A translation device based on Tibetan character structure and image reading method thereof

Abstract

The present application provides a translation device based on Tibetan character structure and an image reading method thereof, wherein syllable vectors are extracted from a Tibetan character structure sequence obtained by performing a convolution recognition operation on a target Tibetan text image to obtain multiple syllable vectors; the inline relationship between syllables is determined according to the confidence distance between each syllable vector, and then the syllable loss amount in the syllable vector extraction process is determined by the probability of co-occurrence between all inline relationships and each syllable vector; the convolution interference amount of character recognition is determined by the ambient light intensity data and the Tibetan character structure sequence, and then the character recognition loss when recognizing Tibetan characters is determined by the convolution interference amount and the syllable loss amount; and then each syllable vector is read and compensated by the character recognition loss to obtain multiple syllable compensation vectors, and image reading is completed based on all syllable compensation vectors. The scheme of the present application can realize adaptive compensation of recognition loss in the process of Tibetan character structure recognition.

Description

Technical Field

[0001] The present application relates to the technical field of character recognition, and more specifically, to a translation device based on Tibetan character structure and an image reading method thereof. Background Art

[0002] Tibetan is a language with a long history and unique cultural background. It uses a combination of consonant letters and vowel symbols to record pronunciation. The characteristics of the Tibetan character structure are its complexity and systematicity, which is mainly reflected in its unique alphabet consisting of 30 consonant letters and 4 vowel symbols. Consonant letters can be arranged in the positions of front added characters, top added characters, bottom added characters, and back added characters in words, and different syllables are formed through combination.

[0003] In terms of translation based on Tibetan character structure, due to the complex structure of Tibetan characters, traditional character recognition methods are difficult to cope with. Modern deep learning technology, especially the application of convolutional neural networks (CNN), provides a new solution. Through image preprocessing, feature extraction and character classification, the system can quickly identify the Tibetan character structure and perform Tibetan translation based on the recognition results. However, in the process of using convolutional neural networks to recognize Tibetan character structures, due to the complexity of the Tibetan character structure and the interference of the external environment, a large amount of recognition loss is generated in the recognition process, resulting in a large deviation in the Tibetan translation results. By dynamically compensating for the recognition loss in the process of recognizing the Tibetan character structure, the translation deviation caused by the recognition loss can be effectively solved. Therefore, how to achieve adaptive compensation for the recognition loss in the process of Tibetan character structure recognition, thereby improving the accuracy of Tibetan translation, has become a difficult problem faced by the industry. Summary of the invention

[0004] The present application provides a translation device based on Tibetan character structure and an image reading method thereof, which can realize adaptive compensation of recognition loss in the process of Tibetan character structure recognition, thereby improving the accuracy of Tibetan translation.

[0005] In a first aspect, the present application provides an image reading method based on a Tibetan character structure, comprising the following steps:

[0006] Starting a Tibetan translation device, and performing a convolution recognition operation on a character structure in a target Tibetan text image through an image reading device in the Tibetan translation device, thereby obtaining a Tibetan character structure sequence;

[0007] Extracting syllable vectors of each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector of each Tibetan character structure;

[0008] The inline relationship of the syllables between each Tibetan character structure and other Tibetan character structures is determined according to the confidence distance between each syllable vector, and then the syllable loss amount generated during character segmentation in the syllable vector extraction process is determined through all the inline relationships and the probability of co-occurrence between each syllable vector;

[0009] Acquire light intensity data when the image reading device recognizes the target Tibetan text image, determine the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation through the light intensity data and the Tibetan character structure sequence, and then determine the character recognition loss of the image reading device during the Tibetan text character recognition process through the convolution interference amount and the syllable loss amount;

[0010] The syllable vector of each Tibetan character structure is read and compensated according to the character recognition loss to obtain the syllable compensation vector of each Tibetan character structure, and the image reading of the Tibetan translation device is completed based on all the syllable compensation vectors.

[0011] Preferably, extracting syllable vectors for each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector for each Tibetan character structure specifically includes:

[0012] Performing character segmentation on the Tibetan character structures in the Tibetan character structure sequence by using a pre-trained Tibetan language recognition model to obtain a plurality of independent Tibetan character structures;

[0013] All Tibetan character structures are converted into vectors to obtain the syllable vector of each Tibetan character structure.

[0014] Preferably, determining the syllable inline relationship between each Tibetan character structure and other Tibetan character structures according to the confidence distance between each syllable vector specifically includes:

[0015] Determine the distance between each syllable vector;

[0016] Determine the confidence distance between corresponding syllable vectors through the distance between each syllable vector;

[0017] Determine the co-occurrence discrimination of each syllable vector based on all confidence distances;

[0018] Determine the correlation coefficient of the corresponding syllable vector according to the frequency of occurrence of each syllable vector;

[0019] The syllable in-line relationship between each Tibetan character structure and other Tibetan character structures is determined by the confidence distance between each syllable vector, all co-occurrence discriminations and all correlation coefficients.

[0020] Preferably, determining the syllable loss amount generated during character segmentation during the syllable vector extraction process through all the inline relationships and the probability of co-occurrence between the syllable vectors specifically includes:

[0021] Acquiring a predetermined amount of Tibetan text data;

[0022] Constructing a co-occurrence matrix of syllable vectors through the Tibetan text data;

[0023] Determining the probability of co-occurrence between various syllable vectors according to the co-occurrence matrix;

[0024] Determine the conversion error when converting each Tibetan character structure into a syllable vector through all probabilities and all internal relationships;

[0025] The amount of syllable loss caused by character segmentation during the syllable vector extraction process is determined based on all conversion errors.

