertical stitching, stitching with CLAHE, stitching with wavelet filtering, stitching with gradient fusion, stitching with image normalization, and stitching with feathering.

[0030] In an optional implementation, the text recognition model training process includes:

[0031] A text recognition model is constructed, comprising a feature extraction layer, a recurrent layer, and a transcription layer linked sequentially.

[0032] Text recognition model training, using the training dataset The text recognition model is iteratively trained using sample Qin Dynasty bamboo slip text data.

[0033] In an optional implementation, the feature extraction layer includes a DRCA module, a STEM module, an Inception-Resnet module, a convolution module, and an adaptive pooling module linked in sequence.

[0034] In an optional implementation, the recurrent layer includes a BiLSTM module and a Transformer Encoder module.

[0035] In an optional implementation, the transcription layer includes a CTCFL-LS module.

[0036] In an optional implementation, the text recognition model uses the Adam optimizer and cosine annealing algorithm as the learning rate adjustment algorithm during training.

[0037] In summary, this invention provides a method for recognizing Qin Dynasty bamboo slips. This method is based on a text recognition model. During the training process of the text recognition model, a large amount of training data is constructed through imitation, which helps to improve the recognition accuracy of the text recognition model. The process of imitating the training data can fully incorporate the writing habits of ancient Chinese word order, which helps to improve the authenticity of the training data, thereby ensuring the recognition ability of the text recognition model. It has good practicality in practical applications. Attached Figure Description

[0038] Figure 1 This is a flowchart of the Qin Dynasty bamboo slips text recognition method according to Embodiment 1 of the present invention.

[0039] Figure 2 This is a flowchart of the text recognition model training process in Embodiment Six of the present invention.

[0040] Figure 3 This is a schematic diagram of the text recognition model structure in Embodiment Six of the present invention.

[0041] Figure 4 This is a schematic diagram of the DRCA module structure in Embodiment 7 of the present invention.

[0042] Figure 5 This is a schematic diagram of the STEM module structure in Embodiment 7 of the present invention.

[0043] Figure 6 This is a schematic diagram of the Inception-Resnet module structure in Embodiment 7 of the present invention.

[0044] Figure 7 This is a schematic diagram of the Reduction-A Block unit structure in Embodiment 7 of the present invention.

[0045] Figure 8This is a schematic diagram of the Reduction-B Block unit structure in Embodiment 7 of the present invention.

[0046] Figure 9 This is a schematic diagram of the CPA structure in Embodiment 7 of the present invention.

[0047] Figure 10 This is a schematic diagram of the first example of the Qin Dynasty bamboo slip text recognition method according to an embodiment of the present invention.

[0048] Figure 11 This is a schematic diagram of a second example of the Qin Dynasty bamboo slip text recognition method according to an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] It should be noted that, since the Qin bamboo slip text recognition method in this embodiment of the invention involves multiple types of data, in order to avoid ambiguity, a basic explanation of each type of data will be given first.

[0051] Specifically, text refers to words or phrases composed of two or more single characters. The data form of text is a grayscale image. Ideographs refer to the corresponding meanings of ancient characters in modern Chinese. Text ideographs refer to the corresponding meanings of words or phrases composed of ancient characters in modern Chinese.

[0052] Example 1

[0053] Figure 1 This is a flowchart of the Qin Dynasty bamboo slips text recognition method according to an embodiment of the present invention.

[0054] This invention discloses a method for recognizing Qin Dynasty bamboo slips, including steps S101 and S102.

[0055] Step S101: Input an unknown Qin Dynasty bamboo slip text into the text recognition model.

[0056] Specifically, unknown Qin bamboo slip text is generally extracted from physical carriers such as books and inscriptions. Correspondingly, the data form of the unknown Qin bamboo slip text identified by the Qin bamboo slip text recognition method in this embodiment of the invention is an image. Based on the actual situation of Qin bamboo slip text (the content consists of two or more single-character phrases) and the convenience of data processing for the text recognition model, the specific form of the image is a grayscale photo with a fixed height (width is not limited). This involves the process of extracting the unknown Qin bamboo slip text from the physical carrier and converting it into a clear binary photo. This process can be implemented based on existing technology, and this invention will not provide further explanation.

[0057] It should be noted that the process of extracting the unknown Qin bamboo slip text from the physical carrier and converting it into a clear binary image should be as consistent as possible with the method of generating the training data used by the text recognition model during training. That is, the image of the unknown Qin bamboo slip text and the image of the training data should have similar data features to improve the recognition accuracy of the text recognition model for the unknown Qin bamboo slip text.

[0058] Step S102: The text recognition model outputs the predicted meaning of the text.

[0059] The text recognition model outputs the predicted meaning of the text based on the input of unknown Qin bamboo slip text.

[0060] Example 2

[0061] The Qin bamboo slip text recognition method disclosed in Example 1 requires the use of a trained text recognition model. The basic training process of this text recognition model includes:

[0062] S201: Initialize the text recognition model;

[0063] Basically, the text recognition model of this invention includes a feature extraction layer, a recurrent layer, and a transcription layer linked in sequence. The feature extraction layer is used to extract feature information from the input source data (Qin bamboo slip text in grayscale image form) and output a feature sequence with a height of 1. The recurrent layer is used to combine the feature sequence and the source data to learn contextual information. The transcription layer is used to output the prediction result.

