JP3244068U - Handwritten character recognition system based on neural network - Google Patents

Abstract

The present invention provides a handwritten character recognition system based on a neural network that has a high recognition rate including languages written from right to left and supports recognition of appropriate stroke order. [Solution] A handwritten character recognition system includes a handwritten handwriting input processor used to receive a handwritten character string handwritten trajectory online and perform preprocessing, and a large number of character strings processed by the handwritten handwriting input processor. It comprises a character string database used to store handwriting trajectories, and a character identification neural network including an encoder, decoder, input interface, and output interface. Each handwritten stroke in the character string handwriting trajectory of the character string database is input to an encoder and a decoder and identified multiple times to obtain multiple possible character strings. The identified string is output to the post-processor via the output interface and normalized. [Selection diagram] Figure 1

Description

TECHNICAL FIELD The present invention relates to a character identification system, and more particularly to a handwritten character identification system based on a neural network.

Online handwritten Chinese character recognition technology is a technology that converts the dynamic handwriting of one line of handwritten characters into a character string. Compared to print recognition techniques, handwriting recognition is more difficult because it requires consideration of different writing habits of each person. Compared to offline handwriting identification, online handwriting identification has more time-series information and less background noise, so it can achieve higher accuracy.

In the prior art, there are two types of online handwriting recognition methods with high recognition rates.One is based on a partition decoding mechanism, which first divides the stroke order, and then applies dynamic programming to find the optimal one by a classifier. Select the division method and the corresponding identification result. MyScript and early Google adopted this method, and its advantage is that it outputs accurate correspondence. The second method does not explicitly partition and is generally based on a bidirectional circular neural network and a CTC (Connectionist Temporal Classification) decoder. Companies that have adopted this method include Samsung and Google, and the advantage of this method is that it is easy to implement and the associated models can be trained using an end-to-end training method. It is a point.

However, the conventional character identification technology described above does not have a high recognition rate, and cannot identify the major natural languages of the world (including languages written from right to left) with high accuracy, or cannot identify the major natural languages of the world (including languages written from right to left), or I couldn't identify the letters written in order.

As a result of extensive studies, the present inventors have found that the above object can be achieved by adopting the configuration of a handwritten character recognition system based on a new neural network, and have completed the present invention.

The present invention was developed through intensive research by the inventor in view of the above-mentioned problems, and its purpose is to provide a handwritten character recognition system based on a neural network.

In order to solve the above problems, a handwritten character identification system based on a neural network according to an aspect of the present invention includes a processor and a memory, the processor is used to perform operations necessary for the system, and the memory is used to perform operations necessary for the system. It is used to store computer program data or calculation programs of the system, related operation results and related data. The system includes:
A handwriting input processor used to receive a character string handwriting trajectory handwritten online, wherein the character string handwriting trajectory is decomposed into a plurality of handwritten handwritings, and in a training stage and a prediction stage of a neural network, The handwriting input processor performs pre-processing on each handwritten stroke of the character string handwritten trajectory, and in a training phase, these handwritten strokes of the character string handwritten trajectory correspond to a known character string and have at least one character. a handwritten handwriting input processor configured such that a character string sample is formed by the character string handwritten trajectory and its corresponding character string;
a character string database connected to the handwritten handwriting input processor and used to store a large number of handwritten character string trajectories processed by the handwritten handwriting input processor; a character string database that stores character string samples formed by column handwritten trajectories and character strings to which they correspond;
A character identification neural network comprising an encoder, a decoder, an input interface located at the encoder, and an output interface located at the decoder, the input interface being connected to the character string database; used to receive each handwritten stroke of the string handwritten trajectory, the encoder finds a context matrix and an initial state used by the decoder in each handwritten stroke, and the decoder receives the context matrix and the initial state. At the same time, the portion and reliability of the pre-output string are output from the output interface, and the portion and reliability are re-inputted into the decoder to identify them. After repeating this several times, possible A plurality of character strings and their corresponding confidence levels are obtained and stored in a result candidate set, the result candidate set is outputted to the outside via the output interface, and in the training stage, each of the character string handwritten trajectories is a character identification neural network in which handwritten handwriting is input to the character identification neural network via the input interface, and a character string to which the character string handwritten trajectory corresponds is a preliminary output of the character identification neural network;
The character strings of the result candidate set, which are connected to the character string database and the output interface of the character identification neural network and output by the output interface of the character identification neural network, are normalized by the post-processor, and the character strings are normalized by the post-processor. The apparatus further includes a post-processor for outputting the character identification neural network.

