CN105335754A - Character recognition method and device - Google Patents

Abstract

The invention relates to a character recognition method and device. The character recognition method includes the following steps that: a convolutional neural network is adopted to perform feature extraction on a character image to be recognized, so that a character feature value is obtained; and an MQDF classifier is adopted to perform classification processing on the character feature value, so that a character corresponding to the character image to be recognized can be obtained. According to the character recognition method and device provided by the embodiments of the invention, the convolutional neural network is adopted to perform feature extraction on the character image to be recognized, so that the obtained character feature value is more precise and has high anti-interference ability, and therefore, the accuracy rate of the character recognized by the MQDF classifier according to the character feature value is high.

Description

Character recognition method and device

Technical field

The disclosure relates to mode identification technology, particularly relates to a kind of character recognition method and device.

Background technology

The research purpose of mode identification technology is the recognition mechanism of the brain according to people, by computer simulation, constructs the task that people can be replaced to complete classification and identification, and then carries out the machine system of automatic information processing.The development of Chinese Character Recognition in pattern-recognition has certain history, and for traditional Chinese Character Recognition, if noise is little, discrimination can reach more than 95%.Most typical application is exactly I.D. identification and business card recognition.

Traditional Chinese characters recognition method based on traditional feature extraction, and then adopts secondary classification function (ModifiedQuadraticDiscriminantFunction, the MQDF) sorter improved to carry out Chinese Character Recognition.

But traditional Chinese Character Recognition extracts feature based on artificial experience usually, does not possess certain noise antijamming capability.Therefore, when noise is larger, the discrimination of Chinese character can be caused lower.

Summary of the invention

For overcoming Problems existing in correlation technique, the disclosure provides a kind of character recognition method and device.

According to the first aspect of disclosure embodiment, a kind of character recognition method is provided, comprises:

Adopt convolutional neural networks to extract feature to character image to be identified, obtain character features value;

Adopt MQDF sorter to carry out classification process to described character features value, obtain the word that described character image to be identified is corresponding.

In one embodiment, described method also comprises:

Model training is carried out to described convolutional neural networks.

In one embodiment, described model training is carried out to described convolutional neural networks, comprising:

The convolutional layer that character image sample to be identified and label are input to described convolutional neural networks is processed, obtains convolution eigenwert;

The full articulamentum that described convolution eigenwert is input to described convolutional neural networks is processed, obtains full connection features value;

The loss function layer that described full connection features value is input to described convolutional neural networks is carried out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in described loss function layer converges to convergence threshold, otherwise repeat to carry out model training to described convolutional neural networks.

In one embodiment, the described convolutional layer described character image sample to be identified and label being input to described convolutional neural networks processes, and obtains convolution eigenwert, comprising:

Described character image sample to be identified and described label are input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert and process, obtain the first convolution eigenwert;

Described (n-1)th convolution eigenwert is input in the n-th convolutional layer of described convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

In one embodiment, described described (n-1)th convolution eigenwert being input in the n-th convolutional layer of described convolutional neural networks processes, and obtains the n-th convolution eigenwert, comprising:

Described (n-1)th convolution eigenwert is input to the n-th true convolutional layer of described n-th convolutional layer, the multiple convolution kernels in the described n-th true convolutional layer carry out convolution to described (n-1)th convolution eigenwert, obtain the (n-1)th true convolution eigenwert;

Described (n-1)th true convolution eigenwert is input in the n-th pond layer and carries out pondization process, obtain the n-th convolution eigenwert.

In one embodiment, described the loss function layer that described full connection features value is input to described convolutional neural networks is carried out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in described loss function layer converges to convergence threshold, otherwise repeat to carry out model training to described convolutional neural networks, comprising:

The loss function layer that described full connection features value is input to described convolutional neural networks is carried out loss function calculating, obtains loss function value, described loss function is the difference functions of classification value and label;

When described loss function value be less than last train the described loss function value obtained time, backpropagation is carried out to adjust the parameter of convolution kernel in described convolutional layer to described convolutional neural networks, and repeat to carry out model training to described convolutional neural networks, until the described loss function value obtained is less than loss function threshold value and restrains.

In one embodiment, described employing convolutional neural networks extracts feature to character image to be identified, obtains character features value, comprising:

The convolutional layer that described character image to be identified is input to described convolutional neural networks is processed, obtains convolution eigenwert;

The full articulamentum that described convolution eigenwert is input to described convolutional neural networks is processed, obtains full connection features value;

The function layer that described full connection features value is input to described convolutional neural networks is carried out function process, obtains described character features value.

In one embodiment, described the convolutional layer that described character image to be identified is input to described convolutional neural networks to be processed, obtains convolution eigenwert, comprising:

Described character image to be identified is input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert and processes, obtain the first convolution eigenwert;

Described (n-1)th convolution eigenwert is input in the n-th convolutional layer of described convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

In one embodiment, described method also comprises:

To the operation that described character image to be identified is normalized, or to the operation that described character image sample to be identified is normalized.

According to the second aspect of disclosure embodiment, a kind of character recognition device is provided, comprises:

Extraction module, is configured to adopt convolutional neural networks to extract feature to character image to be identified, obtains character features value;

Sort module, is configured to adopt MQDF sorter to carry out classification process to the described character features value that described extraction module obtains, obtains the word that described character image to be identified is corresponding.

In one embodiment, described device also comprises:

Training module, is configured to carry out model training to described convolutional neural networks.