[0064] S301: Text recognition model training;

[0065] Specifically, the training method of the text recognition model is the same as that of conventional neural network training methods, machine learning training methods, and other similar training methods. Existing data (i.e. sample Qin bamboo slip text) that is consistent with the input data in the actual application (i.e., unknown Qin bamboo slip text) is input into the text recognition model. The text recognition model obtains the predicted meaning of the text. The generation effect of the predicted meaning of the text is evaluated according to the defined loss function. After adjusting the parameters of the text recognition model using backpropagation, iterative testing of sample Qin bamboo slip text is continued until the number of training times reaches a preset value or the loss function reaches an ideal state.

[0066] After training, the text recognition model can accurately predict the meaning of the text based on the input unknown Qin bamboo slip text.

[0067] It should be noted that current methods for validating the effectiveness of text recognition models generally involve batch input of data and then verifying the accuracy of the model in predicting the meaning of the text. Typically, researchers divide collected sample data into training and validation data. The training data is used to train the text recognition model, and the validation data is used for validation. Both training and validation data originate from the same database and exhibit high similarity. Therefore, in practice, text recognition models trained on the training data generally show high accuracy during validation. However, many existing technical papers suffer from issues such as unreasonable sample data. Specifically, in the field of ancient scripts, due to the lack of sample data, it is often necessary to use methods of fabrication to expand the sample data, and the amount of expanded data far exceeds the original sample data. If an unreasonable expansion method is used, although the final text recognition model may have a high accuracy, this is because the validation and training data are similar, and the recognition accuracy of the text recognition model cannot accurately reflect its actual performance. Similarly, the text recognition model disclosed in this invention is limited by the amount of data and has a similar situation. Therefore, in the embodiments of this invention, the focus is on the means of sample data expansion to improve the quality of sample data. The technical content other than the means of sample data expansion can be implemented according to the existing technology. On the other hand, the embodiments of this invention also provide corresponding reference implementation methods for reference.

[0068] The specific structure of the text recognition model will be described in subsequent embodiments.

[0069] In this embodiment of the invention, the training dataset ,in, This is a dataset of authentic Qin Dynasty bamboo slip texts. To replicate the Qin Dynasty bamboo slips text dataset, For the number The sample Qin bamboo slip text data, each of the sample Qin bamboo slip text data includes the sample Qin bamboo slip text and the corresponding text meaning, and the data form of the sample Qin bamboo slip text is a grayscale image with the same height.

[0070] Specifically, in the real Qin bamboo slips text dataset In the sample Qin bamboo slip text data, the sample Qin bamboo slip text is obtained by extracting authentic Qin bamboo slip text; in the replica Qin bamboo slip text dataset... In the sample Qin bamboo slip text data, the sample Qin bamboo slip text is based on a Qin bamboo slip text generation model.

[0071] For some ancient scripts from more distant periods, due to the scarcity of surviving reference texts, it is necessary to generate simulated sample data based on existing real texts in order to meet the training requirements of the text recognition model. Details regarding simulated sample data will be explained in subsequent embodiments.

[0072] With the support of the training dataset, the text recognition model is iteratively trained using sample Qin bamboo slip text data in the training dataset until training is complete.

[0073] Example 3

[0074] Specifically, in the description of Embodiment 2, the sample Qin bamboo slip text in the sample Qin bamboo slip text data is implemented based on a Qin bamboo slip text generation model.

[0075] In this embodiment of the invention, the simulated Qin bamboo slip text generation model includes a word and sentence creation module and an image stitching module. The word and sentence creation module is used to generate simulated words and sentences that conform to the word order of the Qin bamboo slip text. The image stitching module is used to generate sample Qin bamboo slip text that satisfies the data format of the sample Qin bamboo slip text based on the simulated words and sentences.

[0076] The training process of the word-construction module includes:

[0077] S301: Based on the extractor, extract several sub-texts of real Qin bamboo slips and their corresponding meanings from the real Qin bamboo slip text and store them in a real Qin bamboo slip text sub-text database.

[0078] Specifically, the extractor can be understood as a screenshot tool. Any consecutive words or sentences in the real Qin bamboo slips text have continuity in collocation or meaning, which can show the writing habits of articles in the corresponding era. Therefore, a certain number of words or sentences can be extracted from the real Qin bamboo slips text with continuous text as the object. Specifically, as a screenshot tool, the extractor can extract consecutive words or sentences in the real Qin bamboo slips text by means of manual calibration. In addition, it is also possible to use the blank interval feature between characters, and based on the edge feature of the characters, each ancient character is intelligently divided, and then according to the randomly determined number to determine the length of the words or sentences to be extracted, the real Qin bamboo slips text is intercepted to obtain the real Qin bamboo slips sub-text.

[0079] Each real Qin bamboo slips sub-text extracted is stored in a real Qin bamboo slips sub-text database. Correspondingly, each real Qin bamboo slips sub-text has a calibrated meaning.

[0080] For the reference of subsequent steps, assume that the meaning of one of the real Qin bamboo slips sub-texts is "To be an official, one must be capable".

[0081] S302: Extract the meanings of single characters and words in the real Qin bamboo slips text, and store the corresponding meanings in a word list in sequential order.

[0082] Compared with the extraction in step S301, the extraction object in this step is single characters and words, and in the storage method of the single characters or words in the word list, the meanings need to be numbered with corresponding index numbers in sequential order to facilitate subsequent data processing.