[0026] Preferably, determining the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation by using the light intensity data and the Tibetan character structure sequence specifically includes:

[0027] Determine the fluctuation entropy of the light intensity when the external environment light changes through the light intensity data;

[0028] Performing a deconvolution operation on the Tibetan character structure sequence, thereby extracting a convolution residual in a convolution recognition operation process;

[0029] The amount of convolution interference caused by changes in ambient illumination on character recognition during a convolution recognition operation is determined according to the fluctuation entropy and the convolution residual.

[0030] Preferably, determining the character recognition loss of the image reading device in the process of recognizing Tibetan text characters by the convolution interference amount and the syllable loss amount specifically includes:

[0031] Obtain the historical convolution interference amount and the historical syllable loss amount;

[0032] Determine the convolution loss weight of the convolution interference amount by using the historical convolution interference amount;

[0033] Determine the syllable loss weight of the syllable loss amount according to the historical syllable loss amount;

[0034] The convolution interference amount and the syllable loss amount are weighted and fused according to the convolution loss weight and the syllable loss weight to obtain the character recognition loss of the image reading device in the process of recognizing Tibetan text characters.

[0035] Preferably, the Tibetan language recognition model built into the image reading device in the Tibetan language translation apparatus is a convolutional neural network model.

[0036] In a second aspect, the present application provides a translation device based on a Tibetan character structure, the translation device based on a Tibetan character structure includes an image reading unit, and the image reading unit includes:

[0037] an acquisition module, configured to, after starting the Tibetan translation device, perform a convolution recognition operation on the character structure in the target Tibetan text image through an image reading device in the Tibetan translation device, thereby obtaining a Tibetan character structure sequence;

[0038] A processing module, configured to extract syllable vectors of each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector of each Tibetan character structure;

[0039] The processing module is further used to determine the inline relationship of syllables between each Tibetan character structure and other Tibetan character structures according to the confidence distance between each syllable vector, and then determine the syllable loss amount generated during character segmentation during the syllable vector extraction process through all the inline relationships and the probability of co-occurrence between each syllable vector;

[0040] The processing module is further used to obtain light intensity data when the image reading device recognizes the target Tibetan text image, determine the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation through the light intensity data and the Tibetan character structure sequence, and then determine the character recognition loss of the image reading device during the process of recognizing Tibetan text characters through the convolution interference amount and the syllable loss amount;

[0041] The execution module is used to read and compensate the syllable vector of each Tibetan character structure according to the character recognition loss, obtain the syllable compensation vector of each Tibetan character structure, and complete the image reading of the Tibetan translation device based on all the syllable compensation vectors.

[0042] In a third aspect, the present application provides a computer device, comprising a memory and a processor, wherein the memory stores codes, and the processor is configured to obtain the codes and execute the above-mentioned image reading method based on Tibetan character structure.

[0043] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned image reading method based on the Tibetan character structure is implemented.

[0044] The technical solution provided by the embodiments disclosed in this application has the following beneficial effects:

[0045] In an embodiment of the present application, a Tibetan translation device is started, and a convolution recognition operation is performed on a character structure in a target Tibetan text image through an image reading device in the Tibetan translation device, thereby obtaining a Tibetan character structure sequence; a syllable vector is extracted from each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan recognition model to obtain a syllable vector for each Tibetan character structure; an inline relationship of syllables between each Tibetan character structure and other Tibetan character structures is determined based on a confidence distance between each syllable vector, and then the syllable vector extraction process is determined through the probability of all inline relationships and co-occurrence between each syllable vector. The invention relates to a method for realizing a Tibetan translation device and a method for realizing a Tibetan translation device. The method comprises the following steps: obtaining light intensity data when an image reading device recognizes a target Tibetan text image, determining the amount of convolution interference caused by changes in ambient light on character recognition during a convolution recognition operation through the light intensity data and the Tibetan character structure sequence, and then determining the character recognition loss of the image reading device during the process of recognizing Tibetan text characters through the convolution interference amount and the syllable loss amount; performing reading compensation on the syllable vector of each Tibetan character structure through the character recognition loss to obtain the syllable compensation vector of each Tibetan character structure, and completing the image reading of the Tibetan translation device based on all the syllable compensation vectors.

[0046] It can be seen that the present application determines the character recognition loss of the image reading device in the process of identifying Tibetan text characters through the convolution interference amount and the syllable loss amount, and then uses the character recognition loss to read and compensate the syllable vector of each Tibetan character structure to obtain the syllable compensation vector of each Tibetan character structure; firstly, the overall syllable error caused by character segmentation errors is quantified through the inline relationship between syllable vectors and the probability of co-occurrence of each syllable vector to obtain the syllable loss amount, so that the syllable loss in the character segmentation process can be evaluated through the syllable loss amount; secondly, the overall impact of ambient light changes on the convolution recognition operation is quantified through the convolution interference amount, and then the convolution interference is evaluated. Compensation can reduce the negative impact of changes in ambient light on recognition results, thereby improving recognition accuracy; then, the character recognition loss is determined by the convolution interference amount and the syllable loss amount, and the character recognition loss can reflect the overall recognition loss generated by the image reading device in the process of recognizing Tibetan text characters; finally, the syllable vector of the Tibetan character structure is adaptively compensated by the overall recognition loss, and the obtained syllable compensation vector can more accurately reflect the information of the original text, thereby improving the performance of the overall translation device and achieving high-quality Tibetan text recognition and translation; in summary, this scheme can realize adaptive compensation of recognition loss in the process of Tibetan character structure recognition, thereby improving the accuracy of Tibetan translation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1is an exemplary flow chart of an image reading method based on a Tibetan character structure according to some embodiments of the present application;

