CN105023029A - Online handwritten Tibetan syllable recognition method and device - Google Patents

Abstract

The invention provides an online handwritten Tibetan syllable recognition method and device, relates to the technical field of character recognition, and is used to solve the problem that the prior art cannot efficiently recognize Tibetan syllables continuously handwritten by users. The method includes: preprocessing the dot track of the Tibetan syllables continuously handwritten by the user; successively segmenting the preprocessed Tibetan syllables from the horizontal direction and the vertical direction to obtain the substructure block sequence of the two-layer marking result ; Using a segmentation hypothesis verification method based on a semi-Markov conditional random field, the substructure block sequence of the two-layer marking result is subjected to segmentation hypothesis verification to obtain the optimal segmentation path and the identification result of the component string; according to the optimal Based on the optimal segmentation path and the recognition result of the part string, the category of the handwritten Tibetan syllable input by the user is determined. The invention is suitable for recognizing Tibetan syllables continuously handwritten by users.

Description

A method and device for online handwritten Tibetan syllable recognition

technical field

The invention relates to the technical field of character recognition, in particular to an online handwritten Tibetan syllable recognition method and device.

Background technique

Tibetan input methods mainly include handwriting input and keyboard input. Compared with keyboard input, handwriting input is more in line with people's expression habits. It is an effective real-time tool that is easy to use by users, and it is easy to carry and easy to operate. With the advancement and wide application of mobile terminal devices such as smartphones, tablet computers, electronic whiteboards, and iPads, research on online handwritten Tibetan input (pen input) algorithms has received more and more applications and attention. It mainly focuses on Tibetan character recognition, and there is already a handwriting input method that supports Tibetan characters as input units. However, due to the particularity of the Tibetan language itself, people in Tibetan areas hope that handwritten Tibetan input can support continuous writing, and Tibetan syllables are used as handwritten input units, which is more in line with the writing habits of people in Tibetan areas. There are relatively few studies on text syllable recognition, and there are no related literature or patent reports on this technology.

In the process of realizing the present invention, the inventor found that there are at least the following technical problems in the prior art:

The existing online handwritten Tibetan syllable recognition method cannot efficiently recognize the Tibetan syllables input by users' continuous handwriting, and cannot meet the writing habits and needs of Tibetan users.

Contents of the invention

The invention provides an online handwritten Tibetan syllable recognition method and device, capable of efficiently recognizing Tibetan syllables continuously handwritten by users, and meeting the writing habits and needs of Tibetan users.

The online handwritten Tibetan syllable recognition method provided by the present invention comprises:

Preprocess the dot track of Tibetan syllables input by the user's continuous handwriting;

Segment the preprocessed Tibetan syllables successively from the horizontal direction and the vertical direction, and obtain the substructure block sequence of the two-layer marking result;

Using a segmentation hypothesis verification method based on a semi-Markov conditional random field, the segmentation hypothesis verification is performed on the substructure block sequence of the two-layer marking result, and the optimal segmentation path and the recognition result of the component string are obtained;

According to the optimal segmentation path and the recognition result of the part string, the handwritten Tibetan syllable category input by the user is determined.

The online handwritten Tibetan syllable recognition device provided by the present invention comprises:

A preprocessing unit is used to preprocess the dot track of the Tibetan syllables continuously handwritten by the user;

The over-segmentation unit is used to successively over-segment the preprocessed Tibetan syllables from the horizontal direction and the vertical direction to obtain the substructure block sequence of the two-layer marking result;

The segmentation hypothesis verification unit is used to adopt the segmentation hypothesis verification method based on the semi-Markov conditional random field to perform segmentation hypothesis verification on the substructure block sequence of the two-layer marking result, and obtain the optimal segmentation path and identification of the component string result;

The determination unit is configured to determine the handwritten Tibetan syllable category input by the user according to the optimal segmentation path and the recognition result of the component string.