According to the present invention, it has a higher recognition rate, can more accurately identify major natural languages around the world (including right-to-left languages), and can improve the correspondence between input strokes and output characters. Provide relationships. Supports identification of appropriate stroke order. In actual application, the present invention supports application to open or closed dictionaries by arranging necessary character recognition ranges. The present invention can also be applied to instant handwritten character recognition on general mobile phones and tablet terminals.

1 is a schematic configuration diagram showing the main parts of a handwritten character identification system based on a neural network according to an embodiment of the present invention. FIG. 3 is a schematic diagram illustrating the application of the training stage of the present invention; FIG. 2 is a block diagram showing a sample string of the present invention. FIG. 3 is a schematic diagram illustrating the application of the prediction stage of the present invention; 1 is a schematic configuration diagram showing the hardware structure of a handwritten character identification system based on a neural network according to an embodiment of the present invention; FIG.

The present invention will be described below through embodiments of the invention, but the following embodiments are not intended to limit the invention claimed as a utility model. Furthermore, not all combinations of features described in the embodiments are essential to the proposed solution.

First, a handwritten character identification system 1 based on a neural network according to an embodiment of the present invention mainly includes a processor 2 and a memory 6 (see FIG. 5). The processor 2 is used to perform the operations required by the system 1. The memory 6 is used to store computer program data or calculation programs of the system 1, related operation results, and related data. All operation results and related data are stored in the memory 6.

The handwritten character recognition system based on the neural network of the present invention (see Figures 1 to 5) has a problem in that the model is constructed as one Seq2Seq (sequence to sequence) for character recognition, and the input for the problem is one sequence. (including the x-coordinate, y-coordinate, and handwritten end mark), the output is a line of characters, and the lengths of both are all non-fixed values. The present invention adopts an encoder-decoder framework based on attention mechanism to solve this problem.

A handwritten character identification system based on a neural network according to an embodiment of the present invention mainly includes the following configuration.

<Handwritten handwriting input processor 10>
It is used to receive a character string handwritten trajectory 80 handwritten online, and the character string handwritten trajectory 80 is decomposed into a plurality of handwritten strokes 81 . In the training and prediction stages of the neural network, the handwritten handwriting input processor 10 performs preprocessing on each handwritten handwriting 81 of the character string handwritten trajectory 80 . In the training stage, these handwritten strokes 81 of the character string handwritten trajectory 80 correspond to a known character string 90 and are composed of at least one character 91, and the character string handwritten trajectory 80 and its corresponding character string 90 form a character string. A sample 100 is formed (see Figure 3). The preprocessing method of the handwritten handwriting input processor 10 is as follows.

<Normalization of handwritten handwriting order>
A bounding box is determined for each handwritten handwriting 81, and the bounding box can accommodate the minimum box of the handwritten handwriting 81. So-called normalization is to rearrange all handwritten handwritings 81 from left to right using the left side position of the bounding box of each handwritten handwriting 81 as a reference.

<Estimate row height>
If the length of one handwritten handwriting 81 is close to the width of its bounding box, the width of the bounding box is the measure of the handwritten handwriting; otherwise, the height of the bounding box is the measure of the handwritten handwriting. . A certain quantile of these handwritten strokes is estimated as the line height.

<Perform resampling>
Each handwritten handwriting 81 is displayed with a plurality of sampling points 83, and when the distance between the sampling point 83 and the previous sampling point 83 is some multiple of the estimated line height. , the sampling point 83 is deleted. If the included angle between the sampling point 83 and the sampling points 83 in front and behind the sampling point approaches a flat angle (180 degrees), the sampling point 83 is deleted. By doing so, the number of sampling points 83 can be greatly reduced.

<Normalization of coordinates>
Divide the coordinates of each sampling point 83 by the estimated row height.

In the training stage, each handwritten handwriting 81 of the character string handwritten trajectory 80 is preprocessed by the handwritten handwriting input processor 10 and used as an input to the neural network, and the character string 90 to which the character string handwritten trajectory 80 corresponds is processed by the neural network. Use it as a preliminary output of the network.

<Character string database 20>
It is connected to the handwritten handwriting input processor 10 and is used to store a large amount of character string handwritten trajectories 80 processed by the handwritten handwriting input processor 10. During the training stage, the character string database 20 stores a character string sample 100 formed by the character string handwritten trajectory 80 and a character string 90 corresponding thereto. Each handwritten handwriting 81 of each character string handwritten trajectory 80 is displayed as a plurality of sampling points 83. The first sampling point 83 of each handwritten handwriting 81 is displayed with coordinates (0,0), and the remaining sampling points 83 are displayed by applying the displacement vector to the previous sampling point 83. The last sampling point 83 of has an end mark.

