JP2019528520A - Classification network training apparatus, character recognition apparatus and method for character recognition - Google Patents

Abstract

A classification network training apparatus, character recognition apparatus, and method for character recognition are provided. The apparatus and method build a sample pair for unlabeled samples to train a symmetric network, initialize a classification network using parameters of the trained symmetric network, and initialize a classification network using labeled samples Training can improve the accuracy of classification network recognition and effectively save the cost of labeling. [Selection] Figure 1

Description

  The present invention relates to the field of information technology, and more particularly to a classification network training apparatus, character recognition apparatus, and method for character recognition.

  Due to the necessity of document storage and information development, the demand for digitization of document materials is increasing. For this reason, recognition of characters in document images is becoming increasingly important. Recognition of special characters such as kanji in ancient literature is very important in digitizing classical literature, organizing classical books and preserving culture. However, compared to the recognition of modern Chinese characters, the recognition of ancient Chinese characters is a very difficult problem. First, the number of Chinese characters in ancient literature is far greater than the number of modern Chinese characters. Second, the structure of Chinese characters in ancient literature is much more complex than modern simplified Chinese characters. Thirdly, Chinese characters in ancient literature have multiple aspects, that is, large numbers of Chinese characters have different writing styles at different historical times. Fourth, due to the use of different writing instruments (eg brush) or woodcut printing, ancient Chinese characters have multiple styles. Finally, the degradation of the images of classics taken or scanned is more pronounced compared to modern Chinese characters.

  In recent years, in the field of optical character recognition (OCR), deep learning methods (for example, convolutional neural networks) are significantly superior to conventional methods. A supervised learning method based on a Convolutional Neural Network (CNN), which is currently mainly used, usually requires millions of samples of training data. Since there are not enough labeled samples for kanji recognition in ancient literature, a large number of unlabeled samples are obtained by scanning or photographing, and divided using an automatic character dividing method, and manually labeled by humans. It is necessary to obtain labeled samples for training the convolutional neural network by performing the pasting.

  The above description of the technical background is an explanation for making the technical solution of the present invention clear and complete, and is described for the purpose of understanding those skilled in the art. These technical solutions are merely described as background art portions of the present invention, and are not well known by those skilled in the art.

  When training a convolutional neural network using the above conventional method, a large amount of manual labeling is required, a long time is required, and a lot of labor and cost are required.

  An embodiment of the present invention builds a sample pair for unlabeled samples to train a symmetric network, initializes a classification network using parameters of the trained symmetric network, and initializes classification using labeled samples. Classification network training apparatus for character recognition, character recognition apparatus, which can improve the accuracy rate of classification network recognition by training the network, and can effectively save the cost of labeling, and Provide a method.

  According to a first aspect of an embodiment of the present invention, there is provided a classification network training apparatus for character recognition, an extraction means for extracting features of each unlabeled sample including characters, and features of each extracted unlabeled sample. Using the construction means for constructing the sample pair, the first training means for training the symmetric network based on the constructed sample pair, and the parameters for the trained symmetric network for character recognition An apparatus is provided comprising: initialization means for initializing a classification network; and second training means for training the initialized classification network using labeled samples containing characters.

  According to a second aspect of an embodiment of the present invention, there is provided a character recognition device including a classification network for character recognition trained by the apparatus according to the first aspect of the embodiment of the present invention.

  According to a third aspect of the embodiment of the present invention, there is provided a classification network training method for character recognition, comprising: extracting a feature of each unlabeled sample including characters; and extracting a feature of each extracted unlabeled sample. Based on, constructing a sample pair, training a symmetric network based on the constructed sample pair, and initializing a classification network for character recognition using parameters of the trained symmetric network And training the initialized classification network with labeled samples containing characters.

  An advantageous effect of the present invention is that it builds a sample pair for unlabeled samples to train a symmetric network, initializes a classification network using the parameters of the trained symmetric network, and initializes using labeled samples. By training the classification network, the accuracy of classification network recognition can be improved and the cost of labeling can be effectively saved.