[0083] For the sake of easy understanding, assume that the structure of the word list is {0: "[START]", 1: " ", 2: "为吏", 3: "要", 4: "能"}, separated by commas, the index number 0 is the start tag, the index number 1 is the end tag, and correspondingly, the words corresponding to the index numbers 2, 3, and 4 are "为吏", "要", and "能" respectively.

[0084] S303: Randomly extract one of the real Qin bamboo slips sub-texts from the real Qin bamboo slips sub-text database, and convert the real Qin bamboo slips sub-text into an index sequence according to the serial numbers corresponding to the single characters or words in the word list.

[0085] Specifically, assume that the meaning of the extracted real Qin bamboo slips sub-text is "To be an official, one must be capable", and correspondingly, the converted index sequence is {0, 2, 3, 4, 1}.

[0086] Specifically, the tool for extracting the real Qin bamboo slips sub-text and converting it into an index sequence can be named a tokenizer.

[0087] S304: Map the index sequence into a low-dimensional dense vector based on the Embedding layer;

[0088] Specifically, the Embedding layer is a bridge connecting discrete symbols and continuous vector spaces, representing the semantics of words through low-dimensional dense vectors, and is the cornerstone of deep learning for processing text.

[0089] Specifically, if the embedding dimension of the Embedding layer is set to 3, i.e., embedding_dim = 3, the mapping content corresponding to the index sequence is shown in Table 1.

[0090] Index number words Embedding vector (example) 0 [START] [0.1, 0.2, 0.3] 1 [0.9, -0.5, 0.0] 2 official [0.4, -0.1, 0.8] 3 want [0.5, 0.3, 0.6] 4 able [0.2, 0.7, -0.4]

[0091] Table 1

[0092] Among them, the closer the Embedding vectors are, the more similar they are semantically, that is, they are interchangeable in form, which is beneficial to the sentence-making function required by the embodiments of the present invention.

[0093] S305: Perform per-time-step processing on the low-dimensional dense vector based on the LSTM layer to obtain the hidden states of all time steps in the low-dimensional dense vector and the cell state of the last time step;

[0094] The LSTM layer (Long Short-Term Memory network) is a tool that controls the flow of information through three gates (forget gate, input gate, output gate) and a cell state (Cell State). Its core function is to process sequence data (such as sentences, time series), and it can remember long-term dependencies and avoid the "gradient disappearance" problem of ordinary RNNs.

[0095] Taking the real Qin bamboo slip text sub-text "To be an official requires ability" as an example, assuming the sequence sentence input to the LSTM layer is "To be an official requires ability", the LSTM layer will split it into three time steps, and the specific form is shown in Table 2.

[0096] Time Department Input words Output target (next word) t=1 "Serving as an official" "want" t=2 "want" "able" t=3 "able"

[0097] Table 2

[0098] Per-time-step processing means processing each time step in sequence. Specifically, in the process of processing a time step, the following information is included:

[0099] Input: The word of the current time step (such as "To be an official" at t = 1);

[0100] Combined with history: Combine the current word vector with the hidden state and cell state of the previous step (initially, it can be assumed that the hidden states of the previous step are all 0);

[0101] Calculate the gating signal: Determine which information to retain and which to update through the forget gate, input gate, and output gate;

[0102] Update the memory: Generate a new cell state and hidden state;

[0103] Output: Predict the probability distribution of the next word based on the current hidden state.

[0111] ,

[0112] ,

[0110] , , ,

[0104] Correspondingly, the corresponding data flow state can be referred to Table III.

[0105] Time Department Input words <![CDATA[Hidden state (h t ) change]]> Predicted words Actual target words t=1 "Serving as an official" Initial hidden state → h1 "want" "want" t=2 "want" Calculate h2 based on h1 "able" "able" t=3 "able" Calculate h3 based on h2

[0106] Table III

[0107] Specifically, the hidden state h of each time step t Substantially contains all historical information up to the current step. For example, at time step t = 2, when processing the word "要", the context relationship between "为吏" and "要" has been encoded in the hidden state h2.

[0108] It should be noted that during training, the teacher forcing method is usually used, that is, the true previous word is input at each time step, while in actual applications, it is autoregressive or conditional autoregressive, and the word generated by the model is used as the next input. For example, when inputting "为吏要能", during the training stage, the LSTM layer processes "为吏要能" step by step for "为吏", "要", and "能" respectively; during application, the hidden state of the last word "能" can be directly taken, and the predicted word to be predicted behind it can be selected. The data finally output by the LSTM layer is the hidden state of each time step (which can be combined into vector features or matrix features).

[0109] S306: Map the hidden states of all time steps corresponding to the low-dimensional dense vector to the probability distribution in the word list based on the fully connected layer;

[0110] Map the hidden state to the score of each word in the word list, and convert the score of each word to the probability distribution of each word in the word list.

[0111] S307: Compare the loss between the probability distribution and the true label and adjust the LSTM layer, Embedding, and fully connected layer through backpropagation;

[0112] Steps S303 to S307 are an overall training process, and the function it needs to achieve is as follows: For the phrase "To be an official requires ability", when the word generation module inputs "To be an official", in the word list, the predicted confidence probability of the corresponding "requires" is the highest; in actual applications, according to other training data, there may also be words with predicted confidence probabilities close to that of the word "requires"; for example, in the real sub-text database of Qin bamboo slips characters, there are also real sub-texts of Qin bamboo slips characters with the meanings of "To be an official should be diligent", "To be an official must be clear", and "To be an official should be cautious in using punishment", then the words with relatively high predicted confidence probabilities after "To be an official" include "should", "must", and "cautiously"; ideally, if the real sub-text database of Qin bamboo slips characters only includes four real sub-texts of Qin bamboo slips characters, namely "To be an official requires ability", "To be an official should be diligent", "To be an official must be clear", and "To be an official should be cautious in using punishment", then the predicted confidence probabilities of the four words "should", "must", "cautiously", and "requires" after "To be an official" in the word list are each 25%; in the actual real sub-text database of Qin bamboo slips characters, due to the different frequencies of occurrence of each word during the training process, correspondingly, the predicted confidence probability of each word will also vary. Therefore, in actual applications, the predicted words can be sorted according to the predicted confidence probability, and this sorting order essentially reflects the writing habits of Qin bamboo slips characters at that time.