In one embodiment, described training module comprises:

First process submodule, the convolutional layer being configured to character image sample to be identified and label to be input to described convolutional neural networks processes, and obtains convolution eigenwert;

Second process submodule, the full articulamentum being configured to the described convolution eigenwert that described first process submodule obtains to be input to described convolutional neural networks processes, and obtains full connection features value;

3rd process submodule, the loss function layer being configured to the described full connection features value that described second process submodule obtains to be input to described convolutional neural networks carries out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in described loss function layer converges to convergence threshold, otherwise repeat to carry out model training to described convolutional neural networks.

In one embodiment, described first process submodule comprises:

First processing unit, is configured to described character image sample to be identified and described label to be input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert process, obtains the first convolution eigenwert;

N-th processing unit, be configured to described (n-1)th convolution eigenwert to be input in the n-th convolutional layer of described convolutional neural networks process, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

In one embodiment, described n-th processing unit comprises:

Process of convolution subelement, be configured to the n-th true convolutional layer described (n-1)th convolution eigenwert being input to described n-th convolutional layer, multiple convolution kernels in described n-th true convolutional layer carry out convolution to described (n-1)th convolution eigenwert, obtain the (n-1)th true convolution eigenwert;

Pondization process subelement, the described (n-1)th true convolution eigenwert being configured to described process of convolution subelement to obtain is input in the n-th pond layer carries out pondization process, obtains the n-th convolution eigenwert.

In one embodiment, described 3rd process submodule comprises:

Computing unit, the loss function layer be configured to described full connection features value is input to described convolutional neural networks carries out loss function calculating, and obtain loss function value, described loss function is the difference functions of classification value and label;

Adjustment unit, be configured to when the described loss function value that described computing unit calculates be less than last train the described loss function value obtained time, backpropagation is carried out to adjust the parameter of convolution kernel in described convolutional layer to described convolutional neural networks, and repeat to carry out model training to described convolutional neural networks, until the described loss function value obtained is less than loss function threshold value and restrains.

In one embodiment, described extraction module comprises:

Process of convolution submodule, the convolutional layer be configured to described character image to be identified is input to described convolutional neural networks processes, and obtains convolution eigenwert;

Full connection handling submodule, the full articulamentum that the described convolution eigenwert being configured to be obtained by described process of convolution submodule is input to described convolutional neural networks processes, and obtains full connection features value;

Function process submodule, the function layer being configured to the described full connection features value that described full connection handling submodule obtains to be input to described convolutional neural networks carries out function process, obtains described character features value.

In one embodiment, described process of convolution submodule comprises:

First processing unit, is configured to described character image to be identified to be input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert process, obtains the first convolution eigenwert;

N-th processing unit, be configured to described (n-1)th convolution eigenwert to be input in the n-th convolutional layer of described convolutional neural networks process, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

In one embodiment, described device also comprises:

Normalization module, is configured to the operation be normalized described character image to be identified, or to the operation that described character image sample to be identified is normalized.

According to the third aspect of disclosure embodiment, a kind of character recognition device is provided, comprises:

Processor;

For the storer of storage of processor executable instruction;

Wherein, processor is configured to:

Adopt convolutional neural networks to extract feature to character image to be identified, obtain character features value;

Adopt MQDF sorter to carry out classification process to described character features value, obtain the word that described character image to be identified is corresponding.

The technical scheme that embodiment of the present disclosure provides can comprise following beneficial effect: adopt convolutional neural networks to extract feature to character image to be identified, make the character features value that obtains more accurate, there is good noise antijamming capability, thus the word accuracy rate that employing MQDF sorter is identified according to this word eigenwert is higher.

By carrying out model training to convolutional neural networks, make this convolutional neural networks have suitable model parameter, thus provide condition for the follow-up character features value obtaining enriching.

By carrying out model training to convolutional neural networks, good model parameter can be obtained, thus provide condition for this convolutional neural networks of follow-up employing obtains character features value.

Obtain the n-th convolution eigenwert by multiple convolutional layer, model training is carried out to convolutional neural networks for follow-up condition is provided.

Obtain the n-th convolution eigenwert by the process of Convolution sums pond, implementation is simple, and reduces the dimension of convolution eigenwert, thus can improve follow-up treatment effeciency.

Determine that the mode of convolution kernel parameter in convolutional layer is simple, effectively, be easy to realize.

Character features value can be obtained, for follow-up identification word provides condition by process of convolution, full connection handling sum functions process.

The character features value enriched can be obtained by multiple convolutional layer, for follow-up high-accuracy identify that word provides condition.

By the operation be normalized, be conducive to improving convolutional neural networks to the recognition efficiency of character image to be identified.

Should be understood that, it is only exemplary and explanatory that above general description and details hereinafter describe, and can not limit the disclosure.

Accompanying drawing explanation

Accompanying drawing to be herein merged in instructions and to form the part of this instructions, shows embodiment according to the invention, and is used from instructions one and explains principle of the present invention.

Figure 1A is the process flow diagram of a kind of character recognition method according to an exemplary embodiment.

Figure 1B is a kind of method flow diagram obtaining character features value according to an exemplary embodiment.

Fig. 1 C is a kind of method flow diagram obtaining character features value according to an exemplary embodiment.

Fig. 2 is the process flow diagram of the another kind of character recognition method according to an exemplary embodiment.

Fig. 3 is a kind of method flow diagram convolutional neural networks being carried out to model training according to an exemplary embodiment.

Fig. 4 is a kind of method flow diagram obtaining convolution eigenwert according to an exemplary embodiment.

Fig. 5 is a kind of method flow diagram adjusting the parameter of convolution kernel in convolutional layer according to an exemplary embodiment.

Fig. 6 is the block diagram of a kind of character recognition device according to an exemplary embodiment.