  Certain embodiments of the present invention are disclosed in detail and illustrate the manner in which the principles of the present invention can be employed, as illustrated in the following description and drawings. The embodiment of the present invention is not limited in scope. The embodiments of the present invention include various changes, modifications, and equivalents within the scope and spirit of the appended claims.

  Features described and / or shown in one embodiment may be used in one or many other embodiments in the same or similar manner, and may be combined with features in other embodiments, Features in other embodiments may be substituted.

  As used herein, the term “inclusive / include” means the presence of a feature, element, step or component, and the presence or absence of one or more other features, elements, steps or components. It does not exclude the addition.

The drawings included herein are for the purpose of understanding the embodiments of the present invention, constitute part of the present specification, illustrate the embodiments of the present invention, and are combined with the description of the words. The principle of the present invention will be described. Note that the drawings described here are merely illustrative of embodiments of the present invention, and those skilled in the art can easily obtain other drawings based on these drawings.
It is a figure which shows the training apparatus of the classification network for the character recognition of Example 1 of this invention. It is a figure which shows the construction | assembly part 102 of Example 1 of this invention. It is a figure which shows the 1st determination part 201 of Example 1 of this invention. It is another figure which shows the 1st determination part 201 of Example 1 of this invention. It is a figure which shows the symmetrical network of Example 1 of this invention. It is a figure which shows the character recognition apparatus of Example 2 of this invention. It is a figure which shows the electronic device of Example 3 of this invention. It is a block diagram which shows the system configuration | structure of the electronic device of Example 3 of this invention. It is a figure which shows the training method of the classification network for the character recognition of Example 4 of this invention.

  These and other features of the present invention can be understood from the drawings and the following description. The specification and drawings disclose certain embodiments of the invention, i.e., some embodiments that follow the principles of the invention. Note that the present invention is not limited to the described embodiments, and the present invention includes all modifications, variations, and equivalents within the scope of the claims.

<Example 1>
FIG. 1 is a diagram illustrating a classification network training apparatus for character recognition according to a first embodiment of the present invention. As shown in FIG. 1, the training apparatus 100 includes an extraction unit 101, a construction unit 102, a first training unit 103, an initialization unit 104, and a second training unit 105.

  The extraction unit 101 extracts features of each unlabeled sample including characters.

  The construction unit 102 constructs a sample pair based on the extracted features of each unlabeled sample.

  The first training unit 103 trains a symmetric network based on the constructed sample pair.

  The initialization unit 104 initializes a classification network for character recognition using parameters of the trained symmetric network.

  The second training unit 105 trains the initialized classification network using labeled samples including characters.

  According to this example, a sample pair is constructed for unlabeled samples to train a symmetric network, a classification network is initialized using parameters of the trained symmetric network, and a classification is initialized using labeled samples. Training the network can improve the accuracy of classification network recognition and can effectively save the cost of labeling.

  In this example, unlabeled samples and labeled samples containing characters may be obtained using conventional methods, each sample dividing an image containing multiple characters using a conventional character segmentation method. May be obtained.

  In the present embodiment, the character may be in any form, for example, a modern character, a Chinese character from ancient literature, or a character in another language. The embodiments of the present invention are not limited to the types of characters, and the classification network for character recognition trained according to the embodiments of the present invention may be used to recognize any form of characters. It is not limited to the recognition of special characters such as literature characters.

  In the present embodiment, the number of unlabeled samples including characters and the number of labeled samples may be set according to the actual situation, and the present embodiment is not limited to this.

  In the present embodiment, the extraction unit 101 may extract character features from unlabeled samples that directly include characters, or input unlabeled samples to a trained network using labeled samples that include characters, The output result may be the extracted feature.

  For example, the extraction unit 101 may directly extract character features such as strokes and textures as extraction features using a conventional method.

  For example, the extraction unit 101 may train a network using labeled samples, input unlabeled samples to the trained network, and use the output result as an extraction feature. For example, the network may be a convolutional neural network (CNN). For example, the convolutional neural network may be a classifier. In this case, the extracted feature is a classification result of the input unlabeled sample.