[0048] Figure 2 is a schematic diagram of a process for determining a convolution interference amount according to some embodiments of the present application;

[0049] Figure 3 is a schematic diagram of an exemplary software module of an image reading unit according to some embodiments of the present application;

[0050] Figure 4 It is a structural schematic diagram of a computer device for implementing an image reading method based on a Tibetan character structure according to some embodiments of the present application. DETAILED DESCRIPTION

[0051] The core of this application is to extract syllable vectors from a Tibetan character structure sequence obtained by performing a convolution recognition operation on a target Tibetan text image to obtain a plurality of syllable vectors; determine the inline relationship between syllables based on the confidence distance between each syllable vector, and then determine the syllable loss amount in the syllable vector extraction process based on the probability of co-occurrence between all the inline relationships and each syllable vector; determine the convolution interference amount of character recognition based on the ambient light intensity data and the Tibetan character structure sequence, and then determine the character recognition loss when recognizing Tibetan characters based on the convolution interference amount and the syllable loss amount; and then perform reading compensation on each syllable vector based on the character recognition loss to obtain a plurality of syllable compensation vectors, and complete image reading based on all the syllable compensation vectors; this scheme can realize adaptive compensation of recognition loss in the process of Tibetan character structure recognition.

[0052] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods. Figure 1 , which is an exemplary flow chart of an image reading method based on a Tibetan character structure according to some embodiments of the present application. The image reading method 100 based on a Tibetan character structure mainly includes the following steps:

[0053] In step 101, a Tibetan translation device is started, and a convolution recognition operation is performed on a character structure in a target Tibetan text image by an image reading device in the Tibetan translation device, thereby obtaining a Tibetan character structure sequence.

[0054] It should be noted that the image reading device in the Tibetan translation device in the present application is an electronic device for Tibetan text recognition. In specific implementation, the image reading device in the Tibetan translation device performs a convolution recognition operation on the character structure in the target Tibetan text image, and then obtains the Tibetan character structure sequence. The following method can be used to implement it, namely: after starting the Tibetan translation device, the image reading device in the Tibetan translation device uses a convolution recognition technology to read the character structure in the target Tibetan text image, and then the recognition results of each character are combined into a sequence in order from left to right and from top to bottom, and then the combined sequence is used as the Tibetan character structure sequence; it should also be noted that the convolution recognition technology in the present application is an image recognition technology based on a convolutional neural network, wherein the character structure refers to the Tibetan character structure in the Tibetan text image, and its Tibetan character structure includes: basic characters, vowel symbols, stacked characters and sound node characters.

[0055] It should also be noted that the Tibetan recognition model built into the image reading device in the Tibetan translation device of the present application is a convolutional neural network model, which can quickly recognize characters in Tibetan text images. In other embodiments, other existing recognition models can also be used to recognize characters in Tibetan text images, which is not limited here.

[0056] In step 102, syllable vectors are extracted for each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector for each Tibetan character structure.

[0057] It should be noted that the pre-trained Tibetan recognition model in the present application refers to a pre-trained model for Tibetan character recognition. The Tibetan recognition model can be obtained by performing Tibetan character recognition training on a large amount of Tibetan text image data. In the present application, the algorithm structure of the pre-trained model can adopt a convolutional neural network. In other embodiments, other existing algorithm structures can also be used for implementation, which is not limited here.

[0058] In some embodiments, extracting syllable vectors for each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector for each Tibetan character structure may be achieved by the following steps:

[0059] Performing character segmentation on the Tibetan character structures in the Tibetan character structure sequence by using a pre-trained Tibetan language recognition model to obtain a plurality of independent Tibetan character structures;

[0060] All Tibetan character structures are converted into vectors to obtain the syllable vector of each Tibetan character structure.

[0061] In a specific implementation, the Tibetan character structures in the Tibetan character structure sequence are segmented by a pre-trained Tibetan language recognition model to obtain a plurality of independent Tibetan character structures, which can be achieved in the following manner: the Tibetan character structure sequence is used as an initialization parameter of the pre-trained Tibetan language recognition model, and the pre-trained Tibetan language recognition model is used to segment the Tibetan character structures in the Tibetan character structure sequence according to Tibetan character structure features, thereby obtaining a plurality of independent Tibetan character structures, wherein the Tibetan character structure features refer to the external structure features of Tibetan characters, and the external structure features can be composed of basic Tibetan components, such as basic characters, vowel symbols, stacked characters, and phonetic node characters.

[0062] In the specific implementation, all Tibetan character structures are converted into vectors, and the syllable vector of each Tibetan character structure is obtained, which can be achieved in the following way, that is: all Tibetan character structures can be converted into word vectors by the Word2Vec for Tibetan method in the prior art, so that all the obtained word vectors are used as syllable vectors of the corresponding Tibetan character structure. It should be noted that in other embodiments, other existing word vector conversion technologies can also be used, such as BERT for Tibetan, which adopts the BERT structure and uses Tibetan text data for pre-training to generate syllable vectors of Tibetan character structures, which will not be repeated here.

[0063] In step 103, the inline relationship of the syllables between each Tibetan character structure and other Tibetan character structures is determined based on the confidence distance between each syllable vector, and then the syllable loss amount generated during character segmentation during the syllable vector extraction process is determined through all the inline relationships and the probability of co-occurrence between each syllable vector.