In order to increase the generalized discrimination ability of the entire system, a data augmentation method is used to increase the quantity of string samples 100. The present invention further comprises a corpus and handwriting sample library 35 and a character string sample generation unit 30 connected to the handwriting input processor 10, wherein the corpus and handwriting sample library 35 includes various characters and words of a specific language, Also, samples of handwritten handwriting 81 corresponding to these are stored. The character string sample generation unit 30 extracts a large number of character strings 90 and handwritten handwriting 81 corresponding to them from the corpus and handwritten sample library 35, and generates character string handwritten trajectories 80 corresponding to these character strings 90. A large number of character string samples 100 are synthesized, and after being processed by the handwriting input processor 10, they are stored in the character string database 20 and used for training a neural network. The method by which the character string sample generation unit 30 synthesizes the character string handwritten trajectory 80 includes the following steps.
(1) After rendering the character string 90 using a text typesetting algorithm and recording the position of each character/word in the character string 90, The handwritten strokes 81 are retrieved and placed in corresponding positions, and the normalized handwritten stroke order and line height estimates of each handwritten stroke 81 of these characters/words are obtained and input into the handwritten stroke input processor 10. Then, resampling and coordinate normalization are performed to obtain the corresponding string sample 100.

The levels of characters, words, lines, etc. in the character string 90 are classified, and a random affine transformation is performed on the handwritten handwriting 81 corresponding to the characters, to evaluate the diversity of the corresponding handwritten character string trajectory 80. Further, we obtain more string samples 100 for use in training the neural network.
(2) training a Seq2Seq translation model using a text line sample with an annotation, the annotation indicating a corresponding string 90 of the text line sample, and the Seq2Seq translation model The character string sample 100 is converted into a trajectory 80 and used to generate a corresponding character string sample 100.

The string sample generation unit 30 stores the obtained string samples 100 in the string database 20.

<Character identification neural network 40>
It is configured to include an encoder 41, a decoder 42, an input interface 43 located at the encoder 41, and an output interface 44 located at the decoder 42. The input interface 43 is connected to the character string database 20 and is used to receive each handwritten stroke 81 of the character string handwritten trajectory 80. The encoder 41 is used to find a context matrix and an initial state used by the decoder 42 in each handwritten stroke 81, and the decoder 42 receives the context matrix and the initial state, and also receives the initial state from the output interface 44. The portion and reliability of the output character string are output, and the portion and reliability are regressively input to the decoder 42 for identification. After repeating this several times, a plurality of possible character strings 90 and their corresponding confidence ratings are obtained and stored in the result candidate set 70. The result candidate set 70 is outputted to the outside by the output interface 44.

As shown in FIG. 2, in the training stage, each handwritten stroke 81 of the character string handwritten trajectory 80 is input to the character identification neural network 40 through the input interface 43, and the character string 90 to which the character string handwritten trajectory 80 corresponds is This is the preliminary output of the character identification neural network 40.

<Post processor 50>
It is connected to the character string database 20 and the output interface 44 of the character identification neural network 40 . The character string 90 of the resultant candidate set 70 outputted by the output interface 44 of the character identification neural network 40 is normalized by the post-processor 50 and becomes the output of the character identification neural network 40 . The normalization method is as follows.
(1) Normalize using Unicode NFKD format (Normalization Form KD). Mainly to represent the same characters in different formats and normalize them as the same program code.
(2) By converting the order of Unicode code points from the normal logical order to the visual order (left to right), the positions are approximately aligned with the arranged input. In the prediction stage, predictable identification results are usually normalized, so it is necessary to normalize the identified string 90 using Unicode's NFKC, which changes the order of Unicode code points from the visual order to the logical order. Convert back to order.

For example, characters in parts of the character string 90 that have different sequences are arranged so that they have the same sequence. This makes identification easier when training a neural network.

The encoder 41 includes a multilayer bidirectional LSTM (Long Short-Term Memory) cyclic neural network and a fully-connected neural network (FNN). Each handwritten stroke 81 input to the encoder 41 enters a multi-layer bi-directional LSTM (Long Short-Term Memory) cyclic neural network, and the output of the multi-layer bi-directional LSTM (Long Short-Term Memory) cyclic neural network is a context matrix. After performing global average pooling on the context matrix, the output of the fully-connected neural network is input to the decoder 42 via the fully-connected neural network (FNN). This is the initial state for.

