CN111241984A - Chinese character online Latin type cursive input and intelligent recognition method and system - Google Patents

Abstract

The invention discloses an online Latin type cursive input and intelligent recognition method and system for Chinese characters, which comprises the following specific steps: step 1: deeply mining the information of the cursive Chinese characters, developing a Latin type cursive script font library, and step 2: the Latin type cursive script character convenient input and intelligent cognition algorithm comprises the following steps: the method comprises the steps of feature selection and wavelet neural network, the method is simple and convenient for feature extraction of characters through the feature point extraction method, and the method has high recognition rate and short time consumption when identifying the cursive script, so that the character identification of the cursive script achieves good effect.

Description

Chinese character online Latin type cursive input and intelligent recognition method and system

Technical Field

The invention relates to the field of cross application research of intelligent information processing and wavelet analysis technology of Chinese characters, in particular to an online Latin type cursive input and intelligent identification method and system of Chinese characters.

Background

With the continuous enhancement of the comprehensive national power, the continuous improvement of the international status and the continuous deepening of the reform of the open policy in China, the development concept and the cultural value of China are more and more concerned by the international society. Particularly, after WTO is added in China, Chinese and foreign people are increasingly active, and the demand and the opportunity for non-Chinese people to process and exchange Chinese character information are increased. "the people's daily newspaper overseas edition" 2011 11 month 28 th 04 th edition report: "11/8 th day, the first Confucius school of Switzerland held at the lakeside of Wailang Miscanthus in Japan. In recent years, along with the development of Chinese economy, Chinese releases unprecedented charm, and 'Chinese fever' is hot all over the world. By the end of 2010, foreigners learning chinese worldwide have reached 1 million. From its development trend, chinese is inevitable to move to the world. In the current information and digitalization era, Chinese tends to move to the world, and besides government attention and foreign study measures, a set of Chinese character input technology which is convenient for non-Chinese people to use is also needed. However, although the number of the Chinese character input techniques which have been introduced and put into use is not small and is classified into two categories, namely, the phonetic code input technique and the shape code input technique, the Chinese character input techniques are designed for the group using Chinese as the native language, and the foreign character input techniques for the non-Chinese group are not available. In recent years, although many scholars have paid hard efforts in the field of handwritten character recognition, they have obtained handwriting input products with strong functions such as hanwang and the like and have been widely popularized and applied. However, the use of these products is premised on that users must write Chinese characters, as is well known, the Chinese characters have complex structures and criss-cross strokes, and are completely strange, strange and complicated things like a Chinese character script for non-Chinese people. And non-Chinese people are more difficult to learn to write Chinese characters in the draft of the Chinese character, which completely do not accord with the Latin writing habit and always make the Chinese characters in the draft of the Lively dance. In the existing Chinese character input method, people who are not Chinese characters need to quickly hand write Chinese characters in a hand writing terminal to input and process Chinese character information, namely 'Tianfangpeng'. So far, the bottleneck that non-Chinese people quickly input, process and exchange Chinese character information based on a handwriting terminal platform has not been broken through. Especially, with the popularization of portable Personal Digital Assistant (PDA) devices having only a keypad, such as smart phones, palm computers, etc., it is very inconvenient to directly complete input through the keypad, foreign people learn chinese characters, and if there is no technology capable of conveniently inputting chinese characters on an intelligent terminal, it is impossible to move splendid chinese characters for more than five thousand years to the world, so it is the biggest bottleneck of moving chinese characters to the world. Research and realization of an input method which can break through the bottleneck and really allow non-Chinese people to conveniently input, process and communicate Chinese character information are very important and necessary for inheriting and developing Chinese characters and realizing the Chinese popularization vision of government.