[0064] In some embodiments, determining the syllable inline relationship between each Tibetan character structure and other Tibetan character structures according to the confidence distance between each syllable vector can be implemented by the following steps:

[0065] Determine the distance between each syllable vector;

[0066] Determine the confidence distance between corresponding syllable vectors through the distance between each syllable vector;

[0067] Determine the co-occurrence discrimination of each syllable vector based on all confidence distances;

[0068] Determine the correlation coefficient of the corresponding syllable vector according to the frequency of occurrence of each syllable vector;

[0069] The syllable in-line relationship between each Tibetan character structure and other Tibetan character structures is determined by the confidence distance between each syllable vector, all co-occurrence discriminations and all correlation coefficients.

[0070] In specific implementation, the following method can be used to determine the distance between each syllable vector, namely: select a syllable vector and use the syllable vector as the designated syllable vector, select a syllable vector from the remaining syllable vectors as the target syllable vector, and use the number of syllable vectors between the designated syllable vector and the target syllable vector as the distance between the designated syllable vector and the target syllable vector, thereby obtaining the distance between the designated syllable vector and other syllable vectors, repeat the above steps to obtain the distance between the remaining syllable vectors and other syllable vectors. It should be noted that in the present application, the distance between any two syllable vectors is calculated twice in the process of determining the distance between syllable vectors, and any calculation result can be used as the distance between the corresponding two syllable vectors. In addition, it should be noted that the distance between adjacent syllable vectors is defined as 0. In other embodiments, the distance between adjacent syllable vectors can also be defined as a number infinitely close to 0, which is not limited here.

[0071] In addition, in the specific implementation, the confidence distance between corresponding syllable vectors can be determined by the distance between each syllable vector in the following manner, namely: the natural exponential of the opposite number of the distance between any two syllable vectors can be used as the confidence distance between the corresponding two syllable vectors, that is, the confidence distance = exp (- the distance between the syllable vectors). It should be noted that the confidence distance in the present application is an indicator to measure the degree of credibility of the simultaneous appearance of any two syllable vectors. The larger the confidence distance, the higher the degree of credibility of the simultaneous appearance of any two syllable vectors, and the smaller the confidence distance, the lower the degree of credibility of the simultaneous appearance of any two syllable vectors.

[0072] In specific implementation, the co-occurrence discrimination of each syllable vector can be determined based on all confidence distances in the following manner, namely: select a syllable vector, take the average value of the confidence distance between the syllable vector and other syllable vectors as the co-occurrence discrimination of the syllable vector, repeat the above steps to obtain the co-occurrence discrimination of the remaining syllable vectors. It should be noted that the co-occurrence discrimination in this application is a measurement indicator for measuring the mutual dependence of two syllable vectors (or words) that co-occur in a text. The higher the co-occurrence discrimination, the higher the mutual dependence of the two syllable vectors that co-occur in a text. The lower the co-occurrence discrimination, the lower the mutual dependence of the two syllable vectors that co-occur in a text. The value range of the co-occurrence discrimination is between 0 and 1.

[0073] It should be noted that in this application, by calculating the co-occurrence distinction of each syllable vector, the correlation information between the syllable vectors can be obtained, which can then assist in understanding the inline relationship between the Tibetan character structures, thereby providing co-occurrence information at different levels and helping to build a more accurate syllable relationship model.

[0074] In specific implementation, the correlation coefficient of the corresponding syllable vector can be determined according to the frequency of occurrence of each syllable vector in the following manner, namely: the total number of syllable vectors is obtained, and the quotient of the frequency of occurrence of each syllable vector and the total number is used as the correlation coefficient of the corresponding syllable vector. The correlation coefficient is an indicator of the degree of correlation between the words in the text and the full text content. The larger the correlation coefficient, the higher the correlation between the corresponding word and the full text content, and the smaller the correlation coefficient, the lower the correlation between the corresponding word and the full text content.

[0075] In addition, in a specific implementation, the inline relationship of the syllables between each Tibetan character structure and other Tibetan character structures can be determined by the confidence distance between each syllable vector, all co-occurrence distinctions and all correlation coefficients. The following method can be used to implement the method, namely: initialize a relationship model, use all co-occurrence distinctions and all correlation coefficients as constraint parameters, and then use the confidence distance between each syllable vector as an initialization parameter, and then output the inline relationship of the syllables between each Tibetan character structure and other Tibetan character structures through the relationship model. It should be noted that the relationship model uses a machine learning algorithm (such as a regression algorithm, a neural network) to establish the inline relationship between syllables. The inline relationship is, for example: the inline relationship of the syllables between the i-th Tibetan character structure and other Tibetan character structures = A*the average value of the confidence distance between the i-th syllable vector and other syllable vectors*the correlation coefficient of the i-th syllable vector+B*the co-occurrence distinction of the i-th syllable vector, wherein A and B are weight coefficients, and A and B can be determined according to the inline relationship of Tibetan character structures in a large number of Tibetan texts, which will not be repeated here.

[0076] It should be noted that the inline relationship of syllables in the present application is a comprehensive indicator for measuring the strength of association between syllables in Tibetan text. The greater the inline relationship of the syllable, the greater the degree of association between the syllable and other syllables in the Tibetan text. The smaller the inline relationship of the syllable, the smaller the degree of association between the syllable and other syllables in the Tibetan text. In the present application, by determining the inline relationship of syllables, the relationship between words can be quantified, thereby providing a reference basis for the subsequent character segmentation process.

[0077] In some embodiments, determining the amount of syllable loss generated during character segmentation during syllable vector extraction by using all inline relationships and the probability of co-occurrence between each syllable vector can be implemented by the following steps:

[0078] Acquiring a predetermined amount of Tibetan text data;

[0079] Constructing a co-occurrence matrix of syllable vectors through the Tibetan text data;

[0080] Determining the probability of co-occurrence between various syllable vectors according to the co-occurrence matrix;

[0081] Determine the conversion error when converting each Tibetan character structure into a syllable vector through all probabilities and all internal relationships;

[0082] The amount of syllable loss caused by character segmentation during the syllable vector extraction process is determined based on all conversion errors.