[0113] Correspondingly, when the word generation module inputs "To be an official", in the word list, the predicted confidence probability of the corresponding "requires" is the highest, and there may be words such as "should", "must", and "cautiously" that have similar meanings or functions to the word "requires" and may also have a very high predicted confidence.

[0114] Therefore, the entire training process can be summarized as

[0115] Extract real sub-texts of Qin bamboo slips characters;

[0116] The Embedding layer converts the real sub-texts of Qin bamboo slips characters into dense vectors according to the word list;

[0117] The LSTM layer combines historical memories to update the current memory;

[0118] The fully connected layer corrects the evaluation values for each word in the word list;

[0119] The Softmax layer converts the evaluation values into the confidence of each word;

[0120] For training, the context relationship of the input real Qin bamboo slip text sub-text is known. The training objective of the model is to make the probability distribution predicted by the model approximate the real next word (for example, during the training process of training the model with "To be an official requires ability", the real word to be predicted after "To be an official requires" is "ability". We hope that the confidence of the ideal probability distribution of the word "ability" is 1, and that of other words is 0); Therefore, in actual training, generally guided by the cross-entropy loss, the parameters of the LSTM, Embedding, and fully connected layers are adjusted through backpropagation to improve the prediction accuracy of the model.

[0121] S308: Repeat steps S303 to S307 until the sentence creation module is trained.

[0122] Correspondingly, the structure of the sentence creation module includes a tokenizer, an Embedding layer, an LSTM layer, a fully connected layer, and a Softmax layer connected in sequence. The use of an extractor is also involved, as well as the auxiliary use of the real Qin bamboo slip text sub-text database and the word list. The sentence creation module with this structure adopts a single-layer LSTM structure, and the number of parameters used in the fully connected layer is small, which is suitable for the small data set environment of the embodiments of the present invention; LSTM has high interpretability for the gradual generation logic of text, and can very conveniently debug the intermediate state data during the training process or the generation process; by reasonably setting the two parameters of hidden_dim and embedding_dim therein, the performance and speed of the module can be balanced, and it has good adjustment convenience.

[0123] It should be noted that the overall structure composed of the tokenizer, the Embedding layer, the LSTM layer, the fully connected layer, and the Softmax layer (named the sentence creation module in the present invention) belongs to the category of existing technologies and is currently mainly applied in the field of AI dialogue robots to make the writing habits of AI robots closer to the daily expression habits of humans; Since this part of the content also involves relevant technical details, it can be implemented by referring to the existing technologies in specific implementations. The embodiments of the present invention only describe the main content involved therein. The embodiments of the present invention use the sentence creation module to realize the text sentence creation of Qin bamboo slip characters, thereby expanding the text volume of Qin bamboo slip characters. The expanded text has good simulation, and combined with the subsequent image splicing module, it can achieve a better simulation effect.

[0124] Embodiment Four

[0125] Specifically, according to the training process listed in Embodiment Three, the trained sentence creation module has the characteristic that when a word sentence is input into the sentence creation module, the sentence creation module displays the confidence of each word in the word list. Correspondingly, the process of generating the imitation sentence by imitating the Qin bamboo slip text in the embodiments of the present invention includes:

[0126] S401: Construct short sentences within the range of single characters and words included in the vocabulary list;

[0127] It is important to note that short sentences should be semantically meaningful and correspond to the era in which the ancient script was written, rather than being arbitrary combinations of words.

[0128] S402: Input the short sentence as the first text into the Embedding layer and select the corresponding single character or word as the second text from the word list according to the probability distribution output by the fully connected layer;

[0129] It should be noted that, since the function to be achieved in this embodiment of the invention is sentence creation, and tends to create words and sentences that do not exist in the real Qin bamboo slips subtext database, the corresponding single character or word is selected as the subtext in the word table according to the probability distribution output by the fully connected layer. There may be single characters or words with the second highest confidence as the subtext.

[0130] S403: Combine the first text and the second text to form a new short sentence;

[0131] In this step, the first and second texts are combined to form a new short sentence;

[0132] S404: Repeat steps S402 and S403 until the length of the short sentence in step S403 reaches the preset word count, and the final generated short sentence is the required imitation sentence.

[0133] The new short sentence formed in step S403 is brought back to step S402, which can extend the short sentence. According to the requirements (generally the number of words), the final short sentence is output as a simulated sentence.

[0134] Example 5

[0135] After generating the required simulated phrases in Example 4, the image stitching module is used to generate sample Qin bamboo slip text that satisfies the sample Qin bamboo slip text data format based on the simulated phrases, including:

[0136] S501: Extract individual Qin bamboo slip characters from the actual Qin bamboo slip text and store them along with their meanings in the corresponding Qin bamboo slip character database;

[0137] The data format of individual characters in Qin bamboo slips is grayscale; it should be noted that individual characters in Qin bamboo slips have a background, not just the text part, and the background content also retains certain texture features of the physical carrier.