The on-line handwritten Tibetan syllable recognition method and device provided by the present invention first preprocess the dot track of the Tibetan syllables continuously handwritten by the user, and then sequentially over-segment the preprocessed Tibetan syllables from the horizontal direction and the vertical direction , obtain the substructure block sequence of the two-layer marking result, and use the segmentation hypothesis verification method based on the semi-Markov conditional random field to perform segmentation hypothesis verification on the substructure block sequence of the two-layer marking result to obtain the optimal segmentation The recognition result of the path and the component string, and finally, according to the optimal segmentation path and the recognition result of the component string, determine the category of the handwritten Tibetan syllable input by the user. Compared with the prior art, the present invention can efficiently recognize the Tibetan syllables continuously handwritten by users, and satisfy the writing habits and needs of Tibetan users.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

Fig. 1 is the flowchart of the online handwritten Tibetan syllable recognition method provided by Embodiment 1 of the present invention;

Fig. 2 is a schematic diagram of the structural composition of Tibetan syllables provided by Embodiment 2 of the present invention;

FIG. 3 is an example diagram of character segmentation of Tibetan syllables in the horizontal direction provided by Embodiment 2 of the present invention;

Fig. 4 is an example diagram of the wrong character segmentation detection and correct segmentation of Tibetan syllables in the horizontal direction provided by Embodiment 2 of the present invention;

Fig. 5 is an example diagram of component segmentation of Tibetan characters in the vertical direction provided by Embodiment 2 of the present invention;

FIG. 6 is a schematic structural diagram of an on-line handwritten Tibetan syllable recognition device provided by Embodiment 3 of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Embodiment one

The present embodiment provides a kind of online handwritten Tibetan syllable recognition method, as shown in Figure 1, described method comprises:

S11. Preprocessing the dot track of the Tibetan syllables continuously handwritten by the user.

S12. Segment the preprocessed Tibetan syllables successively from the horizontal direction and the vertical direction to obtain a sub-structural block sequence of two-layer marking results.

S13. Using a segmentation hypothesis verification method based on a semi-Markov conditional random field, perform segmentation hypothesis verification on the substructure block sequence of the two-layer marking result, and obtain an optimal segmentation path and a component string identification result.

S14. Determine the handwritten Tibetan syllable category input by the user according to the optimal segmentation path and the recognition result of the component string.

The online handwritten Tibetan syllable recognition method provided by the embodiment of the present invention first preprocesses the dot track of the Tibetan syllables continuously handwritten by the user, and then sequentially over-segments the preprocessed Tibetan syllables from the horizontal direction and the vertical direction , obtain the substructure block sequence of the two-layer marking result, and use the segmentation hypothesis verification method based on the semi-Markov conditional random field to perform segmentation hypothesis verification on the substructure block sequence of the two-layer marking result to obtain the optimal segmentation The recognition result of the path and the component string, and finally, according to the optimal segmentation path and the recognition result of the component string, determine the category of the handwritten Tibetan syllable input by the user. Compared with the prior art, the present invention can efficiently recognize the Tibetan syllables continuously handwritten by users, and satisfy the writing habits and needs of Tibetan users.

Further, the preprocessing of the point trajectories of Tibetan syllables continuously handwritten by the user may include: removing isolated points, equidistant resampling and Gaussian smoothing on the point trajectories of Tibetan syllables continuously handwritten by the user.

Further, the preprocessed Tibetan syllables are successively segmented from the horizontal direction and the vertical direction, and the substructure block sequence for obtaining the two-layer marking result may include: performing horizontal division on the preprocessed Tibetan syllables. Character segmentation and component segmentation in the vertical direction result in a sequence of substructural blocks for the two-level labeling result.

Further, the segmentation hypothesis verification method based on the semi-Markov conditional random field is used to verify the segmentation hypothesis of the sub-structural block sequence of the two-layer marking result, and to obtain the optimal segmentation path and the recognition result of the component string can be Including: integrating the component classifier, geometric context and language context into a unified recognition framework through different weights, performing different segmentation hypothesis verification on the sub-structural block sequence of the two-layer marking results, and obtaining the optimal segmentation Path and component string recognition results.

Optionally, the weights connecting the part classifier, the geometric context and the language context and the respective parameters of the part classifier, the geometric context and the language context may be trained by a criterion based on minimizing a negative log-likelihood loss function get.

Embodiment two

This embodiment provides an online handwritten Tibetan syllable recognition method. This embodiment uses the MRG-OHTC sample database of the Multilingual Processing Research Group of the National Engineering Research Center for Basic Software of the Institute of Software, Chinese Academy of Sciences. The database includes handwritten Tibetan syllable samples of 150 different writers. Each writer completed the writing of 827 high-frequency syllables pre-selected, including 456 two-character syllables, 309 three-character syllables, and 62 four-character syllables. indivual. Select 130 sets of (writer) samples for training, and the remaining 20 sets of samples for testing. In addition, the 150 sets of samples are labeled at the character level and syllable level with semi-supervised calibration tools.