Fig. 7 is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Fig. 8 is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Fig. 9 A is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Fig. 9 B is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Fig. 9 C is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Figure 10 A is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Figure 10 B is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Figure 11 is the block diagram of the another kind of character recognition device according to an exemplary embodiment.

Figure 12 is a kind of block diagram being applicable to character recognition device according to an exemplary embodiment.

Embodiment

Here will be described exemplary embodiment in detail, its sample table shows in the accompanying drawings.When description below relates to accompanying drawing, unless otherwise indicated, the same numbers in different accompanying drawing represents same or analogous key element.Embodiment described in following exemplary embodiment does not represent all embodiments consistent with the present invention.On the contrary, they only with as in appended claims describe in detail, the example of apparatus and method that aspects more of the present invention are consistent.

Figure 1A is the process flow diagram of a kind of character recognition method according to an exemplary embodiment, and as shown in Figure 1A, this character recognition method can be applicable on various terminal device, and this character recognition method comprises the following steps S101-S102:

In step S101, adopt convolutional neural networks to extract feature to character image to be identified, obtain character features value.

Convolutional neural networks (ConvolutionalNeuralNetwork, CNN) be the one of degree of depth learning network model, degree of depth learning network model is a new field in machine learning research, the neural network that it is foundation, simulation human brain carries out analytic learning, the mechanism that it can imitate human brain carrys out decryption, such as text.Due to the learning algorithm that convolutional neural networks is a real sandwich construction, so it can utilize the relativeness in space to reduce number of parameters, thus improve training performance, and then improve Text region performance.

In this embodiment, owing to carrying out model training to convolutional neural networks, therefore this convolutional neural networks can be adopted to extract feature to character image to be identified, obtain character features value.

Wherein, the word in this embodiment can including, but not limited to Chinese character.

Because the convolutional neural networks being in operational phase comprises convolutional layer, full articulamentum sum functions layer (softmax), therefore adopt this convolutional neural networks to extract feature to character image to be identified, obtain the process of character features value as shown in Figure 1B, can comprise the following steps:

In step S1011, the convolutional layer that character image to be identified is input to convolutional neural networks is processed, obtains convolution eigenwert.

In this embodiment, the convolutional layer of convolutional neural networks can have multiple, such as can comprise first volume lamination, volume Two lamination ... n-th convolutional layer, therefore convolutional layer character image to be identified being input to convolutional neural networks processes, the process obtaining convolution eigenwert can comprise:

Character image to be identified is input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert and processes, obtain the first convolution eigenwert.(n-1)th convolution eigenwert is input in the n-th convolutional layer of convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer.

Wherein, the n-th convolution eigenwert is the convolution eigenwert that the convolutional layer of this convolutional neural networks exports, and is also the convolution eigenwert obtained in step S1011.

Suppose, during n=5, the process obtaining convolution eigenwert can be: be input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert by character image to be identified and process, obtain the first convolution eigenwert.First convolution eigenwert is input in the volume Two lamination of convolutional neural networks and processes, obtain the second convolution eigenwert.Second convolution eigenwert is input in the 3rd convolutional layer of convolutional neural networks and processes, obtain the 3rd convolution eigenwert.3rd convolution eigenwert is input in the Volume Four lamination of convolutional neural networks and processes, obtain Volume Four and amass eigenwert.Volume Four is amassed eigenwert to be input in the 5th convolutional layer of convolutional neural networks and to process, obtain the 5th convolution eigenwert.

In step S1012, full articulamentum convolution eigenwert being input to convolutional neural networks processes, and obtains full connection features value.

After obtaining convolution eigenwert, convolution eigenwert can be input to the full articulamentum of convolutional neural networks, be processed by full articulamentum, obtain full connection features value.

In step S1013, the function layer that full connection features value is input to convolutional neural networks is carried out function process, obtains character features value.

After obtaining full connection features value, full connection features value can be input to the function layer of convolutional neural networks, carry out function process by function layer, obtain character features value.

As can be seen here, can be obtained the character features value of character image to be identified by above-mentioned steps S1011-1013, because this convolutional neural networks has multiple convolutional layer, therefore the character features value obtained is more accurate.

In addition, as shown in Figure 1 C, before step S1011, can also comprise:

In step S1010, to the operation that character image to be identified is normalized.

In this embodiment, operation character image to be identified is normalized can including, but not limited to rotating character image to be identified, the operation such as convergent-divergent, sharpening.

By the operation be normalized character image to be identified, be conducive to improving convolutional neural networks to the recognition efficiency of character image to be identified.

In step s 102, adopt MQDF sorter to carry out classification process to character features value, obtain the word that character image to be identified is corresponding.

In this embodiment, secondary classification function (MQDF) sorter to improving is needed to train, to train the parameter of MQDF sorter.

Wherein, to the process that MQDF sorter is trained can be: character features value sample and tag along sort are input to and carry out classification process with MQDF sorter, obtain classification results; Classification results and tag along sort are compared, if the difference of the two is less than predetermined threshold value, then determine to train MQDF sorter, namely can classify with this MQDF sorter, if the difference of the two is more than or equal to predetermined threshold value, then adjust the parameter of this MQDF sorter, and repeat said process, until the difference of the two is less than predetermined threshold value.

Because tag along sort represents correct classification results, therefore, if the difference of the classification results exported and tag along sort is less than predetermined threshold value, then show that the classification results exported is close to correct classification results, namely shows to train MQDF sorter.

After the parameter training MQDF sorter, the MQDF sorter trained can be adopted to carry out classification process to character features value, thus word corresponding to character image to be identified can be obtained.