  In this embodiment, after the extraction unit 101 extracts the features of each unlabeled sample, the construction unit 102 constructs a sample pair based on the extracted features of each unlabeled sample. Hereinafter, the configuration of the construction unit 102 and the construction method of the sample pair according to the present embodiment will be described as an example.

  FIG. 2 is a diagram illustrating the construction unit 102 according to the first embodiment of this invention. As shown in FIG. 2, the construction unit 102 includes a first determination unit 201.

  The first determination unit 201 determines a first similar sample pair and a first dissimilar sample pair based on the extracted feature of each unlabeled sample.

  In the present embodiment, the number of first similar sample pairs and first dissimilar sample pairs determined by the first determination unit 201 may be set according to an actual request.

  Below, the structure of the 1st determination part 201 and the determination method of a 1st similar sample pair and a 1st dissimilar sample pair are demonstrated exemplarily.

  FIG. 3 is a diagram illustrating the first determination unit 201 according to the first embodiment of this invention. As illustrated in FIG. 3, the first determination unit 201 includes a second calculation unit 301, a fourth determination unit 302, and a fifth determination unit 303.

  The second calculator 301 calculates the distance between the features of any two unlabeled samples among the extracted unlabeled samples.

  The fourth determination unit 302 determines any two unlabeled samples whose distance between features is smaller than a predetermined threshold as the first similar sample pair.

  The fifth determination unit 303 determines any two unlabeled samples whose distance between features is equal to or greater than the predetermined threshold as the first dissimilar sample pair.

  In the present embodiment, the first calculation unit 201 may calculate a distance between features of any two label samples using a conventional method, and the predetermined threshold may be set according to an actual request. .

  In the present embodiment, when the feature of each unlabeled sample extracted by the extraction unit 101 is a classification result of the unlabeled sample, the first determination unit 201 uses the other method to determine the first similar sample pair and the first similar sample pair. One dissimilar sample pair may be determined.

  FIG. 4 is another diagram illustrating the first determination unit 201 according to the first embodiment of this invention. As illustrated in FIG. 4, the first determination unit 201 includes a sixth determination unit 401 and a seventh determination unit 402.

  The sixth determination unit 401 determines any two unlabeled samples having the same classification result as the first similar sample pair.

  The seventh determination unit 402 determines any two unlabeled samples with different classification results as the first dissimilar sample pair.

  Although the above has described the determination method of the 1st similar sample pair and the 1st dissimilar sample pair by the 1st determination part 201 exemplarily, the Example of this invention is not limited to this.

  In the present embodiment, the construction unit 102 illustrated in FIG. 2 may further include a second determination unit 202, a first calculation unit 203, and a third determination unit 204.

  The second determination unit 202 determines any two labeled samples among the labeled samples as the second similar sample pair or the second dissimilar sample pair based on the label of the labeled sample.

  The first calculation unit 203 calculates the ratio of the number of the determined second similar sample pair and the second dissimilar sample pair.

  The third determination unit 204 sets the first similar sample pair and the first dissimilar sample pair so that the number ratio of the first similar sample pair and the first dissimilar sample pair is equal to the number ratio of the second similar sample pair and the second dissimilar sample pair. A ratio of the number of similar sample pairs to the first dissimilar sample pair is determined.

  Thus, based on the ratio of the number of the second similar sample pair and the second dissimilar sample pair of the labeled sample, the ratio of the number of the first similar sample pair and the first dissimilar sample pair of the unlabeled sample is calculated. The determination can further improve the classification accuracy of the trained classification network.

  In the present embodiment, the second determination unit 202 determines any two labeled samples among the labeled samples as the second similar sample pair or the second dissimilar sample pair based on the label of the labeled sample. For example, any two labeled samples having the same label are determined as the second similar sample pair, and any two labeled samples having different labels are determined as the second dissimilar sample pair.