[0138] S502: Based on the fabricated phrases, extract several corresponding Qin bamboo slip characters from the Qin bamboo slip character database and splice them together to form a spliced ​​Qin bamboo slip text.

[0139] Specifically, the splicing methods include vertical splicing, splicing with CLAHE, splicing with wavelet filtering, splicing with gradient fusion, splicing with image normalization, and splicing with feathering.

[0140] S503: Perform corresponding format processing on the spliced ​​Qin bamboo slip text to generate sample Qin bamboo slip text that meets the data format of the sample Qin bamboo slip text.

[0141] Specifically, based on the requirements of the input data of the text recognition model, the spliced ​​data is processed in the corresponding format, such as adjusting the resolution of the spliced ​​Qin bamboo slip text and modifying the contrast of the spliced ​​Qin bamboo slip text.

[0142] Example 6

[0143] Figure 2 A flowchart illustrating the text recognition model training process according to an embodiment of the present invention is shown.

[0144] Figure 3 A schematic diagram of the text recognition model structure according to an embodiment of the present invention is shown.

[0145] Specifically, the training process of the text recognition model in this embodiment of the invention, based on the training dataset, includes:

[0146] S601: Text recognition model construction;

[0147] Specifically, in this embodiment of the invention, the text recognition model is built based on the CRNN (Convolutional Recurrent Neural Network) deep learning model. CRNN is a deep learning model mainly used for scene text recognition, and it is particularly good at recognizing text sequences in images that are of variable length and irregularly arranged.

[0148] Basically, the text recognition model of this invention includes:

[0149] The feature extraction layer typically includes convolutional layers, pooling layers, and batch normalization layers. Its input data is Qin bamboo slip text (image). Through feature extraction, it outputs a feature map whose height is compressed to 1. The width (W') of the feature map is scaled proportionally according to the change in the height. This feature map can be regarded as a sequence of feature vectors with a width of W', and each vector has C-dimensional feature dimensions.

[0150] The recurrent layer typically includes a bidirectional long short-term memory network (BiLSTM), which is mainly used to model the feature map output by the feature extraction layer and capture the contextual dependencies (the order of text) in the sequence. Correspondingly, the input data is the feature map of the feature extraction layer (W' C-dimensional vectors), and the output data is an enhanced feature sequence (with a width of W'). The feature vector at each position in this enhanced feature sequence incorporates the contextual information of that position and its preceding and following neighbors. This sequence contains richer text features corresponding to each position from left to right in the image.

[0151] The transcription layer is used to map the enhanced feature sequence to the character label sequence to obtain the desired text recognition result. Specifically, it generally adopts the Connectionist Temporal Classification (CTC) transcription mechanism, which is an algorithm commonly used in speech recognition, text recognition and other fields to solve the problem of inconsistent input and output sequence lengths and lack of alignment.

[0152] Basically, CTC allows the model to output a sequence longer than the actual label. The output sequence can contain repeated characters and special whitespace characters. The actual label sequence is obtained through a defined mapping rule (removing consecutive repeated characters and removing whitespace characters). During training, the feature sequence output by the recurrent layer and the real feature sequence are input. CTCLoss (CTC loss function) uses dynamic programming (forward-backward algorithm) to calculate the sum of probabilities of all possible paths that can be mapped to the real label sequence. Then, this probability is maximized (i.e., the negative log-likelihood is minimized). This process automatically handles the alignment problem between the input and output data, so that the input data of the entire CRNN model only needs to be labeled at the row level, i.e., the sentence length is not limited.

[0153] The specific structure of each level will be further explained in subsequent embodiments.

[0154] S602: Text recognition model training;

[0155] With the training dataset The text recognition model is iteratively trained using sample Qin bamboo slip text data. The CTC Loss of the transcription layer is used as the loss function, which can train the feature extraction layer, recurrent layer and transcription layer simultaneously. Compared with other techniques, the advantage of using the CRNN model is that it only requires the image of the entire text line (in this embodiment, the sample Qin bamboo slip text) and the corresponding text label (i.e., the text meaning corresponding to the sample Qin bamboo slip text). It does not require labeling the specific position of each character in the image, that is, it does not need to consider the sentence length, which can greatly reduce the cost of sample data labeling and improve the convenience of subsequent applications.

[0156] Example 7

[0157] Refer to the attached diagram. Figure 3 Specifically, the feature extraction layer in this embodiment of the invention includes a DRCA module, a STEM module, an Inception Resnet module, a convolution module, and an adaptive pooling module linked in sequence.

[0158] Specifically, the DRCA (Deformable Convolution Residual Connection CoordinateAttention) module is a composite convolutional neural network architecture combining residual learning and coordinate attention mechanisms. It typically employs a composite structure of deformable convolutional units and CA units, and performs residual linking with the input data before output, thereby achieving cross-scale deformation-aware position detection. (Refer to the attached diagram.) Figure 4 The schematic diagram illustrates the DRCA module structure of an embodiment of the present invention. The DRCA of this embodiment includes a pre-deformable convolutional unit, a CA unit, and a post-deformable convolutional unit. Specifically, the pre-deformable convolutional unit includes a 3x3 convolution, a 3x3 deformable convolution, BN, and ReLU linked sequentially. The post-deformable convolutional unit is also linked sequentially with a 3x3 convolution, a 3x3 deformable convolution, BN, and ReLU. Specifically, during the nonlinear activation of the ReLU in the post-deformable convolutional unit before data output, the input data is fused through residual linking before activation.