Because the character features obtained in this embodiment is more accurate, so be conducive to the recognition accuracy improving word.

Above-mentioned character recognition method embodiment, convolutional neural networks is adopted to extract feature to character image to be identified, make the character features value that obtains more accurate, there is good noise antijamming capability, thus the word accuracy rate making to adopt MQDF sorter to identify according to this word eigenwert is higher.

Fig. 2 is the process flow diagram of the another kind of character recognition method according to an exemplary embodiment, as Fig. 2 shows, before step S101, can also comprise the steps:

In the step s 100, model training is carried out to convolutional neural networks.

In this embodiment, the object of convolutional neural networks being carried out to model training is the parameter adjusting convolutional neural networks, after inputting this convolutional neural networks to make character image to be identified, under the prerequisite that can obtain correct classification, obtain accurate character features value.

Above-mentioned character recognition method embodiment, by carrying out model training to convolutional neural networks, makes this convolutional neural networks have suitable model parameter, thus provides condition for the follow-up character features value obtaining enriching.

Fig. 3 is a kind of method flow diagram convolutional neural networks being carried out to model training according to an exemplary embodiment, due in the model training stage, CNN can comprise multiple convolutional layer, full articulamentum and loss function layer (softmaxloss), therefore can show as Fig. 3 the process that convolutional neural networks carries out model training, comprise the steps:

In step S300, to the operation that character image sample to be identified is normalized.

In this embodiment, operation character image sample to be identified is normalized can including, but not limited to rotating character image sample to be identified, the operation such as convergent-divergent, sharpening.

By the operation be normalized character image sample to be identified, be conducive to the training effectiveness improving convolutional neural networks.

In step S301, the convolutional layer that character image sample to be identified and label (i.e. label) are input to convolutional neural networks is processed, obtains convolution eigenwert.

Because the convolutional layer of convolutional neural networks can have multiple, such as, can comprise first volume lamination, volume Two lamination ... n-th convolutional layer, therefore as shown in Figure 4, this step S301 can comprise:

In step S3011, character image sample to be identified and label are input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert and process, obtain the first convolution eigenwert.

In step S3012, (n-1)th convolution eigenwert is input in the n-th convolutional layer of convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, the n-th convolution eigenwert be convolutional neural networks convolutional layer export convolution eigenwert.

Because each convolutional layer comprises a true convolutional layer and a pond layer, therefore step S3012 can comprise:

(n-1)th convolution eigenwert is input to the n-th true convolutional layer of the n-th convolutional layer, the multiple convolution kernels in the n-th true convolutional layer carry out convolution to the (n-1)th convolution eigenwert, obtain the (n-1)th true convolution eigenwert.(n-1)th true convolution eigenwert is input in the n-th pond layer and carries out pondization process, obtain the n-th convolution eigenwert.

In this embodiment, the object of pond layer carries out dimension-reduction treatment to convolution eigenwert, to improve follow-up treatment effeciency.

Wherein, label is the word for identifying corresponding to character image sample to be identified.Such as, character image sample to be identified be " in " image of word, then label be " in " word.

In step s 302, full articulamentum convolution eigenwert being input to convolutional neural networks processes, and obtains full connection features value.

In step S303, the loss function layer that full connection features value is input to convolutional neural networks is carried out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in loss function layer converges to convergence threshold, otherwise repeat to carry out model training to convolutional neural networks.

As shown in Figure 5, step S303 can comprise:

In step S3031, the loss function layer that full connection features value is input to convolutional neural networks is carried out loss function calculating, obtains loss function value, loss function is the difference functions of classification value and label.

Due to, the convolutional neural networks in this embodiment has the sorter carried, therefore the classification value in this step refers to the classification value of convolutional neural networks.

In step S3032, when loss function value be less than last train the loss function value obtained time, backpropagation is carried out to adjust the parameter of convolution kernel in convolutional layer to convolutional neural networks, and repeat to carry out model training to convolutional neural networks, until the loss function value obtained is less than loss function threshold value and restrains.

When loss function value is less than loss function threshold value and restrains, show that the difference of classification value and label is very little, namely classification value is close to label, and in other words, classification results is close to correct word, and namely classification is correct.

As can be seen here, the above-mentioned mode determining convolution kernel parameter in convolutional layer simple, effectively, be easy to realize.

Above-described embodiment, by carrying out model training to convolutional neural networks, can obtain good model parameter, thus provides condition for this convolutional neural networks of follow-up employing obtains character features value.

Corresponding with aforementioned character recognition method embodiment, the disclosure additionally provides character recognition device embodiment.

Fig. 6 is the block diagram of a kind of character recognition device according to an exemplary embodiment, and as shown in Figure 6, character recognition device comprises: extraction module 61 and sort module 62.

Extraction module 61 is configured to adopt convolutional neural networks to extract feature to character image to be identified, obtains character features value.

Convolutional neural networks (ConvolutionalNeuralNetwork, CNN) be the one of degree of depth learning network model, degree of depth learning network model is a new field in machine learning research, the neural network that it is foundation, simulation human brain carries out analytic learning, the mechanism that it can imitate human brain carrys out decryption, such as text.Due to the learning algorithm that convolutional neural networks is a real sandwich construction, so it can utilize the relativeness in space to reduce number of parameters, thus improve training performance, and then improve Text region performance.

In this embodiment, owing to carrying out model training to convolutional neural networks, therefore this convolutional neural networks can be adopted to extract feature to character image to be identified, obtain character features value.

Wherein, the word in this embodiment can including, but not limited to Chinese character.

Sort module 62 is configured to adopt MQDF sorter to carry out classification process to the character features value that extraction module 61 obtains, and obtains the word that character image to be identified is corresponding.