[0083] In specific implementation, the present application can extract the co-occurrence relationship between each syllable vector from a large amount of Tibetan text data. In the present application, obtaining a predetermined amount of Tibetan text data means obtaining more than 10,000 Tibetan text data. In other embodiments, it can also be other amounts of Tibetan text data, which is not limited here.

[0084] It should be noted that the co-occurrence matrix in the present application is a matrix used to characterize the number of times each syllable co-occurs in a Tibetan text. In specific implementation, the co-occurrence matrix of the syllable vector constructed by the Tibetan text data can be implemented in the following manner, namely: record the number of times each pair of syllables (syllable vectors) appears in the same context, arrange all the numbers in a series from large to small, and use the series as the co-occurrence matrix of the syllable vector, where one syllable corresponds to one syllable vector.

[0085] In specific implementation, the probability of co-occurrence between each syllable vector is determined based on the co-occurrence matrix, which can be achieved in the following manner, namely: randomly selecting two syllable vectors from each syllable vector as the designated syllable vector and the target syllable vector respectively, taking the number of occurrences of the designated syllable vector as the designated number, taking the number of occurrences of the target syllable vector as the target number, obtaining the number of co-occurrences of the designated syllable vector and the target syllable vector from the co-occurrence matrix, and then taking the quotient of the average of the designated number and the target number and the number of co-occurrences of the designated syllable vector and the target syllable vector as the probability of co-occurrence between the designated syllable vector and the target syllable vector.

[0086] In addition, in the specific implementation, the conversion error when each Tibetan character structure is converted into a syllable vector can be determined by all probabilities and all inline relationships. The following method is adopted, namely: a Tibetan character structure is selected from all Tibetan character structures, the probability of the co-occurrence between the syllable vector corresponding to the Tibetan character structure and other syllable vectors is taken as the target probability, the inline relationship of the syllable between the Tibetan character structure and other Tibetan character structures is taken as the target inline relationship, and then the natural exponent of the opposite number of the product between the target probability and the target inline relationship is taken as the conversion error when the Tibetan character structure is converted into a syllable vector, and the above steps are repeated to obtain the conversion error when the remaining Tibetan character structures are converted into syllable vectors; the syllable loss amount generated during character segmentation in the process of syllable vector extraction is determined according to all conversion errors. The following method is adopted, namely: the set composed of all conversion errors is taken as the syllable loss amount generated during character segmentation in the process of syllable vector extraction.

[0087] It should be noted that the conversion error in the present application refers to the error generated when the Tibetan character structure is converted into a syllable vector. The conversion error is usually caused by a segmentation error when the Tibetan character structure is segmented. The larger the conversion error, the larger the error caused by the segmentation error when the Tibetan character structure is segmented. The smaller the conversion error, the smaller the error caused by the segmentation error when the Tibetan character structure is segmented. In addition, the syllable loss in the present application refers to the overall syllable error generated when the Tibetan character structure is converted into a syllable vector due to character segmentation errors during character segmentation in the syllable vector extraction process. The larger the syllable loss, the larger the overall syllable error generated when the Tibetan character structure is converted into a syllable vector due to character segmentation errors during character segmentation in the syllable vector extraction process. The smaller the syllable loss, the smaller the overall syllable error generated when the Tibetan character structure is converted into a syllable vector due to character segmentation errors during character segmentation in the syllable vector extraction process.

[0088] In step 104, the light intensity data when the image reading device recognizes the target Tibetan text image is obtained, and the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation is determined by the light intensity data and the Tibetan character structure sequence, and then the character recognition loss of the image reading device in the process of recognizing Tibetan text characters is determined by the convolution interference amount and the syllable loss amount.

[0089] In a specific implementation, obtaining the light intensity data when the image reading device identifies the target Tibetan text image can be achieved in the following manner, namely: obtaining the light intensity data when the image reading device identifies the target Tibetan text image from the database of the image reading device, wherein the light intensity data is collected by a light sensor according to a preset sampling frequency, and the value range of the preset sampling frequency is between 50-100 Hz. In other embodiments, the value range of the preset sampling frequency can also be set in other intervals according to actual needs, which is not limited here.

[0090] In some embodiments, reference Figure 2 As shown, this figure is a schematic diagram of the process of determining the convolution interference amount in some embodiments of the present application. In this embodiment, the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation can be determined by the light intensity data and the Tibetan character structure sequence, which can be implemented by the following steps:

[0091] First, in step 1041, the fluctuation entropy of the light intensity when the external environment light changes is determined by using the light intensity data;

[0092] Then, in step 1042, a deconvolution operation is performed on the Tibetan character structure sequence to extract a convolution residual in the convolution recognition operation process;

[0093] Finally, in step 1043, the amount of convolution interference caused by the change in ambient light on character recognition during the convolution recognition operation is determined based on the fluctuation entropy and the convolution residual.

[0094] In specific implementation, the following method can be used to determine the fluctuation entropy of light intensity when the external environment light changes through the light intensity data, namely: the light intensity values ​​in the light intensity data are arranged in the order of collection time to obtain a light intensity sequence, and the difference between two adjacent light intensity values ​​in the light intensity sequence is used as the light intensity fluctuation value, and then all the fluctuation values ​​are obtained, and then, the entropy characteristics of all the fluctuation values ​​are described by the information entropy in the prior art to obtain the information entropy of the fluctuation value, and the information entropy of the fluctuation value can be used as the fluctuation entropy of the light intensity.