  In the present embodiment, the first calculation unit 203 calculates the ratio of the number of the second similar sample pairs and the second dissimilar sample pairs determined by the first determination unit 201, and the third determination unit 204 The first similar sample pair and the first dissimilarity such that the ratio of the number of one similar sample pair and the first dissimilar sample pair is equal to the ratio of the number of the second similar sample pair and the second dissimilar sample pair. Determine the ratio of the number to the sample pair.

  For example, when a sufficiently large number of first similar sample pairs and first dissimilar sample pairs have already been determined by the first determining unit 201, the third determining unit 204 determines that the first similar sample pair and the first dissimilar sample pair The first similar sample pair and the first dissimilar sample pair are selected such that the ratio of the number of the first similar sample pair and the second dissimilar sample pair is equal to the number ratio of the second similar sample pair. Alternatively, after the third determining unit 204 determines the ratio of the number of the first similar sample pair and the first dissimilar sample pair, the first determining unit 201 determines whether the first similar sample pair and the first similar sample pair and A first dissimilar sample pair is determined.

  In the present embodiment, the first training unit 103 trains a symmetric network based on the sample pairs constructed by the construction unit 102. For example, the symmetric network is a Siamese network having two convolutional neural networks (CNN) arranged symmetrically. FIG. 5 is a diagram showing a symmetric network according to the first embodiment of the present invention. As shown in FIG. 5, the two convolutional neural networks CNN1 and CNN2 in the Siamese network 500 may be arranged symmetrically, and CNN1 and CNN2 may use the conventional CNN configuration, and the configuration and parameters of both are completely the same. is there.

  In this example, a symmetric network may be trained based on the constructed sample pairs using conventional methods. For example, the constructed sample pair is input to the Siase network for each pair, one sample in the sample pair is input to CNN1, the other sample is input to CNN2, and CNN1 and CNN2 are input to the output side of the Siasese network. The loss of the Siamese network is calculated based on the contrast loss. Then, the loss of the Siasese network is fed back through each layer in CNN1 and CNN2, and the parameters of each layer in CNN1 and CNN2 are corrected. The training is terminated after repeating the above steps until the loss of the Siamese network satisfies a predetermined convergence condition.

  In this embodiment, the initialization unit 104 initializes a classification network for character recognition using parameters of a trained symmetric network. For example, the classification network for character recognition is a convolutional neural network (CNN), and the convolutional neural network may use a conventional configuration. The parameters of the arbitrary one convolutional neural network in the trained Siamese network are used to initialize the convolutional neural network that becomes the classification network.

  In the present embodiment, the initialization parameters may include parameters of each convolution layer of the convolutional neural network, and may further include parameters of all connection layers.

  In a present Example, the 2nd training part 105 trains the initialized classification network using the labeled sample containing a character, and acquires the classification network for the trained character recognition. In the present embodiment, the second training unit 105 may train the initialized classification network using a conventional method.

  For example, a labeled sample including characters is input to the initialized convolutional neural network for each sample, and the loss of the network is calculated on the output side. Then, the loss of the network is fed back through each layer in the convolutional neural network, and the parameters of each layer in the convolutional neural network are corrected. The training is terminated after repeating the above steps until the network loss of the convolutional neural network satisfies a predetermined convergence condition.

  In the present embodiment, the training apparatus 100 illustrated in FIG. 1 may further include a determination unit 106.

  The determination unit 106 determines whether or not the trained classification network satisfies a predetermined condition. When the trained classification network does not satisfy the predetermined condition, the trained classification network is used to include each character including the character. The feature of the unlabeled sample is extracted, and if the trained classification network satisfies the predetermined condition, the trained classification network is output.

  In this embodiment, the determination unit 106 is an optional component, and is indicated by a broken-line frame in FIG.

  In this embodiment, the predetermined condition may be set according to an actual request. For example, the predetermined condition is that the number of iterations reaches a predetermined number, or that the classification accuracy of the trained classification network converges, that is, the classification accuracy of the current trained classification network and the previous training The difference from the classification accuracy rate of the classified network is smaller than a predetermined threshold.