The input to the decoder 42 includes input characters, decoding states, convergence vectors, and the context matrix. The decoder 42 performs multi-step decoding, and when decoding the first step, the initial value of the input character is one built-in starting code, and the initial value of the convergence vector is a zero vector; The initial value of the decoding state is the initial state output by the encoder 41, and the context matrix is output by the encoder 41. The output of the decoder 42 is the next character and its confidence distribution, a new decoding state, and a new convergence vector, the next character being a new input character, and the output of the decoder 42 with a new decoding state and a new convergence vector. and perform the next step of decoding. The decoder 42 repeatedly performs multiple steps of decoding using the method described above. The dimension of said convergence vector indicates each sampling point 83, its action records which sampling point 83 already has a corresponding output character, and the initial coordinate of each dimension is 0. If a corresponding output character exists, the dimensional coordinates corresponding to the sampling point 83 are set to 1. In this way, the convergence vector is applied to reduce the situation of lack or overabundance of characters.

The decoder 42 defines a loss function that is a linear combination of a plurality of functions, and defines a meaningful loss function as necessary during calculation. The loss function is used to calibrate the cross-entropy of the confidence distribution and guide the all-ones mean squared error at the end of the convergence vector. If some known character corresponds to the handwriting 81 (e.g. the data has character-level annotations or is a composite), then further guidance to decode one character will cause the convergence vector to correspond to the handwriting 81. The amount corresponding to increases by 1. By doing this, in the prediction stage, it is determined which handwritten stroke 81 each output character corresponds to based on the change in the convergence vector.

In the prediction stage, the decoder 42 gradually constructs candidate rows using a beam search method, and for each incomplete candidate row, determines its character string, reliability, decoding state, convergence vector, and predictive analysis. Record the stack (predictive analysis stack). The predictive analytics stack is a stack containing grammar strings of predictive analytics. If character-level candidates need to be output, some other candidate characters for each character and the corresponding reliability are also recorded. If it is necessary to output the correspondence between handwritten handwriting 81 and characters, a historical convergence vector is further recorded. In each step of decoding, the decoder 42 performs the following steps for each unfinished candidate row.
(1) Activate the decoder 42 to obtain the next character candidate list and the corresponding reliability, and update the decoding state and convergence vector at the same time.
(2) The output interface 44 outputs each candidate for the next character to a confidence adjustment unit 60 coupled between the output interface 44 and the post-processor 50, and the confidence adjustment unit 60 outputs n- Calculate the probability of the string obtained using the gram method and adjust the confidence level of each candidate for the next character.
(3) For each candidate next character, calculate the reliability and predictive analysis stack after combination with the original character string.
(a) If the candidate is a termination symbol and the predictive analysis stack is blank, obtain one completed candidate line.
(b) If the candidate is not a termination symbol and the predictive analysis stack is not in an error state, obtain one incomplete candidate line and use it for processing during the next step of decoding. At the beginning of each next step of decoding, the decoder 42 reserves only a few candidate lines with the highest reliability. When the number of already completed candidate lines reaches a predetermined value, decoding is stopped, and at this time, the completed candidate lines are some possible strings 90, and these strings 90 and their reliability are evaluated as above results. It is stored in the candidate set 70 and output to the outside through the output interface 44.

In the prediction stage (see FIG. 4), each handwritten handwriting 81 of the character string handwritten trajectory 80 to be identified is processed by the handwritten handwriting input processor 10, and then input to the character identification neural network 40 using the character string database 20 for identification. The candidate result set 70 obtained after processing the identification results by the post-processor 50 includes some character strings 90 that may correspond to the handwritten character string trajectory 80 and their reliability.

The invention may be embodied in various other forms without departing from its spirit or essential characteristics. Therefore, the above-described embodiments are merely illustrative in all respects, and should not be interpreted in a limiting manner. The scope of the present invention is indicated by the claims of the utility model registration, and is not restricted in any way by the main text of the specification. Furthermore, all modifications and changes that fall within the scope of equivalents of the claims for utility model registration are within the scope of the present invention.