[0095] It should be noted that the fluctuation entropy of light intensity in the present application is an indicator to measure the uncertainty of light intensity fluctuation when the external environment changes. The larger the fluctuation entropy, the greater the uncertainty of light intensity fluctuation when the external environment changes. The smaller the fluctuation entropy, the smaller the uncertainty of light intensity fluctuation when the external environment changes.

[0096] It should also be noted that in this application, a convolutional neural network is used to perform convolution processing on the Tibetan text image to extract character structure features. Therefore, in this application, a deconvolution operation is performed on the character structure features to analyze the convolution residual during the convolution processing.

[0097] In a specific implementation, performing a deconvolution operation on the Tibetan character structure sequence and then extracting the convolution residual during the convolution recognition operation can be achieved in the following manner, namely: performing a deconvolution operation on all character structure features in the Tibetan character structure sequence to obtain a deconvolution feature of each character structure feature, and determining the convolution residual during the convolution recognition operation by the difference between all the deconvolution features and the character structure features in the Tibetan text image. It should be noted that the convolution residual in the present application is an indicator for measuring the intensity of additional noise introduced due to external interference during the convolution recognition operation. The larger the convolution residual, the greater the intensity of additional noise introduced due to external interference during the convolution operation, and the smaller the convolution residual, the smaller the intensity of additional noise introduced due to external interference during the convolution operation.

[0098] It should be noted that in the process of collecting Tibetan text images using an optical sensor, the spatial variation of light intensity caused by light interference or diffraction results in the appearance of alternating light and dark speckles on the collected Tibetan text image, and the speckles are randomly distributed. When the intensity of the external ambient light changes, the speckles appearing on the Tibetan text image will change randomly, resulting in random changes in the intensity of additional noise introduced by external interference. Therefore, in this application, the convolution residual in the convolution recognition operation is adjusted by the fluctuation entropy of the light intensity when the external ambient light changes, which can effectively reduce the random errors caused by changes in the intensity of the external ambient light.

[0099] In specific implementation, the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation can be determined according to the fluctuation entropy and the convolution residual in the following manner, that is, the fluctuation entropy can be used as an adjustment coefficient, and the product of the adjustment coefficient and the convolution residual can be used as the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation. As a preferred embodiment, the adjustment coefficient can be implemented in the following manner, that is, the natural exponent of the opposite number of the fluctuation entropy can be used as the adjustment coefficient.

[0100] It should be noted that the convolution interference amount in this application refers to the overall impact of ambient lighting changes on the convolution recognition operation. The larger the convolution interference amount, the greater the overall impact of ambient lighting changes on the convolution recognition operation. The smaller the convolution interference amount, the smaller the overall impact of ambient lighting changes on the convolution recognition operation.

[0101] In some embodiments, determining the character recognition loss of the image reading device in the process of recognizing Tibetan text characters by the convolution interference amount and the syllable loss amount can be implemented by the following steps:

[0102] Obtain the historical convolution interference amount and the historical syllable loss amount;

[0103] Determine the convolution loss weight of the convolution interference amount by using the historical convolution interference amount;

[0104] Determine the syllable loss weight of the syllable loss amount according to the historical syllable loss amount;

[0105] The convolution interference amount and the syllable loss amount are weighted and fused according to the convolution loss weight and the syllable loss weight to obtain the character recognition loss of the image reading device in the process of recognizing Tibetan text characters.

[0106] It should be noted that in the present application, the historical convolution interference amount and the historical syllable loss amount can be obtained from the image reading device, the historical convolution interference amount refers to the set of convolution interference amounts extracted when convolution recognition is performed on all Tibetan text images in the historical time period, and the historical syllable loss amount refers to the set of syllable loss amounts generated during character segmentation during the syllable vector extraction process for all Tibetan text images in the historical time period.

[0107] In specific implementation, determining the convolution loss weight of the convolution interference amount through the historical convolution interference amount can be implemented in the following manner, namely: the relationship between the historical convolution interference amount and the character recognition loss can be analyzed by the regression analysis algorithm in the prior art, and the relationship weight obtained by the analysis is used as the convolution loss weight of the convolution interference amount. As a preferred embodiment, the convolution loss weight can be determined in the following manner, namely: all historical convolution interference amounts are used as the input of the convolution loss weight function, and the output of the convolution loss weight function is used as the convolution loss weight, wherein the convolution loss weight function can be determined by analyzing the relationship between the historical convolution interference amount and the character recognition loss using a regression analysis algorithm; determining the syllable loss weight of the syllable loss amount through the historical syllable loss amount can be implemented in the following manner, namely: all historical syllable loss amounts are used as the input of the syllable loss weight function, and the output of the syllable loss weight function is used as the syllable loss weight, wherein the syllable loss weight function can be determined by analyzing the relationship between the historical syllable loss amount and the character recognition loss using a regression analysis algorithm, which will not be elaborated here.

[0108] In addition, in a specific implementation, the convolution interference amount and the syllable loss amount are weighted and fused according to the convolution loss weight and the syllable loss weight to obtain the character recognition loss of the image reading device in the process of recognizing Tibetan text characters. This can be achieved in the following manner, namely: the product of the convolution loss weight and the convolution interference amount is used as the convolution loss, the product of the syllable loss weight and the syllable loss amount is used as the syllable loss, and then the sum of the convolution loss and the syllable loss is used as the character recognition loss of the image reading device in the process of recognizing Tibetan text characters.

[0109] It should be noted that the character recognition loss in the present application refers to the overall recognition loss generated by the image reading device in the process of recognizing Tibetan text characters. The overall recognition loss includes the convolution interference amount and the syllable loss amount. The larger the character recognition loss, the larger the overall recognition loss generated by the image reading device in the process of recognizing Tibetan text characters. The smaller the character recognition loss, the smaller the overall recognition loss generated by the image reading device in the process of recognizing Tibetan text characters.