[0159] Specifically, the STEM module is mainly responsible for rapid downsampling and basic feature extraction, providing low-resolution, high-semantic feature representations for subsequent networks; see attached diagram. Figure 5The schematic diagram illustrates the STEM module structure of an embodiment of the present invention. Specifically, the STEM module can have various different configurations, such as an Inception-style structure. This structure typically processes data through multiple branches before concatenating the channels. The branch structure can be batch normalized convolutional fusion (BN_Conv), continuous batch normalized convolutional fusion, convolution (Conv), continuous convolution, max pooling, or adaptive pooling. The layers can also be designed in multiple levels as needed. That is, after data is processed by branch paths to extract features and concatenate the channels, the process repeats again. The structural design of each branch can be determined according to requirements. For example, in an optional implementation of this invention, after the input data is processed by the DRCA module, it will be divided into two paths (one input...) After being processed by the batch normalized convolutional fusion unit (one input and one Maxpool unit), the data enters the first Concat unit for channel concatenation. The data fused by the first Concat unit is then split into two paths again (one input and one output linked sequentially). Batch normalized convolutional fusion units and Batch normalized convolutional fusion units, with another input sequentially linked... Batch normalized convolutional fusion units Batch normalized convolutional fusion units Batch normalized convolutional fusion units and After processing by the batch normalized convolutional fusion unit, the data is concatenated through a second concat layer. The data fused by the second concat unit is then split into two paths again (one input path and one output path). After being processed by the batch normalized convolutional fusion unit and another input Maxpool unit, the signal enters the third Concat unit for channel splicing, and preliminary features are extracted through the STEM module.

[0160] It should be noted that the DRCA module, as the first processing module for the input Qin bamboo slip text, does not compress the image of the Qin bamboo slip text. Therefore, the DRCA module can continue high-resolution image analysis, but in practical applications, it will consume more computer resources. In this embodiment of the invention, by introducing the coordinate attention mechanism, combined with the pre-deformable convolutional unit and the post-deformable convolutional unit, it can achieve a good function of recognizing and segmenting text objects (the actual function used is target detection and target extraction).

[0161] The STEM module primarily functions as a spatial compression, feature extraction, and semantic transition module. Specifically, the parallel multi-path structure can extract and retain multi-scale features, while the batch normalized convolutional fusion unit can perform spatial compression (downsampling). Furthermore, in this process, through feature extraction, the transformation from raw pixels to abstract features is gradually realized, providing a foundation for subsequent data processing.

[0162] Figure 6 This is a schematic diagram of the Inception-ResNet module structure according to an embodiment of the present invention. Specifically, the Inception-ResNet module is a fusion of the Inception architecture proposed by Google and residual connections. By introducing skip connections from ResNet into the Inception module, training stability and feature transfer efficiency are significantly improved. The specific units involved include Inception-ResNet-A units, Inception-ResNet-B units, Inception-ResNet-C units, and Reduction Block units. The Reduction Block units can have different structures, such as Reduction-A Block units and Reduction-B Block units, depending on the units they are linked to. The Reduction Block unit linked after the Inception-ResNet-A unit is called the Reduction-A Block unit, and the Reduction Block unit linked after the Inception-ResNet-B unit is called the Reduction-B Block unit.

[0163] In this embodiment of the invention, the Inception-ResNet network includes two Inception-ResNet-A units, one Reduction-A Block unit, five Inception-ResNet-B units, one Reduction-B Block unit, and two Inception-ResNet-C units linked sequentially.

[0164] Specifically, the specific structures of the Inception-ResNet-A unit, Inception-ResNet-B unit, and Inception-ResNet-C unit can be implemented with reference to existing technologies, and will not be described in detail in this embodiment of the invention.

[0165] Figure 7 This is a schematic diagram of the Reduction-A Block unit structure according to an embodiment of the present invention.

[0166] Specifically, in this embodiment of the invention, the Reduction-A Block unit includes three parallel branches. The first branch consists of a Maxpool structure; the second branch consists of sequentially linked... It consists of a convolutional structure (batch-normalized convolutional fusion structure) and a ZeroPad structure (zero-padding boundaries); the third branch is composed of sequentially linked... Convolutional structure Convolutional structures and It consists of convolutional layers and ZeroPad structures. The outputs of each branch are concatenated along the channel dimension and then fed into the CPA (Channel-Pixel Attention) structure for further enhancement.

[0167] Figure 8 This is a schematic diagram of the Reduction-B Block unit structure according to an embodiment of the present invention.

[0168] The Reduction-B Block structure includes four parallel branches. The first branch consists of a Maxpool structure; the second branch consists of sequentially linked... Convolutional structure It consists of convolutional structures and ZeroPad structures; the third branch has the same structure as the second branch; the fourth branch is composed of sequentially linked... Convolutional structure Convolutional structure It consists of a convolutional structure and a ZeroPad structure. The outputs of each branch are concatenated along the channel dimension and then fed into the CPA structure for further enhancement.

[0169] Figure 9 This is a schematic diagram of the CPA structure according to an embodiment of the present invention.