1 Handwritten character identification system 2 Processor 6 Memory 10 Handwritten handwriting input processor 20 Character string database 30 Character string sample generation unit 35 Handwritten sample library 41 Encoder 42 Decoder 43 Input interface 44 Output interface 50 Post processor 60 Reliability adjustment unit 70 Result candidate set 80 Character string handwritten trajectory 81 Handwritten handwriting 83 Sampling point 90 Character string 100 Character string sample

Claims (12)

comprising a processor and a memory, the processor is used to perform operations necessary for the system, and the memory is used to store computer program data or arithmetic programs, related operation results and related data of the system; A handwritten character recognition system based on a neural network,
A handwriting input processor used to receive a character string handwriting trajectory handwritten online, wherein the character string handwriting trajectory is decomposed into a plurality of handwritten handwritings, and in a training stage and a prediction stage of a neural network, The handwriting input processor performs pre-processing on each handwritten stroke of the character string handwritten trajectory, and in a training phase, these handwritten strokes of the character string handwritten trajectory correspond to a known character string and have at least one character. a handwritten handwriting input processor configured such that a character string sample is formed by the character string handwritten trajectory and its corresponding character string;
a character string database connected to the handwritten handwriting input processor and used to store a large number of handwritten character string trajectories processed by the handwritten handwriting input processor; a character string database that stores character string samples formed by column handwritten trajectories and character strings to which they correspond;
A character identification neural network comprising an encoder, a decoder, an input interface located at the encoder, and an output interface located at the decoder, the input interface being connected to the character string database; used to receive each handwritten stroke of the string handwritten trajectory, the encoder finds a context matrix and an initial state used by the decoder in each handwritten stroke, and the decoder receives the context matrix and the initial state. At the same time, the portion and reliability of the pre-output string are output from the output interface, and the portion and reliability are re-inputted into the decoder to identify them. After repeating this several times, possible A plurality of character strings and their corresponding confidence levels are obtained and stored in a result candidate set, the result candidate set is outputted to the outside via the output interface, and in the training stage, each of the character string handwritten trajectories is a character identification neural network in which handwritten handwriting is input to the character identification neural network via the input interface, and a character string to which the character string handwritten trajectory corresponds is a preliminary output of the character identification neural network;
The character strings of the result candidate set, which are connected to the character string database and the output interface of the character identification neural network and output by the output interface of the character identification neural network, are normalized by a post-processor to A handwritten character recognition system based on a neural network, further comprising: a post-processor for outputting the output of the recognition neural network. The preprocessing method of the handwritten handwriting input processor is to normalize the handwritten handwriting order, estimate the line height, sample again, and finally normalize the coordinates, and the coordinates of each sampling point are the line height. 2. The handwritten character identification system based on a neural network according to claim 1, wherein the handwritten character identification system is estimated by dividing by . further comprising a character string sample generation unit connected to a corpus and handwriting sample library and the handwritten handwriting input processor, wherein the corpus and handwriting sample library includes samples of various characters and words of a specific language and their corresponding handwritten handwritings; is stored, and the character string sample generation unit extracts a large number of character strings and their corresponding handwritten strokes from the corpus and handwritten sample library, and synthesizes these character strings with their corresponding character string handwritten trajectories. 2. The method according to claim 1, wherein a large number of character string samples are formed by the handwriting input processor and stored in the character string database after being processed by the handwriting input processor and used for training the neural network. Handwritten character identification system. The character string sample generation unit synthesizes the character string handwritten trajectory by rendering the character string using a text composition algorithm and recording the position of each character/word in the character string. , by searching the handwritten strokes of each character/word from the corpus and handwriting sample library and placing them in the corresponding positions, the normalized handwritten stroke order and line height of each handwritten stroke in these characters/words can be determined. obtain an estimate of , and input it into the handwritten handwriting input processor to perform sampling and coordinate normalization again to obtain a corresponding character string sample;