The specific process of the online handwritten Tibetan syllable recognition method provided by the present embodiment is as follows:

(1) Point trajectory preprocessing

The input of an online handwritten Tibetan syllable is expressed as a point sequence of the handwritten trajectory: (x ₁ , y ₁ ), (x ₂ , y ₂ ), ... (x _n , y _n ), where n represents the point in the input syllable trajectory The number of points and the order of the points in the trajectory are arranged according to the time of writing, and the strokes are separated by the end mark point. First remove isolated points, that is, strokes composed of single noise points, to eliminate the impact of isolated noise points on character and component segmentation and component recognition, then equidistantly resample the trajectory of syllables, and finally use Gaussian filtering to smooth the points, to overcome fluctuations in the midpoint of the trajectory. In equidistant resampling, the distance of the points is set to 0.5; in Gaussian smoothing, the variance is set to 1.2.

(2) Over-segmentation

The preprocessed Tibetan syllables are over-segmented into two layers of labeling results, and each layer of labeling results consists of a sequence of substructural blocks. A substructure block is a complete part or a part of a part. A Tibetan syllable is composed of 1 to 4 Tibetan characters combined in the horizontal direction, and each character is composed of one or more components vertically superimposed vertically, as shown in Figure 2. The components refer to the sub-stroke sequences of characters, which are structural primitives that are easily extracted by computer segmentation algorithms and are more stable than stroke structures. Since characters are composed of components, and different characters share the same components, the number of categories of components is much smaller than the number of categories of characters. Combined with the general writing order of Tibetan syllables, the specific steps of over-segmentation are as follows:

a. Character segmentation in the horizontal direction:

First, Tibetan syllables are segmented into character sequences from the horizontal writing direction. Initially assume that each stroke is a sub-structure block, iteratively merge any two sub-structure blocks with a large overlap in the horizontal direction until there is no mergeable one. Assuming that each stroke is a substructure block initially, based on this information, if there is an interval between two substructure blocks (stroke sequences) in the horizontal direction or if the two substructure blocks overlap in the horizontal direction and the overlap is less than 0.1, the two substructure blocks Block segmentation; if two substructure blocks overlap in the horizontal direction and the degree of overlap is greater than 0.1, merge the two substructure blocks. The horizontal overlapping degree refers to the measurement of the overlapping degree of two sub-structural blocks in the horizontal direction.

Usually, when Tibetan syllables are handwritten, there are obvious intervals between the general characters in the horizontal direction, which can be merged through the above iterative method, as shown in Figure 3, which is the correct character segmentation result. However, due to the arbitrariness of writing, the width of the vowel in the horizontal direction is relatively large, and a character with a vowel in a Tibetan syllable often overlaps with other characters in the horizontal direction, as shown in Figure 4. Through the detection The position of the vowel is forced to break, so as to solve the problem of wrong merge.

b. Parts division in the vertical direction:

Based on the result of character segmentation in the horizontal direction, for each character, component segmentation is performed in the vertical direction. Use a calculation method similar to the overlapping degree in the horizontal direction for merging. Compared with the spacing or overlapping between characters, there are generally small blank spaces in the vertical direction between components or large overlaps between components. Experience in setting overlap merging The value is 0.2, and the component segmentation results in the vertical direction are shown in Figure 5.

Components in Tibetan characters may connect with each other, and the corner detection method is used to break the connected strokes, so as to ensure the correct segmentation of components.