  Thus, if the current trained classification network does not meet the predetermined condition, the trained classification network is used to extract the features of each unlabeled sample containing characters. That is, in this case, until the trained classification network satisfies a predetermined condition, the extraction unit 101 extracts features of each unlabeled sample including characters based on the current trained classification network, and reconstructs the sample pair Train the symmetric network, initialize the classification network, and train the classification network. Such an iterative process can further improve the recognition accuracy rate of the trained classification network.

  According to this example, a sample pair is constructed for unlabeled samples to train a symmetric network, a classification network is initialized using parameters of the trained symmetric network, and a classification is initialized using labeled samples. Training the network can improve the accuracy of classification network recognition and can effectively save the cost of labeling.

<Example 2>
The embodiment of the present invention further provides a character recognition device including a classification network for character recognition trained by the training device described in the first embodiment.

  FIG. 6 is a diagram showing a character recognition apparatus according to the second embodiment of the present invention. As shown in FIG. 6, the character recognition device 600 includes a classification network 601 for character recognition.

  The classification network 601 is trained by a training device. Here, the configuration and function of the training device are the same as those described in the first embodiment, and the description thereof is omitted here.

  For example, the character recognition device includes a storage unit, and the storage unit stores a classification network for character recognition trained by the training device described in the first embodiment.

  According to this example, a sample pair is constructed for unlabeled samples to train a symmetric network, a classification network is initialized using parameters of the trained symmetric network, and a classification is initialized using labeled samples. Training the network can improve the accuracy of classification network recognition and can effectively save the cost of labeling.

<Example 3>
The embodiment of the present invention further provides an electronic device, and FIG. 7 is a diagram showing the electronic device of Embodiment 3 of the present invention. As shown in FIG. 7, the electronic device 700 includes a training device 701 or a character recognition device 702. Here, the configuration and function of the training device 701 are the same as those described in the first embodiment, and the configuration and function of the character recognition device 702 are the same as those described in the second embodiment. Is omitted.

  FIG. 8 is a block diagram showing a system configuration of the electronic apparatus according to the third embodiment of the present invention. As shown in FIG. 8, the electronic apparatus 800 may include a central processing unit (central control unit) 801 and a storage device 802, and the storage device 802 is connected to the central processing unit 801. The figures are merely exemplary, and other types of configurations may be used or supplemented or substituted to implement telecommunications functions or other functions.

  As illustrated in FIG. 8, the electronic device 800 may further include an input unit 803, a display 804, and a power source 805.

  In one aspect, the functionality of the training device of Example 1 may be integrated into the central processing unit 801. Here, the central processing unit 801 extracts features of each unlabeled sample including characters, constructs a sample pair based on the extracted features of each unlabeled sample, and based on the constructed sample pair. Train a symmetric network, initialize a classification network for character recognition using parameters of the trained symmetric network, and train the initialized classification network using labeled samples containing characters May be configured.

  Here, the central processing unit 801 determines whether or not the trained classification network satisfies a predetermined condition. If the trained classification network does not satisfy the predetermined condition, the central processing unit 801 determines the trained classification network. May be further configured to extract the features of each unlabeled sample that includes the character and output the trained classification network if the trained classification network satisfies the predetermined condition.

  Here, extracting the feature of each unlabeled sample containing the character was trained using extracting the character feature from the unlabeled sample containing the character directly, or using the labeled sample containing the character. Inputting the unlabeled sample into the network and characterizing the output result.

  Here, the step of constructing a sample pair based on the extracted feature of each unlabeled sample is a first similar sample pair and a first dissimilar sample pair based on the extracted feature of each unlabeled sample. Determining. Determining any two labeled samples of each labeled sample as a second similar sample pair or a second dissimilar sample pair based on the label of the labeled sample; and the determined second similar sample Calculating a ratio of the number of pairs to the second dissimilar sample pair; and a ratio of the number of the first similar sample pair to the first dissimilar sample pair to determine the second similar sample pair and the second dissimilar sample pair. Determining a ratio of the number of the first similar sample pair and the first dissimilar sample pair to be equal to a ratio of the number of dissimilar sample pairs.