The original intention of contemporary grassland saint in the right-handed ancestor to study the chinese draft gives us many inspirations: the Chinese characters are difficult to be recognized and written by the Chinese characters more than seventy years ago, aiming at 'seeking convenience for making, making the function of culture as much as possible and saving time of all the people and developing the traditional interest of the whole family', the standard grass book 'is created by the long of hundreds of grass books, and the aim is' to accept thousands of years of culture in China to make the Chinese characters highly developed and make every son and sun enjoy a lot of time every day (saving) so as to increase the success rate of the career. There has been a wonderful metaphor for various fonts in old mr: the regular script such as walking, the running script such as a ship and the cursive script such as an airplane are necessary to be mined in order to lead the Chinese character input technology to move to the world, so that non-Chinese people who learn to use Chinese characters can conveniently input the Chinese characters, the airplane performance which can be quickly identified by cursive marks is fully utilized, the western Latin writing habit mode is fused, the Chinese character information convenient input and intelligent cognition system of the simulated Latin handwriting method is realized, the obstacles of the non-Chinese people for Chinese character information convenient input, Chinese character information processing and communication are really cleared away, and the online handwriting input system which is a more natural and humanized man-machine interaction mode is necessary and is an important way for solving the worldwide problem of the foreign Chinese character input of the non-Chinese people all over the world. The invention provides a Chinese character online latin type cursive input and intelligent identification method and system which accord with the writing habit mode of non-Chinese population, so that the non-Chinese population familiar with latin language characters can input Chinese characters as fast and convenient as writing latin characters, the bottleneck that the non-Chinese population carries out convenient input, processing and communication of Chinese character information based on a handwriting terminal is broken through, a channel is really opened for the non-Chinese population to carry out convenient input, processing and communication of Chinese character information, and the method and system plays a due role for the wide popularization of Chinese characters in international communication.

Disclosure of Invention

The invention aims to solve the defects and provides an online Latin type cursive input and intelligent identification method and system for Chinese characters.

The purpose of the invention is realized by the following technical scheme:

the method and system for online Latin type cursive input and intelligent identification of Chinese characters specifically comprise the following steps: step 1: deeply digging the information of the cursive Chinese characters, and developing a Latin type cursive script font library: the core theory of 'word home position' is taken as guidance, according to precious documents such as 'standard cursive script' of mr. on the right in the contemporary grass, standard cursive script 'of Zhang right,' cursive script new interpretation 'of Sungqingchang, and' universal 'and universal' universal cursive script 'of week' and the like, the culture gene of Chinese cursive script is penetrated based on the data mining technology, deep mining of cursive Chinese character information is seriously carried out, the gene codes of cursive script symbols of the past generations in China are decoded, the connection relation of strokes of Chinese character cursive script is cleared, the Chinese character cursive rule is researched and innovated, a set of effective Latin type anti-cursive handwriting rules and schemes of the ontology linguistics with Chinese language system are researched and innovated by using 'new eye light, new thinking and new methods', and a Latin type anti-cursive script character handwriting character library which forms Chinese characters is developed; the Chinese characters are divided into single-body characters and multi-body characters, the Latin type anti-draft handwriting rules of the single-body characters are relatively simple to make, and only the direct deep excavation is needed according to the standard draft; the combined characters have upper and lower structures and left and right structures, and in order to make the combined characters accord with the Latin writing habit of non-Chinese people after being drafted by Latin type, the combined characters need to be processed by a structure transformation technology besides mining corresponding standard cursive Chinese character information;

step 2: the Latin type cursive script character convenient input and intelligent cognition algorithm comprises the following steps:

(1) selecting characteristics: feature extraction is a key to realize recognition of characters represented by a set of features, and the object of feature extraction is to obtain a two-dimensional image to a one-dimensional feature vector X closely related to image information^T(x₁,...,x_m) The purpose of feature extraction is to reduce intra-class variance by increasing inter-class separation, the feature extraction comprising two steps: extracting a first sequence of related features from the normalized feature coordinates, calculating Wavelet Transform (WT) of the features to obtain feature vectors in a compressed form, extracting 6 time domain features in the first step, wherein the first two are x and y coordinates, and then extracting angle features, directions and curvatures, wherein the angle features are extracted because the angle features cannot be changed during translation and scaling, the wavelet transform of the time domain features needs to be performed separately, and only approximation coefficients are used for obtaining the feature vectors;

the first two features are normalized values in the X and Y coordinates, and the local writing directions at points X (t), Y (t) are represented by the sine and cosine of angle α (t), angle α (t) being formed by the horizontal line connecting the forward and backward nodes;

the curvature or angular difference of points x (t), y (t) is described by the sine and cosine of the angle θ (t) along which a counterclockwise rotation of the forward vector may coincide with the backward vector, which is a method of measuring the angular difference between the forward and backward vectors, θ (t) may be calculated according to the following equations (1), (2) and (3):