In this embodiment, secondary classification function (MQDF) sorter to improving is needed to train, to train the parameter of MQDF sorter.

Wherein, to the process that MQDF sorter is trained can be: character features value sample and tag along sort are input to and carry out classification process with MQDF sorter, obtain classification results; Classification results and tag along sort are compared, if the difference of the two is less than predetermined threshold value, then determine to train MQDF sorter, namely can classify with this MQDF sorter, if the difference of the two is more than or equal to predetermined threshold value, then adjust the parameter of this MQDF sorter, and repeat said process, until the difference of the two is less than predetermined threshold value.

Because tag along sort represents correct classification results, therefore, if the difference of the classification results exported and tag along sort is less than predetermined threshold value, then show that the classification results exported is close to correct classification results, namely shows to train MQDF sorter.

After the parameter training MQDF sorter, the MQDF sorter trained can be adopted to carry out classification process to character features value, thus word corresponding to character image to be identified can be obtained.

Because the character features obtained in this embodiment is more accurate, so be conducive to the recognition accuracy improving word.

Device is as shown in Figure 6 for realizing above-mentioned method flow as shown in Figure 1A, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, convolutional neural networks is adopted to extract feature to character image to be identified by extraction module, make the character features value that obtains more accurate, there is good noise antijamming capability, thus the word accuracy rate that employing MQDF sorter is identified according to this word eigenwert is higher.

Fig. 7 is the block diagram of the another kind of character recognition device according to an exemplary embodiment, and as shown in Figure 7, on above-mentioned basis embodiment illustrated in fig. 6, this device also can comprise: training module 60.

Training module 60 is configured to carry out model training to convolutional neural networks.

In this embodiment, the object of convolutional neural networks being carried out to model training is the parameter adjusting convolutional neural networks, after inputting this convolutional neural networks to make character image to be identified, under the prerequisite that can obtain correct classification, obtain accurate character features value.

Device is as shown in Figure 7 for realizing above-mentioned method flow as shown in Figure 2, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, carries out model training by training module to convolutional neural networks, makes this convolutional neural networks have suitable model parameter, thus provides condition for the follow-up character features value obtaining enriching.

Fig. 8 is the block diagram of the another kind of character recognition device according to an exemplary embodiment, as shown in Figure 8, on above-mentioned basis embodiment illustrated in fig. 7, training module 60 comprises: the first process submodule 601, second processes submodule 602 and the 3rd process submodule 603.

The convolutional layer that first process submodule 601 is configured to character image sample to be identified and label to be input to convolutional neural networks processes, and obtains convolution eigenwert.

Because the convolutional layer of convolutional neural networks can have multiple, such as, can comprise first volume lamination, volume Two lamination ... n-th convolutional layer, therefore the process that the first process submodule 601 obtains convolution eigenwert can comprise:

Character image sample to be identified and label are input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert and process, obtain the first convolution eigenwert.

(n-1)th convolution eigenwert is input in the n-th convolutional layer of convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, the n-th convolution eigenwert be convolutional neural networks convolutional layer export convolution eigenwert.

The full articulamentum that second process submodule 602 is configured to the convolution eigenwert that the first process submodule 601 obtains to be input to convolutional neural networks processes, and obtains full connection features value.

The loss function layer that 3rd process submodule 603 is configured to the full connection features value that the second process submodule 602 obtains to be input to convolutional neural networks carries out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in loss function layer converges to convergence threshold, otherwise repeat to carry out model training to convolutional neural networks.

The process of the 3rd process submodule 603 training pattern can comprise: first, the loss function layer that full connection features value is input to convolutional neural networks is carried out loss function calculating, obtains loss function value, loss function is the difference functions of classification value and label.

Because the convolutional neural networks in this embodiment has the sorter carried, therefore the classification value in this step refers to the classification value of convolutional neural networks.

Secondly, when loss function value be less than last train the loss function value obtained time, backpropagation is carried out to adjust the parameter of convolution kernel in convolutional layer to convolutional neural networks, and repeat to carry out model training to convolutional neural networks, until the loss function value obtained is less than loss function threshold value and restrains.

When loss function value is less than loss function threshold value and restrains, show that the difference of classification value and label is very little, namely classification value is close to label, and in other words, classification results is close to correct word, and namely classification is correct.

As can be seen here, the above-mentioned mode determining convolution kernel parameter in convolutional layer simple, effectively, be easy to realize.Device is as shown in Figure 8 for realizing above-mentioned method flow as shown in Figure 3, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, by carrying out model training to convolutional neural networks, can obtain good model parameter, thus provides condition for this convolutional neural networks of follow-up employing obtains character features value.

Fig. 9 A is the block diagram of the another kind of character recognition device according to an exemplary embodiment, and as shown in Figure 9 A, on above-mentioned basis embodiment illustrated in fig. 8, the first process submodule 601 can comprise: the first processing unit 6011 and the n-th processing unit 6012.

First processing unit 6011 is configured to character image sample to be identified and label to be input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert process, and obtains the first convolution eigenwert.

N-th processing unit 6012 is configured to the (n-1)th convolution eigenwert to be input in the n-th convolutional layer of convolutional neural networks process, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, the n-th convolution eigenwert be convolutional neural networks convolutional layer export convolution eigenwert.Device is as shown in Figure 9 A for realizing above-mentioned method flow as shown in Figure 4, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, obtains the n-th convolution eigenwert by the first processing unit and the n-th processing unit, carries out model training provide condition for follow-up convolutional neural networks.