  Here, the step of determining the first similar sample pair and the first dissimilar sample pair based on the characteristics of each extracted unlabeled sample includes any two unlabeled samples of the extracted unlabeled samples. Calculating a distance between features of the sample; determining any two unlabeled samples whose feature distance is less than a predetermined threshold as the first similar sample pair; and determining a distance between features as the predetermined threshold. Determining any two unlabeled samples as the first dissimilar sample pair as described above.

  Here, the feature of each extracted unlabeled sample is a classification result of the unlabeled sample, and based on the feature of each extracted unlabeled sample, the first similar sample pair and the first dissimilar sample pair Determining any two unlabeled samples having the same classification result as the first similar sample pair, and determining any two unlabeled samples having different classification results to the first dissimilar sample. Determining as a pair.

  Here, the symmetric network is a Siamese network having two convolutional neural networks arranged symmetrically, and the classification network for character recognition is a convolutional neural network.

  Here, the step of initializing a classification network for character recognition using the parameters of the trained symmetric network uses the parameters of any one convolutional neural network in the trained sham network to Initializing a convolutional neural network to be a network.

  In another aspect, the training device described in the first embodiment may be configured with a central processing unit 801, for example, the training device is a chip connected to the central processing unit 801, and the control of the central processing unit 801. The function of the training apparatus may be realized by

  The electronic apparatus 800 in the present embodiment may not include all the components shown in FIG.

  As shown in FIG. 8, the central processing unit 801 is also referred to as a controller or an operation control unit, and may include a microprocessor or other processing unit and / or a logic unit. The operation of each part of the device 800 is controlled.

  Storage device 802 may be, for example, one or more of a buffer, flash memory, hard disk, removable medium, volatile memory, non-volatile memory, or other suitable device. Further, the central processing unit 801 may execute a program stored in the storage device 802 to realize information storage or processing. Since other components are similar to those of the prior art, description thereof is omitted here. Each unit of the electronic device 800 may be realized by specific hardware, firmware, software, or a combination thereof without departing from the scope of the present invention.

  According to this example, a sample pair is constructed for unlabeled samples to train a symmetric network, a classification network is initialized using parameters of the trained symmetric network, and a classification is initialized using labeled samples. Training the network can improve the accuracy of classification network recognition and can effectively save the cost of labeling.

<Example 4>
The embodiment of the present invention further provides a classification network training method for character recognition, which corresponds to the classification network training apparatus for character recognition of the first embodiment. FIG. 9 is a diagram illustrating a classification network training method for character recognition according to a fourth embodiment of the present invention. As shown in FIG. 9, the method includes the following steps.

  Step 901: Extract features of each unlabeled sample containing characters.

  Step 902: Build a sample pair based on the extracted features of each unlabeled sample.

  Step 903: Train a symmetric network based on the constructed sample pairs.

  Step 904: Initialize a classification network for character recognition using the parameters of the trained symmetric network.

  Step 905: Train the initialized classification network with labeled samples containing letters.

  Step 906: It is determined whether the trained classification network satisfies a predetermined condition. If the determination result is “NO”, the process proceeds to step 901, and features of each unlabeled sample including characters are extracted using the trained classification network. If the determination result is “YES”, the process proceeds to step 907.

  Step 907: Output the trained classification network.

  In this embodiment, a feature extraction method, a sample pair construction method, a symmetric network training method, a classification network initialization method, a classification network training method, and whether or not the trained classification network satisfies a predetermined condition. The determination method is the same as that described in the first embodiment, and the description thereof is omitted here.

  According to this example, a sample pair is constructed for unlabeled samples to train a symmetric network, a classification network is initialized using parameters of the trained symmetric network, and a classification is initialized using labeled samples. Training the network can improve the accuracy of classification network recognition and can effectively save the cost of labeling.

  According to the embodiment of the present invention, when the program is executed in the classification network training apparatus or electronic device for character recognition, the above-described fourth embodiment is applied to the classification network training apparatus or electronic device for character recognition. The computer-readable program which performs the training method as described in 1 is further provided.