Δy(t)＝y(t+1)-y(t-1) (3)

the following can be obtained:

the wavelet transform decomposes the original image into 4 approximation sub-bands and 3 detail sub-bands in horizontal, vertical and diagonal directions based on approximation coefficients and detail coefficient representation signals, the process is iterated continuously to predetermined layers to obtain multi-resolution representation of the image, the number of layers depends on the inverse size of the aperture of the wavelet filter applied to the image, when the wavelet transform needs to sample twice from the first layer to the second layer, the maximum layer value of the applied wavelet transform depends on the number of data points in the data set, and for an image of s x s pixels, the relation of the layers M and s can be represented as 2^M＝s/2；

The method adopts a multi-resolution method to generate features, the features are generated through calculation in three stages of local, global and intermediary to realize dynamic character recognition, the features are extracted in a coarse resolution firstly, the high resolution of sub-images must be considered in each iteration next until all classifications must reach acceptance criteria, and a 3-time wavelet filter is used, and the character images are decomposed into 10 sub-band images after 3-scale sym4 wavelet decomposition

Wherein,

the subband images represent the basic shape of the character image and are negligible.

Showing the vertical high frequency components or horizontal details,

representing horizontal high frequency components or vertical details;

representing diagonal details, then subband images

And subband images

Correlating, from subband images

Extracting wavelet relative energy distribution characteristics from detail image components

Extracting chain code histogram feature from k 1,2,3,

and

in the feature calculation of (2), the respective bounding boxes are divided into 8 × 8 blocks, and the number of black pixels is calculated and connected

And

forming a feature vector;

(2) WNN: WNN (wavelet neural network) is a multilayer feedforward network, which takes wavelet theory as the basis and takes discrete wavelet function as the activation function of nodes, and the WNN classification process is divided into three steps, namely, the first step is network initialization, the second step is training of weighting coefficients by using gradient descent algorithm, and the last step is training according to trainingThe wavelet network can be further divided into a series of stages from the perspective of image application, and WNN has a 3-layer structure which is n on an input layer respectively_inA node, n of a hidden layer_hN of one node and output layer_outA node, which selects Mexican hat wavelet as the base, and is defined as expression (7)

The kth input neuron is defined by equation (8):

wherein x_j(j＝1,2,…,n_inIs an input variable, W_j,kRepresenting the weight of the ith input and the kth hidden node connecting line, in order to grasp the level and position of the wavelet, a multi-scale wavelet function is used as the conversion function of the hidden node, the expansion parameter a of the first hidden node is set to 1, namely psi_1,b1(x) ψ (x-b1), the expansion parameter a of the second hidden node is set to 2, i.e.

In which the output result of the wavelet is reduced to

Similarly, the dilation parameter a for the jth hidden node is set to j, so the output of the hidden layer of WNN can be given by equation (9):

the output of the kth neuron is defined by equation (10):

the output of WNN is defined as equation (11):

wherein

ω_l,k,k＝1,2,…,n_h，l＝1,2,…,n_outAnd represents the weight of the link between the kth hidden node and the first output node. In training, the weight, balance and scaling parameters are adjusted to minimize the error T function by equation (12):

where I is 1, …, I is the number of training patterns, K is 1, …, K is the number of targets, D_ikAnd O_ikRespectively represent Node_ikA desired output value and an activation net output value.

The invention has the following beneficial effects:

the method for extracting the characters enables the characters to be simple and convenient in feature extraction through the feature point extraction method, and when the method is used for identifying the cursive script, the identification rate is high, the time consumption is short, and the character identification of the cursive script achieves a good effect.