[0170] Specifically, in this embodiment of the invention, the CPA structure is mainly implemented through the parallel connection of the channel attention branch (CAM) and the pixel attention branch (PA). When data is input into the CPA structure, it is first processed by GroupNorm. The processed feature maps are then processed by the channel attention branch (CAM) and the pixel attention branch (PA) respectively, followed by channel concatenation, and then output after two 1x1 ordinary convolutions. In this embodiment of the invention, the introduction of the CPA structure into the Reduction Block unit achieves more refined feature selection performance than a single attention mechanism.

[0171] Specifically, the convolutional module and adaptive pooling module located in the feature extraction layer are used to compress the height dimension of the feature map (the height is compressed to 1), and the feature map is output after reasonable dimensionality reduction.

[0172] Furthermore, it should be noted that in the application of the CPA structure, the embodiments of the present invention use residual links between the input and output data of the CPA structure to retain sufficient feature sequence information; therefore, the actual function of the CPA structure can be understood as being used to enhance the feature characteristics of the text sequence.

[0173] Example 8

[0174] Refer to the attached diagram. Figure 3 The schematic diagram of the text recognition model structure shows that the recurrent layer in this embodiment of the invention includes a BiLSTM module (bidirectional recurrent neural network module) and a Transformer Encoder module linked in sequence.

[0175] The Transformer Encoder is a key part of the Transformer architecture, mainly used to encode input sequences and extract high-level feature representations. Specifically, the Transformer Encoder is generally composed of multiple identical layers stacked together. Each layer contains two main sub-layers: a multi-head differential attention mechanism and a feed-forward neural network. Each sub-layer is followed by a residual connection and layer normalization.

[0176] In this embodiment of the invention, the input sequence is first processed using a BiLSTM module, and then the output of the BiLSTM module is used as the input of the Transformer Encoder module. The purpose of this design is that the BiLSTM module, as a low-level feature extractor, can effectively capture the bidirectional local dependencies of the sequence (such as forward and backward contextual information), while the Transformer Encoder module, as a high-level feature extractor, uses a multi-head differential attention mechanism to capture the global dependencies within the sequence, thus achieving the capture of both local and global dependencies. Specifically, the BiLSTM module, as a local feature extractor, can accurately capture direction-sensitive local patterns (such as part-of-speech variations and phrase structures) and is highly sensitive to sequence order (natural time awareness); the Transformer Encoder module, as a global dependency modeler, establishes associations between arbitrary positions through a self-attention mechanism, enabling accelerated training through parallel computation and simultaneously achieving multi-head attention and multi-view feature extraction.

[0177] Specifically, the specific structures of the BiLSTM module and the Transformer Encoder module can be implemented with reference to existing technologies, and will not be further described in this embodiment of the invention.

[0178] Example 9

[0179] Refer to the attached diagram. Figure 3 The schematic diagram of the text recognition model structure shows that the transcription layer in this embodiment of the invention includes a CTCFL-LS module. CTCFL-LS (Connectionist Temporal Classification Focal Loss with LabelSmoothing) is an improved loss function for sequence recognition tasks (such as speech recognition, OCR, and handwriting recognition).

[0180] Specifically, CTC (Connectionist Temporal Classification) is a tool for sequence-to-sequence tasks, especially when the alignment of input and output is unknown (such as speech recognition and handwritten character recognition). CTC can calculate the probability of the target sequence given an input sequence by summing all possible alignment paths through dynamic programming.

[0181] Focal Loss is a tool that makes the model focus more on difficult-to-classify samples by reducing the weight of easily classified samples. Essentially, Focal Loss improves upon standard cross-entropy by introducing an adjustable focusing parameter, thereby increasing attention to specific objects.

[0182] Label Smoothing is a regularization technique that prevents models from becoming overconfident by softening hard labels.

[0183] CTC is essentially used as a loss function. After combining Focal Loss and Label Smoothing techniques, the total loss function of CTCFL-LS is the average loss over all samples, specifically expressed as:

[0184]

[0185] Among them, For batch size, Indicates the first The CTC loss value is calculated based on the predicted probabilities of each sample after label smoothing. This represents the class balancing factor, used to address the problem of uneven class distribution among samples. The difficult sample focusing adjustment factor is used to emphasize samples that are difficult to classify and reduce the loss weight of easy-to-classify samples.

[0186] In this embodiment of the invention, by integrating the sequence alignment capability of CTC (Connectionist Temporal Classification), the sample imbalance handling capability of Focal Loss, and the generalization capability of Label Smoothing, the problems of sequence alignment, class imbalance, and model overconfidence can be solved. In practical applications, CTCFL-LS can bring a 1-2% improvement in absolute accuracy, especially in high noise (such as the Qin bamboo slips text extracted from physical carriers in this embodiment of the invention) and long-tail distribution scenarios.

[0187] Specifically, in the text recognition model, the optimizer is the Adam optimizer, and the learning rate adjustment algorithm is the cosine annealing algorithm. Both the optimizer and the learning rate adjustment algorithm are applied to the training process of the text recognition model.

[0188] In summary, for details please refer to the appendix. Figure 10 A schematic diagram of the first example of the Qin bamboo slip text recognition method according to an embodiment of the present invention and an appendix. Figure 11 The schematic diagram illustrates a second example of the Qin bamboo slip text recognition method according to an embodiment of the present invention. As shown in the example, this ancient character recognition method has relatively accurate recognition results for ancient character texts. In practical applications, there is no order of magnitude difference in computation time for Qin bamboo slip texts of different lengths, and it can quickly identify and output the meaning of unknown Qin bamboo slip texts.