  According to an embodiment of the present invention, there is provided a storage medium for storing a computer-readable program for causing a computer to execute the training method described in Embodiment 4 in a classification network training apparatus or electronic device for character recognition. Provide further.

  The training method executed in the classification network training device for character recognition described with reference to the embodiments of the present invention may be implemented in hardware, a software module executed by a processor, or a combination of both. . For example, one or more of the functional block diagrams shown in FIG. 1 or one or more combinations of the functional block diagrams may correspond to each software module of the computer program flow, and may correspond to each hardware module. May correspond. These software modules may correspond to the respective steps shown in FIG. These hardware modules may be realized by hardwareizing these software modules using, for example, a field programmable gate array (FPGA).

  A software module may be located in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, mobile hard disk, CD-ROM, or any other form of storage medium known to those skilled in the art. The processor may be connected to the processor so that the processor reads information from or writes information to the storage medium, or the storage medium may be a component of the processor. The processor and the storage medium are located in the ASIC. The software module may be stored in the memory of the mobile terminal or may be stored in a memory card inserted in the mobile terminal. For example, when a device (eg, a mobile terminal) uses a relatively large capacity MEGA-SIM card or a large capacity flash memory device, the software module is stored in the MEGA-SIM card or the large capacity flash memory device. Also good.

  One or more functional blocks and / or one or more combinations of functional blocks in the functional block diagram described in FIG. 1 may be a general purpose processor, digital signal processor (for performing the functions described herein) DSP, application specific integrated circuit (ASIC), field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, or any suitable combination thereof May be. One or more functional blocks and / or one or more combinations of functional blocks in the functional block diagram depicted in FIG. 1 may be, for example, a combination of computing devices, such as a combination of a DSP and a microprocessor, A combination of processors, one or more microprocessors combined with DSP communications, or any other configuration may be implemented.

  Although the present invention has been described above with reference to specific embodiments, the above description is merely illustrative and does not limit the scope of protection of the present invention. Various modifications and changes may be made to the present invention without departing from the spirit and principle of the present invention, and these modifications and changes are also within the scope of the present invention.

In this embodiment, the second calculation unit 301 may calculate the distance between any two unlabeled sample characteristics using conventional methods, even the predetermined threshold is set according to the actual requirements Good.

Claims (19)