Fig. 9 B is the block diagram of the another kind of character recognition device according to an exemplary embodiment, as shown in Figure 9 B, on the basis of above-mentioned Fig. 9 A illustrated embodiment, the n-th processing unit 6012 can comprise: process of convolution subelement 60121 and pondization process subelement 60122.

Process of convolution subelement 60121 is configured to the n-th true convolutional layer the (n-1)th convolution eigenwert being input to the n-th convolutional layer, and the multiple convolution kernels in the n-th true convolutional layer carry out convolution to the (n-1)th convolution eigenwert, obtains the (n-1)th true convolution eigenwert.

The (n-1)th true convolution eigenwert that pondization process subelement 60122 is configured to process of convolution subelement 60221 to obtain is input in the n-th pond layer carries out pondization process, obtains the n-th convolution eigenwert.

In this embodiment, the object of pond layer carries out dimension-reduction treatment to convolution eigenwert, to improve follow-up treatment effeciency.

Wherein, label is the word for identifying corresponding to character image sample to be identified.Such as, character image sample to be identified be " in " image of word, then label be " in " word.

Device is as shown in Figure 9 B for realizing above-mentioned method flow as shown in Figure 4, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, obtain the n-th convolution eigenwert by process of convolution subelement and pondization process subelement, implementation is simple, and reduces the dimension of convolution eigenwert, thus can improve follow-up treatment effeciency.

Fig. 9 C is the block diagram of the another kind of character recognition device according to an exemplary embodiment, and as shown in Figure 9 C, on above-mentioned basis embodiment illustrated in fig. 8, the 3rd process submodule 603 can comprise: computing unit 6031 and adjustment unit 6032.

Computing unit 6031, the loss function layer being configured to full connection features value to be input to convolutional neural networks carries out loss function calculating, obtains loss function value, and loss function is the difference functions of classification value and label.

Due to, the convolutional neural networks in this embodiment has the sorter carried, therefore classification value refers to the classification value of convolutional neural networks.

Adjustment unit 6032, be configured to when the loss function value that computing unit 6031 calculates be less than last train the loss function value obtained time, backpropagation is carried out to adjust the parameter of convolution kernel in convolutional layer to convolutional neural networks, and repeat to carry out model training to convolutional neural networks, until the loss function value obtained is less than loss function threshold value and restrains.

When loss function value is less than loss function threshold value and restrains, show that the difference of classification value and label is very little, namely classification value is close to label, and in other words, classification results is close to correct word, and namely classification is correct.

As can be seen here, the above-mentioned mode determining convolution kernel parameter in convolutional layer simple, effectively, be easy to realize.

Device is as shown in Figure 9 C for realizing above-mentioned method flow as shown in Figure 5, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, by convolution kernel parameter in computing unit and adjustment unit determination convolutional layer, determines that mode is simple, effectively, is easy to realize.

Figure 10 A is the block diagram of the another kind of character recognition device according to an exemplary embodiment, as shown in Figure 10 A, on above-mentioned basis embodiment illustrated in fig. 6, extraction module 61 can comprise: process of convolution submodule 611, full connection handling submodule 612 sum functions process submodule 613.

Process of convolution submodule 611, the convolutional layer being configured to character image to be identified to be input to convolutional neural networks processes, and obtains convolution eigenwert.

In this embodiment, the convolutional layer of convolutional neural networks can have multiple, such as, can comprise first volume lamination, volume Two lamination ... n-th convolutional layer, therefore the process that process of convolution submodule 611 obtains convolution eigenwert can comprise:

Character image to be identified is input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert and processes, obtain the first convolution eigenwert.(n-1)th convolution eigenwert is input in the n-th convolutional layer of convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer.

Wherein, the n-th convolution eigenwert is the convolution eigenwert that the convolutional layer of this convolutional neural networks exports, and is also the convolution eigenwert that process of convolution submodule 611 obtains.

Suppose, during n=5, the process that process of convolution submodule 611 obtains convolution eigenwert can be: be input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert by character image to be identified and process, obtain the first convolution eigenwert.First convolution eigenwert is input in the volume Two lamination of convolutional neural networks and processes, obtain the second convolution eigenwert.Second convolution eigenwert is input in the 3rd convolutional layer of convolutional neural networks and processes, obtain the 3rd convolution eigenwert.3rd convolution eigenwert is input in the Volume Four lamination of convolutional neural networks and processes, obtain Volume Four and amass eigenwert.Volume Four is amassed eigenwert to be input in the 5th convolutional layer of convolutional neural networks and to process, obtain the 5th convolution eigenwert.

Full connection handling submodule 612, the full articulamentum that the convolution eigenwert being configured to process of convolution submodule 611 to obtain is input to convolutional neural networks processes, and obtains full connection features value.

After obtaining convolution eigenwert, convolution eigenwert can be input to the full articulamentum of convolutional neural networks, be processed by full articulamentum, obtain full connection features value.

Function process submodule 613, the function layer that the full connection features value being configured to be obtained by full connection handling submodule 612 is input to convolutional neural networks carries out function process, obtains character features value.

After obtaining full connection features value, full connection features value can be input to the function layer of convolutional neural networks, carry out function process by function layer, obtain character features value.Device is as shown in Figure 10 A for realizing above-mentioned method flow as shown in Figure 1B, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, can obtain character features value, for follow-up identification word provides condition by process of convolution submodule, full connection handling submodule sum functions process submodule.

Figure 10 B is the block diagram of the another kind of character recognition device according to an exemplary embodiment, and as shown in Figure 10 B, on the basis of above-mentioned Figure 10 A illustrated embodiment, process of convolution submodule 611 can comprise: the first processing unit 6111 and the n-th processing unit 6112.