(3) Segmentation hypothesis verification based on semi-Markov conditional random field

The Tibetan syllable recognition is regarded as the recognition of two-layer component strings, that is, the recognition of component strings divided in the horizontal direction and vertical direction. The key problem is how to obtain Correct part string segmentation point and part recognition result. The present invention adopts the segmentation hypothesis verification method based on the semi-Markov conditional random field, integrates the component classifier, geometric context and language context into a unified recognition framework, verifies different segmentation assumptions, and obtains the optimal segmentation path And the recognition result of the component string. The above models are explained as follows:

a. Part classifier

The component classifier adopts a multi-feature and multi-classification fusion model based on a deep neural network. It uses a deep neural network to represent Tibetan characters from different angles, and then classifies them with different statistical classifiers to realize multi-feature and multi-classification fusion of Tibetan characters. Part identification method. For online features, the online handwritten Tibetan components are composed of stroke sequences. First, the original features are extracted through the coordinate normalization method (NCFE), and then the deep belief network (DBN) is used to obtain higher-level features through multi-layer nonlinear transformation. The online feature-based classification results are obtained by using Nearest Prototype Classifier (NPC) classification. For offline features, the Tibetan components composed of stroke sequences are first converted into binary images, and the bottommost pixels are used as the input of the feature representation, and the deep convolutional neural network (DCNN) is used to extract features, and the modified quadratic discriminant function is used. The classifier (MQDF) classifies and obtains the classification results based on offline features, and finally combines the online and offline classification results to obtain the Tibetan component recognition results.

The embodiment of the present invention uses Tibetan components as the basic recognition unit. Compared with character categories, the total number of component categories is about 1/5 of the character category, which makes the storage capacity of the dictionary of the component classifier small, which can meet the storage requirements of mobile devices. requirements; moreover, the similar parts in smaller part categories are also greatly reduced, which helps to improve the final syllable recognition accuracy.

b. Geometric context

The geometric context includes the geometric context between characters within a syllable and the geometric context between parts within a character. The geometric context between characters within a syllable refers to information such as the height, width, and position of the candidate character pattern relative to the entire Tibetan syllable, and the distance and relative position between adjacent candidate characters. Aiming at the geometric context between characters in a syllable, a unary geometric feature is established for each type of character, and a binary geometric feature is established between every two consecutive characters in a syllable, which are modeled with different quadratic discriminant functions Unary and binary features. This embodiment uses a total of 6 unary geometric features, including the width of the candidate character, the height, the length of the diagonal of the circumscribed rectangle, the distance between the center, upper boundary and lower boundary of the circumscribed rectangle and the horizontal center line of the character string, these 6 features It needs to be normalized with the average Tibetan character height. A total of 4 binary geometric features are used, including the upper boundary, lower boundary, upper boundary and lower boundary of adjacent Tibetan characters, and the difference between the horizontal centerline. The above features are normalized by the average Tibetan character height change.

The geometric context between parts within a character refers to information such as the height, width, and relative position of the candidate part pattern relative to the entire Tibetan character. For each type of character, a unary geometric feature is established for each component in the character, and a binary geometric feature is established between every two consecutive components (arranged according to the upper boundary) in the character, using different Gaussian probability densities Functions to model unary and binary geometric features. The unary and binary geometric feature extraction methods at the component level are similar to those at the character level, and the extracted features are all normalized by the average Tibetan component height.

c. Language context

There is a certain relationship between the categories of characters in Tibetan syllables and the categories of components in Tibetan characters, that is, the language context. The language context builds the language model from the character layer and the component layer respectively. For the language model of the character layer and the component layer, a binary grammar is used to describe it, and the feature function of the language context is defined as the logarithm of the probability of the binary grammar, and it is a binary feature function related to the character or component category.

The modeling of the above three models needs to calibrate the Tibetan syllable samples at the character layer and component layer first, and obtain character and component samples from the Tibetan syllable samples and determine the categories of characters and components through the calibration results. The categories of components are 562 and 120 respectively. The calibration method adopts a method based on semi-supervised learning, which greatly reduces the workload of manual intervention.

After constructing the part classifier, geometric context and language context models, the segmentation hypothesis verification method based on semi-Markov conditional random fields is used to further verify the segmentation hypothesis on the sub-structural block sequence of the two-layer labeling result obtained in step (2). verify.