Drawings

FIG. 1 is a schematic diagram of Pinyin, regular script, cursive script, Latin-type cursive script and English simple translation of some Chinese characters of the present invention;

FIG. 2 is a schematic diagram of similar Latin draft character fonts corresponding to different Chinese characters;

FIG. 3 is a schematic view of a known seal cutting seal according to the present invention;

FIG. 4 is a diagram of handwriting direction and curvature analysis of the present invention;

FIG. 5 is an exploded view of the Latin sketch "know" wavelet of the present invention;

FIG. 6 is a schematic of the 10 sub-bands of the sym4 wavelet after three-level decomposition of the present invention;

FIG. 7 is a diagram of a wavelet network architecture of the present invention;

FIG. 8 is a diagram of the results of the on-line Latin-type cursive input and intelligent cognitive implementation of Chinese character learning.

Detailed Description

The invention is further described with reference to the accompanying drawings in which:

as shown in fig. 1, the method and system for online latin type cursive input and intelligent recognition of chinese characters specifically comprises the following steps: step 1: deeply digging the information of the cursive Chinese characters, and developing a Latin type cursive script font library: the core theory of ' word home ' is taken as guidance, according to precious documents such as ' standard cursive script ' of mr. on the right in the contemporary grass, standard cursive script ' of Zhang right, ' cursive script new interpretation ' of Sungqingchang, and ' universal cursive vocabulary ' of all the week use, the culture gene of Chinese cursive script is penetrated based on the data mining technology, deep mining of cursive Chinese character information is seriously carried out, the gene codes of cursive script symbols of the past generations in China are decoded, the connection relation of strokes of Chinese character cursive script is cleared, the Chinese character cursive rule is cleared, and a set of effective Latin type anti-cursive handwriting rules and schemes which are practical for Chinese people as well as non-Chinese people and have Chinese characteristics are researched and innovated by a new method with ' new eye light, new thought and new method ', and a Latin type anti-cursive handwriting character word library for forming Chinese characters is developed.

The Latin cursive characters are adopted because the Latin cursive characters, cursive characters and regular characters of the Chinese characters in the figure 2 are compared and contrasted, and the Latin cursive characters are written in a Latin cursive manner, so that the Chinese Latin cursive characters are consistent with the stroke and operation rules of the Latin cursive characters, almost the same reason is found, and the writing habits of non-Chinese people are completely met if one track is produced. The cursive script and the Latin cursive script still keep three Chinese elements of 'sound, shape and meaning' of the Chinese characters, and are a model of 'ancient use, ocean use', and Chinese and western combination and combination.

Through intensive research, the seal cutting type seal symmetrical mode which is mutually corresponding, symmetrical and complementary is formed by taking the simplified characters of the script characters as main bodies. FIG. 3 shows a seal cutting type seal symmetrical mode of the 'Zhi' word.

The Chinese characters are divided into single-body characters and multi-body characters, the Latin type anti-draft handwriting rules of the single-body characters are relatively simple to make, and only the direct deep excavation is needed according to the standard draft; the combined characters have upper and lower structures and left and right structures, and in order to make them accord with the Latin writing habit of non-Chinese people after they are drawn by Latin cursive writing, besides digging corresponding standard cursive Chinese character information, they also need to be processed by structure transformation technology.

Step 2: the Latin type cursive script character convenient input and intelligent cognition algorithm comprises the following steps:

(1) feature extraction

Feature extraction is a key to realize recognition of characters represented by a set of features, and the object of feature extraction is to obtain a two-dimensional image to a one-dimensional feature vector X closely related to image information^T(x₁,...,x_m) To (3) is performed. The purpose of feature extraction is to reduce intra-class variance by increasing inter-class separation. This requires that features extracted from samples of the same class should be approximate, while features extracted from samples of different classes should be different.