A classification network training device for character recognition,
Extraction means for extracting features of each unlabeled sample containing characters;
A construction means for constructing a sample pair based on the characteristics of each extracted unlabeled sample;
First training means for training a symmetric network based on the constructed sample pair;
Initialization means for initializing a classification network for character recognition using parameters of the trained symmetric network;
And a second training means for training the initialized classification network with a labeled sample containing characters.   It is determined whether or not the trained classification network satisfies a predetermined condition. If the trained classification network does not satisfy the predetermined condition, each label including the characters is not included using the trained classification network. The apparatus according to claim 1, further comprising: a determination unit that extracts sample characteristics and outputs the trained classification network if the trained classification network satisfies the predetermined condition. The extraction means includes
Extracting character features from the unlabeled sample containing direct characters, or
The apparatus according to claim 1, wherein the unlabeled sample is input to a trained network using labeled samples including characters, and the output result is extracted. The construction means includes
The apparatus according to claim 1, comprising first determining means for determining a first similar sample pair and a first dissimilar sample pair based on characteristics of each extracted unlabeled sample. The construction means includes
Second determining means for determining any two labeled samples of each labeled sample as a second similar sample pair or a second dissimilar sample pair based on the label of the labeled sample;
First calculating means for calculating a ratio of the number of the determined second similar sample pairs and the second dissimilar sample pairs;
The first similar sample such that a ratio of the number of the first similar sample pair and the first dissimilar sample pair is equal to a ratio of the number of the second similar sample pair and the second dissimilar sample pair. The apparatus according to claim 4, further comprising third determining means for determining a ratio of the number of pairs to the first dissimilar sample pair. The first determining means includes
Second calculating means for calculating a distance between features of any two unlabeled samples of each extracted unlabeled sample;
Fourth determining means for determining any two unlabeled samples having a distance between features smaller than a predetermined threshold as the first similar sample pair;
The apparatus according to claim 4, further comprising: fifth determination means for determining any two unlabeled samples whose distance between features is equal to or greater than the predetermined threshold as the first dissimilar sample pair. The feature extracted by the extraction means is a classification result of the unlabeled sample,
The first determining means includes
Sixth determination means for determining any two unlabeled samples having the same classification result as the first similar sample pair;
The 7th determination means which determines the arbitrary two unlabeled samples from which a classification result differs as said 1st dissimilar sample pair, The apparatus of Claim 4 characterized by the above-mentioned. The symmetric network is a Siamese network having two convolutional neural networks arranged symmetrically,
The apparatus of claim 1, wherein the classification network for character recognition is a convolutional neural network.   9. The apparatus according to claim 8, wherein the initialization means initializes a convolutional neural network to be the classification network using parameters of any one convolutional neural network in a trained sham network.   A character recognition device comprising a classification network for character recognition trained by the device according to claim 1. A classification network training method for character recognition,
Extracting the features of each unlabeled sample containing characters;
Building a sample pair based on the characteristics of each extracted unlabeled sample;
Training a symmetric network based on the constructed sample pairs;
Initializing a classification network for character recognition using parameters of the trained symmetric network;
Training the initialized classification network with labeled samples containing characters. Determining whether the trained classification network satisfies a predetermined condition;
If the trained classification network does not satisfy the predetermined condition, using the trained classification network to extract features of each unlabeled sample containing the characters;
12. The method of claim 11, further comprising: outputting the trained classification network if the trained classification network satisfies the predetermined condition. Extracting the features of each unlabeled sample containing said characters,
Extracting character features from the unlabeled sample containing direct characters, or
11. The method of claim 10, comprising: inputting the unlabeled sample into a network trained with labeled samples including characters and characterizing the output result. Building a sample pair based on the characteristics of each extracted unlabeled sample comprises:
11. The method of claim 10, comprising determining a first similar sample pair and a first dissimilar sample pair based on the characteristics of each extracted unlabeled sample. Building a sample pair based on the characteristics of each extracted unlabeled sample comprises:
Determining any two labeled samples of each labeled sample as a second similar sample pair or a second dissimilar sample pair based on the label of the labeled sample;
Calculating a ratio of the determined number of the second similar sample pairs and the second dissimilar sample pairs;
The first similar sample such that a ratio of the number of the first similar sample pair and the first dissimilar sample pair is equal to a ratio of the number of the second similar sample pair and the second dissimilar sample pair. 15. The method of claim 14, further comprising: determining a ratio of the number of pairs to the first dissimilar sample pair. Based on the characteristics of each extracted unlabeled sample, determining a first similar sample pair and a first dissimilar sample pair comprises:
Calculating the distance between the features of any two unlabeled samples of each extracted unlabeled sample;
Determining any two unlabeled samples whose distance between features is less than a predetermined threshold as the first similar sample pair;
And determining any two unlabeled samples having a distance between features greater than or equal to the predetermined threshold as the first dissimilar sample pair. The characteristic of each extracted unlabeled sample is the classification result of the unlabeled sample,
Based on the characteristics of each extracted unlabeled sample, determining a first similar sample pair and a first dissimilar sample pair includes:
Determining any two unlabeled samples with the same classification result as the first similar sample pair;
And determining any two unlabeled samples with different classification results as the first dissimilar sample pair. The symmetric network is a Siamese network having two convolutional neural networks arranged symmetrically,
The method according to claim 11, wherein the classification network for character recognition is a convolutional neural network. Initializing a classification network for character recognition using the parameters of the trained symmetric network comprises:
19. The method of claim 18, comprising initializing a convolutional neural network to be the classification network using parameters of any one convolutional neural network in a trained sham network.

Images