Classify the levels of characters, words, lines, etc. in the character string, and perform random affine transformation on the handwritten handwriting to which they correspond to further increase the diversity of the corresponding character string handwritten trajectory. A handwritten character recognition system based on a neural network according to claim 3. The string sample generation unit synthesizes the string handwritten trajectory by training a Seq2Seq translation model using text line samples with annotations, and the annotations indicate corresponding strings in the text line samples. 4. The handwritten character based on neural network according to claim 3, wherein the Seq2Seq translation model is used to convert a character string into a corresponding character string handwritten trajectory and generate a corresponding character string sample. identification system. The method by which the post-processor performs normalization is characterized in that the normalization is performed in Unicode NFKD format (Normalization Form KD), and the characters are mainly expressed as the same character in different formats and normalized as the same program code. A handwritten character recognition system based on the neural network according to item 1. The method by which the post-processor performs normalization is to convert the order of Unicode code points from the normal logical order to the visual order, roughly aligning the positions with the arranged input, and applying the Unicode code to the identified string. 2. The neural network-based handwritten character identification system according to claim 1, wherein the system performs NFKC normalization and converts the order of Unicode code positions from visual order to logical order. The encoder includes a multi-layer bi-directional LSTM (Long Short-Term Memory) circular neural network and a fully-connected neural network (FNN), and each handwritten stroke input to the encoder is processed by a multi-layer bi-directional LSTM (Long Short-Term Memory). Short-Term Memory) cyclic neural network, the output of the multi-layer bidirectional LSTM cyclic neural network is a context matrix, and after performing Global Average Pooling on the context matrix, The neural network according to claim 1, wherein the neural network is passed through the fully-connected neural network (FNN), and the output of the fully-connected neural network is an initial state to be input to the decoder. A handwritten character identification system based on Each handwritten stroke of each character string handwritten trajectory is displayed by a plurality of sampling points, the first sampling point of each handwritten stroke is displayed with coordinates (0,0), and the remaining sampling points are the previous sampling points. The displacement vector for the point is applied and displayed, and the last sampling point of each handwritten stroke has an end mark;
The input of the decoder includes an input character, a decoding state, a convergence vector, and the context matrix, and the decoder performs multiple steps of decoding, and when decoding the first step, The initial value is a built-in one starting code, the initial value of the convergence vector is a zero vector, the initial value of the decoding state is the initial state output by the encoder, and the context matrix is output by the encoder. and the output of the decoder is the next character and its confidence distribution, a new decoding state, and a new convergence vector, where the next character is a new input character, and the new decoding state and new convergence vector are both The recursive input is input to the decoder and the next step is decoded, and the decoder repeatedly performs multiple steps of decoding by applying the above-mentioned method, the dimension of the convergence vector indicates each sampling point, and its action is Record which sampling point already has a corresponding output character, the initial coordinate of each dimension is 0, and if the corresponding output character exists, set the dimension coordinate corresponding to said sampling point to 1; A handwritten character identification system based on a neural network according to claim 8. A loss function that is a linear combination of multiple functions is defined in the decoder, and a meaningful loss function is defined as necessary during calculation, and the loss function is based on the cross-entropy of the confidence distribution. entropy) and is used to guide a mean squared error of all 1 at the end of the convergence vector, further guiding to decode one character if a known character corresponds to handwriting; 10. The method of claim 9, wherein the amount of the convergence vector corresponding to a point related to handwritten handwriting is increased by 1, and in the prediction step, it is determined based on the change in the convergence vector to which handwritten handwriting each output character corresponds. A handwritten character recognition system based on a neural network described in . In the prediction stage, the decoder gradually constructs candidate rows using a beam search method, and for each incomplete candidate row, stores its character string, reliability, decoding state, convergence vector, and predictive analysis stack. (predictive analysis stack), said predictive analysis stack is a stack containing grammatical symbol strings of predictive analytics method (predictive analytics), and if character-level candidates need to be output, some other The neural network according to claim 9, further comprising recording candidate characters and corresponding confidence levels, and further recording historical convergence vectors when it is necessary to output a correspondence between handwritten strokes and characters. based handwritten character identification system. For each unfinished candidate row in each step of decoding, the decoder
(1) applying the decoder to obtain the next character candidate list and the corresponding reliability, and simultaneously updating the decoding state and convergence vector;
(2) outputting each candidate of the next character via the output interface to a reliability adjustment unit coupled between the output interface and the post-processor, the reliability adjustment unit using an n-gram method; calculating the probability of a character string to be obtained by applying , and adjusting the confidence level of each candidate for the next character;
(3) calculating a confidence level and a predictive analysis stack after combination with the original string for each candidate next character;
(a) if the candidate is a termination symbol and the predictive analysis stack is blank, obtaining a completed candidate line;
(b) if the candidate is not a termination symbol and the predictive analysis stack is not in an error state, obtaining one incomplete candidate line and processing it during decoding of the next step; , execute
Each time at the start of decoding the next step, the decoder only reserves some candidate rows with the highest confidence, and if the quantity of already completed candidate rows reaches a predetermined value, it stops decoding and at this time 12. The candidate rows are some possible character strings, and these character strings and their reliability are stored in the result candidate set and outputted to the outside by the output interface. A handwritten character recognition system based on the described neural network.