First processing unit 6111 is configured to character image to be identified to be input in the first volume lamination of convolutional neural networks as the 0th convolution eigenwert process, and obtains the first convolution eigenwert.

N-th processing unit 6112 is configured to the (n-1)th convolution eigenwert to be input in the n-th convolutional layer of convolutional neural networks process, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, the n-th convolution eigenwert be convolutional neural networks convolutional layer export convolution eigenwert.

Device is as shown in Figure 10 B for realizing above-mentioned method flow as shown in Figure 1B, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, can obtain the character features value enriched by the first processing unit and the n-th processing unit, for follow-up high-accuracy identify that word provides condition.

Figure 11 is the block diagram of another character recognition device according to an exemplary embodiment, and as shown in figure 11, on above-mentioned Fig. 6, Fig. 7 or basis embodiment illustrated in fig. 8, this device also can comprise:

Normalization module 63, is configured to the operation be normalized character image to be identified, or to the operation that character image sample to be identified is normalized.

In this embodiment, the operation that normalization module 63 is normalized character image to be identified can including, but not limited to rotating character image to be identified, the operation such as convergent-divergent, sharpening.

Device is as shown in figure 11 for realizing above-mentioned method flow as shown in Fig. 1 C or Fig. 3, and the related content related to describes identical, does not repeat herein.

Above-mentioned character recognition device embodiment, the operation be normalized by normalization module, is conducive to improving convolutional neural networks to the recognition efficiency of character image to be identified.

About the device in above-described embodiment, wherein the concrete mode of modules, submodule executable operations has been described in detail in about the embodiment of the method, will not elaborate explanation herein.

Figure 12 is a kind of block diagram being applicable to character recognition device according to an exemplary embodiment.Such as, device 1200 can be mobile phone, computing machine, digital broadcast terminal, messaging devices, game console, tablet device, Medical Devices, body-building equipment, personal digital assistant etc.

With reference to Figure 12, device 1200 can comprise following one or more assembly: processing components 1202, storer 1204, power supply module 1206, multimedia groupware 1208, audio-frequency assembly 1210, the interface 1212 of I/O (I/O), sensor module 1214, and communications component 1216.

The integrated operation of the usual control device 1200 of processing components 1202, such as with display, call, data communication, camera operation and record operate the operation be associated.Treatment element 1202 can comprise one or more processor 1220 to perform instruction, to complete all or part of step of above-mentioned method.In addition, processing components 1202 can comprise one or more module, and what be convenient between processing components 1202 and other assemblies is mutual.Such as, processing element 1202 can comprise multi-media module, mutual with what facilitate between multimedia groupware 1208 and processing components 1202.

Storer 1204 is configured to store various types of data to be supported in the operation of equipment 1200.The example of these data comprises for any application program of operation on device 1200 or the instruction of method, contact data, telephone book data, message, picture, video etc.Storer 1204 can be realized by the volatibility of any type or non-volatile memory device or their combination, as static RAM (SRAM), Electrically Erasable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory EPROM (EPROM), programmable read only memory (PROM), ROM (read-only memory) (ROM), magnetic store, flash memory, disk or CD.

The various assemblies that electric power assembly 1206 is device 1200 provide electric power.Electric power assembly 1206 can comprise power-supply management system, one or more power supply, and other and the assembly generating, manage and distribute electric power for device 1200 and be associated.

Multimedia groupware 1208 is included in the screen providing an output interface between described device 1200 and user.In certain embodiments, screen can comprise liquid crystal display (LCD) and touch panel (TP).If screen comprises touch panel, screen may be implemented as touch-screen, to receive the input signal from user.Touch panel comprises one or more touch sensor with the gesture on sensing touch, slip and touch panel.Described touch sensor can the border of not only sensing touch or sliding action, but also detects the duration relevant to described touch or slide and pressure.In certain embodiments, multimedia groupware 1208 comprises a front-facing camera and/or post-positioned pick-up head.When equipment 1200 is in operator scheme, during as screening-mode or video mode, front-facing camera and/or post-positioned pick-up head can receive outside multi-medium data.Each front-facing camera and post-positioned pick-up head can be fixing optical lens systems or have focal length and optical zoom ability.

Audio-frequency assembly 1210 is configured to export and/or input audio signal.Such as, audio-frequency assembly 1210 comprises a microphone (MIC), and when device 1200 is in operator scheme, during as call model, logging mode and speech recognition mode, microphone is configured to receive external audio signal.The sound signal received can be stored in storer 1204 further or be sent via communications component 1216.In certain embodiments, audio-frequency assembly 1210 also comprises a loudspeaker, for output audio signal.

I/O interface 1212 is for providing interface between processing components 1202 and peripheral interface module, and above-mentioned peripheral interface module can be keyboard, some striking wheel, button etc.These buttons can include but not limited to: home button, volume button, start button and locking press button.

Sensor module 1214 comprises one or more sensor, for providing the state estimation of various aspects for device 1200.Such as, sensor module 1214 can detect the opening/closing state of equipment 1200, the relative positioning of assembly, such as described assembly is display and the keypad of device 1200, the position of all right pick-up unit 1200 of sensor module 1214 or device 1200 assemblies changes, the presence or absence that user contacts with device 1200, the temperature variation of device 1200 orientation or acceleration/deceleration and device 1200.Sensor module 1214 can comprise proximity transducer, be configured to without any physical contact time detect near the existence of object.Sensor module 1214 can also comprise optical sensor, as CMOS or ccd image sensor, for using in imaging applications.In certain embodiments, this sensor module 1214 can also comprise acceleration transducer, gyro sensor, Magnetic Sensor, pressure transducer or temperature sensor.