Based on the substructure block sequence of the two-layer labeling results, the candidate segmentation-recognition grid is generated through the classification of candidate component patterns, and a semi-Markov conditional random field model is constructed in the grid. Assuming that the Tibetan syllable over-segmentation result (two-layer labeling result) is X, and the category of the candidate path in the grid is Y (category sequence), the corresponding segmentation is S: Y (candidate component sequence), according to the semi-Markov Conditional random field model, the conditional probability P(S,Y|X) of the candidate path (S,Y) is expressed as:

$\begin{array}{c}\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; Z</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}\underset{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; \&Pi;</span>}}{\mathrm{<span\; class="notranslate">\; \&Psi;</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; Z</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; \{</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span>\end{array}$

where c represents the largest clique in the random field, Y _c represents the category of c, Ψ _c (X, Yc) is the potential function defined on c, and the normalization factor Z(X) is the potential of all candidate paths in the grid The sum of functions, E(X,S,Y) represents the energy function:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; )</span>$

f _k (X _c , Y _c ) is the kth feature function defined on c, which is used to describe the component classification model, the unary and binary feature functions between characters in a syllable, and the unary and Binary feature functions, character-based language models, and component-based language models. The segmentation hypothesis verification method based on the semi-Markov conditional random field integrates each sub-model into a unified identification framework through the weight λ _k , and the weight λ _k and the parameters of each sub-model are based on minimizing the negative logarithm likelihood The criterion of the loss function is trained.

(4) Identification output

Based on the part string segmentation hypothesis verification result obtained in step (3), check the syllable string representation dictionary and the character part string representation dictionary to obtain the character category included in the syllable, thereby determining the input Tibetan syllable category.

Table 1 and Table 2 respectively list the effect of using the segmentation hypothesis verification method based on semi-Markov conditional random fields to fuse geometric context and language context. It can be seen from the table that both geometric context and language context have improved Tibetan syllable recognition In the linguistic context, component-based bi-grams contribute more to Tibetan syllable recognition accuracy than character-based bi-grams, mainly because the entire syllable recognition framework is built at the component level.

Table 1 Effect of geometric context on syllable recognition accuracy

Part Classifier geometry context Recognition accuracy (%) √ 73.51 √ √ 74.87

Table 2 The influence of language context on the accuracy of syllable recognition

Component-based bi-grams character-based bi-gram Recognition accuracy (%) √ 79.65 √ 77.72

√ √ 81.23

Embodiment Three

An embodiment of the present invention provides an online handwritten Tibetan syllable recognition device, as shown in Figure 6, the device includes:

The preprocessing unit 11 is used to preprocess the dot track of the Tibetan syllables continuously handwritten by the user;

The over-segmentation unit 12 is used to successively over-segment the preprocessed Tibetan syllables from the horizontal direction and the vertical direction to obtain the substructure block sequence of the two-layer marking result;

The segmentation hypothesis verification unit 13 is used to adopt the segmentation hypothesis verification method based on the semi-Markov conditional random field to perform segmentation hypothesis verification on the sub-structural block sequence of the two-layer marking result, and obtain the optimal segmentation path and component string recognition result;

The determining unit 14 is configured to determine the handwritten Tibetan syllable category input by the user according to the optimal segmentation path and the recognition result of the component string.

The online handwritten Tibetan syllable recognition device provided by the embodiment of the present invention first preprocesses the dot track of the Tibetan syllables continuously handwritten by the user, and then sequentially over-segments the preprocessed Tibetan syllables from the horizontal direction and the vertical direction , obtain the substructure block sequence of the two-layer marking result, and use the segmentation hypothesis verification method based on the semi-Markov conditional random field to perform segmentation hypothesis verification on the substructure block sequence of the two-layer marking result to obtain the optimal segmentation The recognition result of the path and the component string, and finally, according to the optimal segmentation path and the recognition result of the component string, determine the category of the handwritten Tibetan syllable input by the user. Compared with the prior art, the present invention can efficiently recognize the Tibetan syllables continuously handwritten by users, and satisfy the writing habits and needs of Tibetan users.

Further, the preprocessing unit 11 is configured to perform isolated point removal, equidistant resampling, and Gaussian smoothing on point trajectories of Tibetan syllables continuously handwritten by the user.

Further, the over-segmentation unit 12 is configured to perform horizontal character segmentation and vertical component segmentation on the preprocessed Tibetan syllables successively to obtain a sub-structural block sequence of two-layer marking results.

Further, the segmentation hypothesis verification unit 13 is used to integrate the component classifier, geometric context and language context into a unified recognition framework through different weights, and perform Different segmentation assumptions are verified to obtain the optimal segmentation path and recognition results of component strings.

Optionally, the weights connecting the part classifier, the geometric context and the language context and the respective parameters of the part classifier, the geometric context and the language context are obtained by training based on the criterion of minimizing the negative log-likelihood loss function .