[0189] In summary, this invention provides a method for recognizing Qin Dynasty bamboo slips. This method is based on a text recognition model. During the training process of the text recognition model, a large amount of training data is constructed through imitation, which helps to improve the recognition accuracy of the text recognition model. The process of imitating the training data can fully incorporate the writing habits of ancient Chinese word order, which helps to improve the authenticity of the training data, thereby ensuring the recognition ability of the text recognition model. It has good practicality in practical applications.

[0190] The foregoing has provided a detailed description of a Qin Dynasty bamboo slip text recognition method provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for recognizing Qin Dynasty bamboo slips text, characterized in that, A trained text recognition model is used to identify the predicted meaning of unknown Qin bamboo slip text. The text recognition model is based on a training dataset. Trained; The training dataset ,in, This is a dataset of authentic Qin Dynasty bamboo and wooden slips text. To replicate the Qin Dynasty bamboo slips text dataset, For the number The sample Qin bamboo slip text data, each of the sample Qin bamboo slip text data includes the sample Qin bamboo slip text and the corresponding text meaning, and the data form of the sample Qin bamboo slip text is a grayscale image with the same height. In the real Qin bamboo slip text dataset, the sample Qin bamboo slip text in the sample Qin bamboo slip text data is obtained by extracting real Qin bamboo slip text; In the simulated Qin bamboo slip text dataset, the sample Qin bamboo slip texts in the sample Qin bamboo slip text data are based on a simulated Qin bamboo slip text generation model; The simulated Qin bamboo slip text generation model includes a word and sentence creation module and an image stitching module. The word and sentence creation module is used to generate simulated words and sentences that conform to the word order of Qin bamboo slip text. The image stitching module is used to generate sample Qin bamboo slip text that meets the data format of the sample Qin bamboo slip text based on the simulated words and sentences. The training process of the word-formation module includes: S301, extracting several real Qin bamboo slip texts and their corresponding meanings from the real Qin bamboo slip texts based on the extractor, and storing them in a real Qin bamboo slip text subtext database; S302, extracting the meanings of single characters and words in the real Qin bamboo slip texts, and storing them in a word list by sequential numbering; S303, randomly extracting one real Qin bamboo slip text from the real Qin bamboo slip text subtext database, and converting the meaning of the real Qin bamboo slip text subtext into an index sequence according to the number in the word list; S304, based on Embedding... The layer maps the index sequence to a low-dimensional dense vector; S305, the low-dimensional dense vector is processed time-by-time based on the LSTM layer to obtain the hidden states of all time parts in the low-dimensional dense vector; S306, the hidden states of all time parts corresponding to the low-dimensional dense vector are mapped to the probability distribution in the word list based on the fully connected layer; S307, the loss between the probability distribution and the real label is compared and the LSTM layer, Embedding layer and fully connected layer are adjusted by backpropagation; S308, steps S303 to S307 are repeated until the word-creation module is trained.

2. The Qin bamboo slips text recognition method as described in claim 1, characterized in that, The process of generating the simulated phrases from the simulated Qin bamboo slip text includes: S201: Construct short sentences within the range of single characters and words included in the vocabulary list; S202: Input the short sentence as the first text into the Embedding layer and select the corresponding single character or word as the second text from the word list according to the probability distribution output by the fully connected layer; S203: Combine the first text and the second text to form a new short sentence; S204: Repeat steps S202 and S203 until the length of the short sentence in step S203 reaches the preset word count, and the final generated short sentence is the required imitation sentence.

3. The Qin bamboo slips text recognition method as described in claim 1, characterized in that, The image stitching module is used to generate sample Qin bamboo text that satisfies the format of the sample Qin bamboo text based on the imitated phrases, including: Extract individual Qin bamboo slip characters from the authentic Qin bamboo slip text and store them along with their meanings in the corresponding Qin bamboo slip character database. Based on the fabricated phrases, several corresponding Qin bamboo slip characters are extracted from the Qin bamboo slip character database and spliced ​​together to form a spliced ​​Qin bamboo slip text. The spliced ​​Qin bamboo slip text is processed according to the corresponding format to generate sample Qin bamboo slip text that meets the data format of the sample Qin bamboo slip text.

4. The Qin bamboo slips text recognition method as described in claim 3, characterized in that, The splicing methods include vertical splicing, splicing with CLAHE, splicing with wavelet filtering, splicing with gradient fusion, splicing with image normalization, and splicing with feathering.

5. The Qin bamboo slips text recognition method as described in claim 1, characterized in that, The text recognition model training process includes: A text recognition model is constructed, comprising a feature extraction layer, a recurrent layer, and a transcription layer linked sequentially. Text recognition model training, using the training dataset The text recognition model is iteratively trained using sample Qin Dynasty bamboo slip text data.

6. The Qin bamboo slips text recognition method as described in claim 5, characterized in that, The feature extraction layer includes a DRCA module, a STEM module, an Inception-Resnet module, a convolution module, and an adaptive pooling module linked in sequence.

7. The Qin bamboo slips text recognition method as described in claim 5, characterized in that, The recurrent layer includes a BiLSTM module and a Transformer Encoder module.

8. The Qin bamboo slips text recognition method as described in claim 5, characterized in that, The transcription layer includes the CTCFL-LS module.

9. The method for recognizing Qin bamboo slip text as described in any one of claims 5 to 8, characterized in that, The text recognition model uses the Adam optimizer and cosine annealing algorithm for learning rate adjustment during training.

Images