The feature extraction comprises two steps. A first sequence of relevant features is extracted from the normalized feature coordinates. The Wavelet Transform (WT) of these features is computed to yield a feature vector in compressed form. In the first step, 6 time domain features are extracted. The first two are the x, y coordinates themselves. Angular features, directions and curvatures are then extracted. The angular feature is extracted because it does not change when panning and zooming. The wavelet transformation of these time domain features needs to be performed separately, and only the approximation coefficients are used to derive the feature vector.

The local writing directions at points X (t), Y (t) are represented by the sine and cosine of angle α (t), angle α (t) being formed by the horizontal line connecting the forward and backward nodes.

The curvature or angular difference of the points x (t), y (t) is described by the sine and cosine of the angle θ (t). The counterclockwise rotation of the forward vector along this angle may coincide with the backward vector. This is a method of measuring the angular difference between the forward and backward vectors, as shown in fig. 4. θ (t) can be calculated according to the following equations (1), (2) and (3).

Δy(t)＝y(t+1)-y(t-1) (3)

The following can be obtained:

the wavelet transform represents a signal based on approximation coefficients and detail coefficients. As shown in fig. 5, the original image is decomposed into 4 approximation sub-bands and 3 detail sub-bands in the horizontal, vertical and diagonal directions. The process iterates to a predetermined layer to obtain a multi-resolution representation of the image. The number of layers depends on the inverse size of the wavelet filter aperture applied to the image. When the wavelet transform requires two samples from the first layer to the second layer, the maximum layer value to which the wavelet transform is applied depends on the number of data points in the data set. For an image of s × s pixels, the relationship between the levels M and s can be expressed as 2^M＝s/2。

The feature generation is carried out by adopting a multi-resolution method, and the dynamic character recognition is realized by calculating the features in three stages of local, global and intermediary to generate the features. The features are first extracted in a coarse resolution, and the high resolution of the sub-images must be taken into account in each subsequent iteration until all the classifications have been verifiedAnd (5) receiving the standard. A 3-order wavelet filter is used. As shown in fig. 6, the character image is decomposed into 10 sub-band images after being wavelet-decomposed by 3-scale sym4

Wherein,

the subband images represent the basic shape of the character image and are negligible.

Showing vertical high frequency components or horizontal direction details.

Representing horizontal high frequency components or vertical details.

Diagonal details are shown. Then, the subband image

And subband images

Correlating, from subband images

Extracting the relative energy distribution characteristics of the wavelets. From detail image components

And extracting chain code histogram features from the k-1, 2 and 3.

and

The respective bounding boxes are divided into 8 x 8 blocks in the feature calculation of (1), and the number of black pixels is calculated. Connection of

And

a feature vector is formed.

The zero crossings of the wavelet transform provide the locations of signal changes. The total number of zero crossings is taken as a feature at different levels. The ideal number of multi-resolution levels is obtained by extracting features through wavelet packet transformation of the character image (using the optimal basic algorithm). The scheme of extracting multi-resolution features by using Haar wavelets to consider two feature vectors realizes the recognition of unconstrained handwritten characters. The first uses features at only one level of resolution and the second uses all features at both levels of resolution. A two-dimensional wavelet transform was performed using a spline wavelet CDF 3/7, and an unconstrained handwritten character was recognized using 4 subband images of the coefficients as feature vectors.

(2) WNN: WNN (wavelet neural network) is a multi-layer feedforward network, which takes wavelet theory as the basis and takes discrete wavelet function as the activation function of nodes, the wavelet neural network makes full use of the partial resolution characteristic of wavelet transformation and the nonlinear mapping of artificial neural network, so the defect of BP neural network can be overcome.

The WNN classification process is divided into three steps: the first step is network initialization, the second step is training the weighting coefficient by using a gradient descent algorithm, and the last step is realizing feature classification according to the trained weighting coefficient. From the perspective of image application, the wavelet network can be further divided into a series of stages, wherein each stage is depicted as fig. 7, WNN has a 3-layer structure, n on the input layer_inA node, n of a hidden layer_hN of one node and output layer_outAnd (4) each node.