The online handwritten Tibetan syllable recognition method and device provided by the embodiments of the present invention are suitable for recognizing Tibetan syllables continuously handwritten by users, but are not limited thereto.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented through computer programs to instruct related hardware, and the programs can be stored in computer-readable storage media. During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM), etc.

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present invention. All should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims (10)

1. the recognition methods of on-line handwritten Tibetan language syllable, is characterized in that, comprising:

Pre-service is carried out to the locus of points of the Tibetan language syllable of the continuous handwriting input of user;

Successively from horizontal direction and vertical direction, over-segmentation is carried out to pretreated Tibetan language syllable, obtains the minor structure block sequence of two-layer mark result;

Adopt the segmentation hypothesis verification method based on half Markov condition random field, segmentation hypothesis verification is carried out to the minor structure block sequence of described two-layer mark result, obtain optimum split path and the recognition result of parts string;

According to the split path of described optimum and the recognition result of parts string, determine the hand-written Tibetan language syllable classification that described user inputs.

2. method according to claim 1, it is characterized in that, the locus of points of the described Tibetan language syllable to the continuous handwriting input of user carries out pre-service and comprises: carry out removal isolated point, equidistant resampling and Gaussian smoothing to the locus of points of the Tibetan language syllable of the continuous handwriting input of user.

3. method according to claim 1, it is characterized in that, described successively from horizontal direction and vertical direction, over-segmentation is carried out to pretreated Tibetan language syllable, the minor structure block sequence obtaining two-layer mark result comprises: successively carry out the Character segmentation of horizontal direction and the parts segmentation of vertical direction to pretreated Tibetan language syllable, obtain the minor structure block sequence of two-layer mark result.

4. method according to claim 1, it is characterized in that, described employing is based on the segmentation hypothesis verification method of half Markov condition random field, segmentation hypothesis verification is carried out to the minor structure block sequence of described two-layer mark result, the recognition result obtaining optimum split path and parts string comprises: by different weights by part classification device, geometrically hereafter and under Linguistic context is integrated into a unified identification framework, different segmentation hypothesis verifications is carried out to the minor structure block sequence of described two-layer mark result, obtain optimum split path and the recognition result of parts string.

5. method according to claim 4, it is characterized in that, connect described part classification device, geometrically hereafter and the weights of Linguistic context and described part classification device, geometrically hereafter and Linguistic context parameter separately by obtaining based on the criterion training minimizing negative log likelihood loss function.

6. an on-line handwritten Tibetan language syllable recognition device, is characterized in that, comprising:

Pretreatment unit, the locus of points for the Tibetan language syllable to the continuous handwriting input of user carries out pre-service;

Over-segmentation unit, for successively carrying out over-segmentation from horizontal direction and vertical direction to pretreated Tibetan language syllable, obtains the minor structure block sequence of two-layer mark result;

Segmentation hypothesis verification unit, for adopting the segmentation hypothesis verification method based on half Markov condition random field, carries out segmentation hypothesis verification to the minor structure block sequence of described two-layer mark result, obtains optimum split path and the recognition result of parts string;

Determining unit, for according to the split path of described optimum and the recognition result of parts string, determines the hand-written Tibetan language syllable classification that described user inputs.

7. device according to claim 6, is characterized in that, described pretreatment unit, and the locus of points for the Tibetan language syllable to the continuous handwriting input of user carries out removal isolated point, equidistant resampling and Gaussian smoothing.

8. device according to claim 6, is characterized in that, described over-segmentation unit, for successively carrying out the Character segmentation of horizontal direction and the parts segmentation of vertical direction to pretreated Tibetan language syllable, obtains the minor structure block sequence of two-layer mark result.

9. device according to claim 6, it is characterized in that, described segmentation hypothesis verification unit, for by different weights by part classification device, geometrically hereafter and under Linguistic context is integrated into a unified identification framework, different segmentation hypothesis verifications is carried out to the minor structure block sequence of described two-layer mark result, obtains optimum split path and the recognition result of parts string.

10. device according to claim 9, it is characterized in that, connect described part classification device, geometrically hereafter and the weights of Linguistic context and described part classification device, geometrically hereafter and Linguistic context parameter separately by obtaining based on the criterion training minimizing negative log likelihood loss function.