[0110] In step 105, the syllable vector of each Tibetan character structure is read and compensated by the character recognition loss to obtain a syllable compensation vector of each Tibetan character structure, and the image reading of the Tibetan translation device is completed based on all the syllable compensation vectors.

[0111] In some embodiments, the syllable vector of each Tibetan character structure is read and compensated according to the character recognition loss, and the syllable compensation vector of each Tibetan character structure can be obtained by the following steps:

[0112] Initialize a loss compensation model;

[0113] Using the character recognition loss as a constraint parameter of the loss compensation model;

[0114] The syllable vector of each Tibetan character structure is compensated by the loss compensation model to obtain a syllable compensation vector of each Tibetan character structure.

[0115] It should be noted that the loss compensation model in the present application is a pre-trained model for adaptively compensating for character recognition loss in the character structure recognition process of Tibetan text images. The loss compensation model is a loss compensation model obtained by preprocessing a large amount of Tibetan text image data, performing feature extraction, model training and verification. The algorithm structure in the model training process can adopt a convolutional neural network. In other embodiments, other existing algorithm structures can also be adopted, which is not limited here.

[0116] In specific implementation, the syllable vector of each Tibetan character structure is compensated by the loss compensation model, and the syllable compensation vector of each Tibetan character structure can be achieved in the following way, namely: the syllable vector of each Tibetan character structure is used as the initialization parameter of the loss compensation model, and then the character recognition loss of each syllable vector is compensated by the loss compensation model, so as to obtain each compensated syllable vector, and each compensated syllable vector is used as a syllable compensation vector.

[0117] It should be noted that in the present application, completing the image reading of the Tibetan translation device based on all syllable compensation vectors means taking all syllable compensation vectors as the result of the image reading of the Tibetan translation device, and taking the result of the image reading as the initialization parameter of the Tibetan text translation model to complete the translation of Tibetan characters in the Tibetan text image. It should also be noted that in the present application, the recognition loss in the process of recognizing the character structure of the Tibetan text is adaptively compensated, thereby reducing the conversion error of the syllable vector, thereby improving the accuracy of the Tibetan translation.

[0118] On the other hand, in some embodiments, the present application provides a translation device based on a Tibetan character structure, the translation device based on a Tibetan character structure includes an image reading unit, referring to Figure 3 , which is a schematic diagram of exemplary software modules of an image reading unit according to some embodiments of the present application, the image reading unit 400 includes: an acquisition module 401, a processing module 402 and an execution module 403, which are described as follows:

[0119] Acquisition module 401, in the present application, is mainly used to perform convolution recognition operation on the character structure in the target Tibetan text image through the image reading device in the Tibetan translation device after starting the Tibetan translation device, thereby obtaining a Tibetan character structure sequence;

[0120] Processing module 402, in the present application, is used to extract syllable vectors for each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector for each Tibetan character structure;

[0121] The processing module 402 in the present application is also used to determine the inline relationship of the syllables between each Tibetan character structure and other Tibetan character structures according to the confidence distance between each syllable vector, and then determine the syllable loss amount generated during the character segmentation process during the syllable vector extraction process through all the inline relationships and the probability of co-occurrence between each syllable vector;

[0122] The processing module 402 in the present application is also used to obtain the light intensity data when the image reading device recognizes the target Tibetan text image, and determine the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation through the light intensity data and the Tibetan character structure sequence, and then determine the character recognition loss of the image reading device in the process of recognizing Tibetan text characters through the convolution interference amount and the syllable loss amount;

[0123] Execution module 403, in the present application, execution module 403 is mainly used to read and compensate the syllable vector of each Tibetan character structure based on the character recognition loss, obtain the syllable compensation vector of each Tibetan character structure, and complete the image reading of the Tibetan translation device based on all the syllable compensation vectors.

[0124] In addition, the present application also provides a computer device, which includes a memory and a processor, wherein the memory stores codes, and the processor is configured to obtain the codes and execute the above-mentioned image reading method based on the Tibetan character structure.

[0125] In some embodiments, reference Figure 4 , which is a schematic diagram of the structure of a computer device for implementing an image reading method based on a Tibetan character structure according to some embodiments of the present application. The image reading method based on a Tibetan character structure in the above embodiment can be Figure 4 The computer device 500 shown in the figure is implemented, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503 and at least one communication interface 504.

[0126] The processor 501 may be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0127] The communication bus 502 may be used to transmit information between the above-mentioned components.

[0128] The memory 503 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, an optical disc storage (including a compressed optical disc, a laser disc, an optical disc, a digital versatile disc, a Blu-ray disc, etc.), a magnetic disk or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory 503 may exist independently and be connected to the processor 501 via the communication bus 502. The memory 503 may also be integrated with the processor 501.

[0129] The memory 503 is used to store the program code for executing the solution of the present application, and the execution is controlled by the processor 501. The processor 501 is used to execute the program code stored in the memory 503. The program code may include one or more software modules. The image reading method based on the Tibetan character structure in the above embodiment can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0130] The communication interface 504 uses any transceiver or other device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0131] In a specific implementation, as an embodiment, a computer device may include multiple processors, each of which may be a single-CPU processor or a multi-CPU processor. The processor here may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0132] The above-mentioned computer device may be a general-purpose computer device or a special-purpose computer device. In a specific implementation, the computer device may be a desktop computer, a portable computer, a network server, a personal digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device or an embedded device. The embodiment of the present application does not limit the type of computer device.

[0133] In addition, the present application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned image reading method based on the Tibetan character structure is implemented.