The selection of the mother wavelet is important in wavelet analysis. The wavelet is localized as a basis function, which is derived by shifting and expanding the mother wavelet. These wavelets form a basis and then represent signals such as images at progressively increasing resolutions of the hierarchy. This multi-resolution analysis enables us to perform image analysis over different frequency bands. Wavelet transformation is one of the most suitable techniques for time and frequency domain analysis of non-stationary signals. It uses local basis functions to capture the local features of the signal. Thus, it provides a better approximation of the signal than fourier transforms, sine transforms, cosine transforms, etc. Because characters differ greatly at each local point, the ability to capture local information is critical. Wavelet analysis provides direct access to information that may be masked in other time and frequency domain analysis methods such as fourier transforms. In our study, the Mexican hat wavelet was chosen as the basis, which is defined as expression (7).

The kth input neuron is defined by equation (8):

wherein x_j(j＝1,2,…,n_inIs an input variable, W_j,kRepresents the weight of the ith input and the kth hidden node connection line. In order to grasp the level and position of the wavelet, a multi-scale wavelet function is used as a conversion function of the hidden node. The dilation parameter a of the first hidden node is set to 1, i.e.. psi_1,b1(x) ψ (x-b1), the expansion parameter a of the second hidden node is set to 2, i.e.

In which the output result of the wavelet is reduced to

Similarly, the dilation parameter a for the jth hidden node is set to j. Thus, the output of the hidden layer of WNN can be given by equation (9):

the output of the kth neuron is defined by equation (10):

the output of WNN is defined as equation (11):

wherein

ω_l,k,k＝1,2,…,n_h，l＝1,2,…,n_outAnd representing the weight of the connecting line of the kth hidden node and the first output node, and in training, adjusting the weight, balancing and scaling parameters by the following formula (12) to minimize an error T function:

where I is 1, …, I is the number of training patterns, K is 1, …, K is the number of targets, D_ikAnd O_ikRespectively represent Node_ikA desired output value and an activation net output value.

Claims (1)

1. The Chinese character on-line Latin type cursive input and intelligent identification method and system are characterized in that: the method comprises the following specific steps: step 1: deeply digging the information of the cursive Chinese characters, and developing a Latin type cursive script font library: the core theory of 'word home position' is taken as guidance, according to precious documents such as 'standard cursive script' of mr. on the right in the contemporary grass, standard cursive script 'of Zhang right,' cursive script new interpretation 'of Sungqingchang, and' universal 'and universal' universal cursive script 'of week' and the like, the culture gene of Chinese cursive script is penetrated based on the data mining technology, deep mining of cursive Chinese character information is seriously carried out, the gene codes of cursive script symbols of the past generations in China are decoded, the connection relation of strokes of Chinese character cursive script is cleared, the Chinese character cursive rule is researched and innovated, a set of effective Latin type anti-cursive handwriting rules and schemes of the ontology linguistics with Chinese language system are researched and innovated by using 'new eye light, new thinking and new methods', and a Latin type anti-cursive script character handwriting character library which forms Chinese characters is developed; the Chinese characters are divided into single-body characters and multi-body characters, the Latin type anti-draft handwriting rules of the single-body characters are relatively simple to make, and only the direct deep excavation is needed according to the standard draft; the combined characters have upper and lower structures and left and right structures, and in order to make the combined characters accord with the Latin writing habit of non-Chinese people after being drafted by Latin type, the combined characters need to be processed by a structure transformation technology besides mining corresponding standard cursive Chinese character information;

step 2: the Latin type cursive script character convenient input and intelligent cognition algorithm comprises the following steps:

(1) selecting characteristics: feature extraction is a key to realize recognition of characters represented by a set of features, and the object of feature extraction is to obtain a two-dimensional image to a one-dimensional feature vector X closely related to image information^T(x₁,...,x_m) The purpose of feature extraction is to reduce intra-class variance by increasing inter-class separation, the feature extraction comprising two steps: extracting a first sequence of related features from the normalized feature coordinates, calculating Wavelet Transform (WT) of the features to obtain feature vectors in a compressed form, extracting 6 time domain features in the first step, wherein the first two are x and y coordinates, and then extracting angle features, directions and curvatures, wherein the angle features are extracted because the angle features cannot be changed during translation and scaling, the wavelet transform of the time domain features needs to be performed separately, and only approximation coefficients are used for obtaining the feature vectors;

the first two features are normalized values in the X and Y coordinates, and the local writing directions at points X (t), Y (t) are represented by the sine and cosine of angle α (t), angle α (t) being formed by the horizontal line connecting the forward and backward nodes;

the curvature or angular difference of points x (t), y (t) is described by the sine and cosine of the angle θ (t) along which a counterclockwise rotation of the forward vector may coincide with the backward vector, which is a method of measuring the angular difference between the forward and backward vectors, θ (t) may be calculated according to the following equations (1), (2) and (3):

Δy(t)＝y(t+1)-y(t-1) (3)

the following can be obtained:

the wavelet transform decomposes the original image into 4 approximation sub-bands and 3 detail sub-bands in horizontal, vertical and diagonal directions based on approximation coefficients and detail coefficient representation signals, the process is iterated continuously to predetermined layers to obtain multi-resolution representation of the image, the number of layers depends on the inverse size of the aperture of the wavelet filter applied to the image, when the wavelet transform needs to sample twice from the first layer to the second layer, the maximum layer value of the applied wavelet transform depends on the number of data points in the data set, and for an image of s x s pixels, the relation of the layers M and s can be represented as 2^M＝s/2；

By using a multi-resolution methodPerforming feature generation, generating features through calculation in three stages of local, global and intermediary to realize dynamic character recognition, firstly extracting features in coarse resolution, and taking high resolution of sub-images into consideration in each iteration until all classifications reach acceptance criteria, and decomposing the character image into 10 sub-band images after 3-scale sym4 wavelet decomposition by using a 3-order wavelet filter

Wherein,

the subband images represent the basic shape of the character image and are negligible.

Showing the vertical high frequency components or horizontal details,

representing horizontal high frequency components or vertical details;

representing diagonal details, then subband images

And subband images

Correlating, from subband images

Extracting wavelet relative energy distribution characteristics from detail image components

The chain code histogram feature is extracted from the data,

in the feature calculation of (2), the respective bounding boxes are divided into 8 × 8 blocks, and the number of black pixels is calculated and connected

And

forming a feature vector;

(2) WNN: WNN (wavelet neural network) is a multilayer feedforward network, which takes wavelet theory as the basis and takes discrete wavelet function as the activation function of nodes, the WNN classification process is divided into three steps, the first step is network initialization, the second step is training weighting coefficient by using gradient descent algorithm, and the last step is realizing characteristic classification according to the training weighting coefficient, from the perspective of image application, the wavelet network can be further divided into a series of stages, WNN has 3 layers of structures, which are n on an input layer respectively_inA node, n of a hidden layer_hN of one node and output layer_outA node, which selects Mexican hat wavelet as the base, and is defined as expression (7)

The kth input neuron is defined by equation (8):

wherein x_j(j＝1,2,…,n_inIs an input variable, W_j,kDenotes the ithThe weight of the connecting line of the secondary input and the kth hidden node, in order to grasp the level and the position of the wavelet, a multi-scale wavelet function is used as a conversion function of the hidden node, the expansion parameter a of the first hidden node is set to be 1, namely psi_1,b1(x) ψ (x-b1), the expansion parameter a of the second hidden node is set to 2, i.e.

In which the output result of the wavelet is reduced to

Similarly, the dilation parameter a for the jth hidden node is set to j, so the output of the hidden layer of WNN can be given by equation (9):

the output of the kth neuron is defined by equation (10):

the output of WNN is defined as equation (11):

wherein

Representing the weight of the k hidden node and the first output node connecting line, in training, the weight, balance and scaling parameters are adjusted by the formula (12) to minimize the error T function:

where I is 1, …, I is the number of training patterns, K is 1, …, K is the number of targets, D_ikAnd O_ikRespectively represent Node_ikA desired output value and an activation net output value.

Images