[0134] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0135] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An image reading method based on Tibetan character structure, characterized in that: The steps include: Starting a Tibetan translation device, and performing a convolutional recognition operation on a character structure in a target Tibetan text image using an image reading device in the Tibetan translation device, thereby obtaining a Tibetan character structure sequence; Extracting syllable vectors for each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector for each Tibetan character structure; The syllable in-line relationship between each Tibetan character structure and other Tibetan character structures is determined based on the confidence distance between each syllable vector. The syllable loss caused by character segmentation during the syllable vector extraction process is then determined based on all the in-line relationships and the probability of co-occurrence between each syllable vector. obtaining light intensity data when the image reading device recognizes the target Tibetan text image, determining, using the light intensity data and the Tibetan character structure sequence, an amount of convolution interference caused by changes in ambient light on character recognition during a convolution recognition operation, and further determining, using the convolution interference amount and the syllable loss amount, a character recognition loss during the Tibetan text character recognition process by the image reading device; The syllable vector of each Tibetan character structure is read and compensated according to the character recognition loss to obtain the syllable compensation vector of each Tibetan character structure, and the image reading of the Tibetan translation device is completed based on all the syllable compensation vectors.

2. The method according to claim 1, wherein Extracting syllable vectors for each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector for each Tibetan character structure specifically includes: performing character segmentation on the Tibetan character structures in the Tibetan character structure sequence using a pre-trained Tibetan language recognition model to obtain a plurality of independent Tibetan character structures; All Tibetan character structures are converted into vectors to obtain the syllable vector of each Tibetan character structure.

3. The method according to claim 1, wherein The syllable inline relationship between each Tibetan character structure and other Tibetan character structures is determined based on the confidence distance between each syllable vector, specifically including: Determine the distance between each syllable vector; Determine the confidence distance between corresponding syllable vectors by the distance between each syllable vector; Determine the co-occurrence discrimination of each syllable vector based on all confidence distances; Determine the correlation coefficient of the corresponding syllable vector according to the frequency of occurrence of each syllable vector; The syllable-to-syllable relationship between each Tibetan character structure and other Tibetan character structures is determined by the confidence distance between each syllable vector, all co-occurrence discriminations and all correlation coefficients.

4. The method according to claim 1, wherein The syllable loss caused by character segmentation during syllable vector extraction is determined by all inline relationships and the probability of co-occurrence between each syllable vector. Specifically, the following are included: Acquiring a predetermined amount of Tibetan text data; Constructing a co-occurrence matrix of syllable vectors using the Tibetan text data; Determining the probability of co-occurrence between the syllable vectors according to the co-occurrence matrix; Determine the conversion error when converting each Tibetan character structure into a syllable vector through all probabilities and all internal relationships; The amount of syllable loss caused by character segmentation during the syllable vector extraction process is determined based on all conversion errors.

5. The method according to claim 1, wherein Determining the convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation by using the light intensity data and the Tibetan character structure sequence specifically includes: Determine the fluctuation entropy of the light intensity when the external environment light changes through the light intensity data; performing a deconvolution operation on the Tibetan character structure sequence, thereby extracting a convolution residual during the convolution recognition operation; The convolution interference amount of the ambient light change on the character recognition during the convolution recognition operation is determined according to the fluctuation entropy and the convolution residual.

6. The method according to claim 1, wherein Determining the character recognition loss of the image reading device in the process of recognizing Tibetan text characters based on the convolution interference amount and the syllable loss amount specifically includes: Obtaining historical convolution interference and historical syllable loss; Determining a convolution loss weight of the convolution interference amount by using a historical convolution interference amount; Determining a syllable loss weight of the syllable loss amount according to the historical syllable loss amount; The convolution interference amount and the syllable loss amount are weightedly fused according to the convolution loss weight and the syllable loss weight to obtain the character recognition loss of the image reading device in the process of recognizing Tibetan text characters.

7. The method according to claim 1, wherein The Tibetan language recognition model built into the image reading device in the Tibetan language translation device is a convolutional neural network model.

8. A translation device based on the Tibetan character structure, comprising an image reading unit, characterized in that: The image reading unit includes: an acquisition module, configured to, after starting the Tibetan translation device, perform a convolution recognition operation on the character structure in the target Tibetan text image through an image reading device in the Tibetan translation device, thereby obtaining a Tibetan character structure sequence; a processing module, configured to extract a syllable vector for each Tibetan character structure in the Tibetan character structure sequence based on a pre-trained Tibetan language recognition model to obtain a syllable vector for each Tibetan character structure; The processing module is further configured to determine the syllable inline relationship between each Tibetan character structure and other Tibetan character structures based on the confidence distance between each syllable vector, and further determine the syllable loss amount generated during character segmentation during the syllable vector extraction process based on all the inline relationships and the probability of co-occurrence between each syllable vector; The processing module is further configured to obtain light intensity data when the image reading device recognizes the target Tibetan text image, determine the amount of convolution interference caused by ambient light changes on character recognition during the convolution recognition operation based on the light intensity data and the Tibetan character structure sequence, and further determine the character recognition loss of the image reading device during the Tibetan text character recognition process based on the convolution interference and the syllable loss. An execution module is used to read and compensate the syllable vector of each Tibetan character structure based on the character recognition loss to obtain the syllable compensation vector of each Tibetan character structure, and complete the image reading of the Tibetan translation device based on all the syllable compensation vectors.

9. A computer device comprising a memory and a processor, wherein the memory stores a code, wherein: The processor is configured to obtain the code and execute the image reading method based on the Tibetan character structure according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the image reading method based on the Tibetan character structure according to any one of claims 1 to 7 is implemented.

Images