Communications component 1216 is configured to the communication being convenient to wired or wireless mode between device 1200 and other equipment.Device 1200 can access the wireless network based on communication standard, as WiFi, 2G or 3G, or their combination.In one exemplary embodiment, communication component 1216 receives from the broadcast singal of external broadcasting management system or broadcast related information via broadcast channel.In one exemplary embodiment, described communication component 1216 also comprises near-field communication (NFC) module, to promote junction service.Such as, can based on radio-frequency (RF) identification (RFID) technology in NFC module, Infrared Data Association (IrDA) technology, ultra broadband (UWB) technology, bluetooth (BT) technology and other technologies realize.

In the exemplary embodiment, device 1200 can be realized, for performing said method by one or more application specific integrated circuit (ASIC), digital signal processor (DSP), digital signal processing appts (DSPD), programmable logic device (PLD) (PLD), field programmable gate array (FPGA), controller, microcontroller, microprocessor or other electronic components.

In the exemplary embodiment, additionally provide a kind of non-transitory computer-readable recording medium comprising instruction, such as, comprise the storer 1204 of instruction, above-mentioned instruction can perform said method by the processor 1220 of device 1200.Such as, described non-transitory computer-readable recording medium can be ROM, random access memory (RAM), CD-ROM, tape, floppy disk and optical data storage devices etc.

Those skilled in the art, at consideration instructions and after putting into practice disclosed herein disclosing, will easily expect other embodiment of the present disclosure.The application is intended to contain any modification of the present disclosure, purposes or adaptations, and these modification, purposes or adaptations are followed general principle of the present disclosure and comprised the undocumented common practise in the art of the disclosure or conventional techniques means.Instructions and embodiment are only regarded as exemplary, and true scope of the present disclosure and spirit are pointed out by claim below.

Should be understood that, the disclosure is not limited to precision architecture described above and illustrated in the accompanying drawings, and can carry out various amendment and change not departing from its scope.The scope of the present disclosure is only limited by appended claim.

Claims (19)

1. a character recognition method, is characterized in that, comprising:

Adopt convolutional neural networks to extract feature to character image to be identified, obtain character features value;

Adopt MQDF sorter to carry out classification process to described character features value, obtain the word that described character image to be identified is corresponding.

2. character recognition method according to claim 1, is characterized in that, described method also comprises:

Model training is carried out to described convolutional neural networks.

3. character recognition method according to claim 2, is characterized in that, describedly carries out model training to described convolutional neural networks, comprising:

The convolutional layer that character image sample to be identified and label are input to described convolutional neural networks is processed, obtains convolution eigenwert;

The full articulamentum that described convolution eigenwert is input to described convolutional neural networks is processed, obtains full connection features value;

The loss function layer that described full connection features value is input to described convolutional neural networks is carried out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in described loss function layer converges to convergence threshold, otherwise repeat to carry out model training to described convolutional neural networks.

4. character recognition method according to claim 3, is characterized in that, the described convolutional layer described character image sample to be identified and label being input to described convolutional neural networks processes, and obtains convolution eigenwert, comprising:

Described character image sample to be identified and described label are input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert and process, obtain the first convolution eigenwert;

Described (n-1)th convolution eigenwert is input in the n-th convolutional layer of described convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

5. character recognition method according to claim 4, is characterized in that, described described (n-1)th convolution eigenwert being input in the n-th convolutional layer of described convolutional neural networks processes, and obtains the n-th convolution eigenwert, comprising:

Described (n-1)th convolution eigenwert is input to the n-th true convolutional layer of described n-th convolutional layer, the multiple convolution kernels in the described n-th true convolutional layer carry out convolution to described (n-1)th convolution eigenwert, obtain the (n-1)th true convolution eigenwert;

Described (n-1)th true convolution eigenwert is input in the n-th pond layer and carries out pondization process, obtain the n-th convolution eigenwert.

6. character recognition method according to claim 3, it is characterized in that, described the loss function layer that described full connection features value is input to described convolutional neural networks is carried out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in described loss function layer converges to convergence threshold, otherwise repeat to carry out model training to described convolutional neural networks, comprising:

The loss function layer that described full connection features value is input to described convolutional neural networks is carried out loss function calculating, obtains loss function value, described loss function is the difference functions of classification value and label;

When described loss function value be less than last train the described loss function value obtained time, backpropagation is carried out to adjust the parameter of convolution kernel in described convolutional layer to described convolutional neural networks, and repeat to carry out model training to described convolutional neural networks, until the described loss function value obtained is less than loss function threshold value and restrains.

7. character recognition method according to claim 1, is characterized in that, described employing convolutional neural networks extracts feature to character image to be identified, obtains character features value, comprising:

The convolutional layer that described character image to be identified is input to described convolutional neural networks is processed, obtains convolution eigenwert;

The full articulamentum that described convolution eigenwert is input to described convolutional neural networks is processed, obtains full connection features value;

The function layer that described full connection features value is input to described convolutional neural networks is carried out function process, obtains described character features value.

8. character recognition method according to claim 7, is characterized in that, is describedly processed by the convolutional layer that described character image to be identified is input to described convolutional neural networks, obtains convolution eigenwert, comprising:

Described character image to be identified is input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert and processes, obtain the first convolution eigenwert;

Described (n-1)th convolution eigenwert is input in the n-th convolutional layer of described convolutional neural networks and processes, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

9. the character recognition method according to any one of claim 1-3, is characterized in that, described method also comprises:

To the operation that described character image to be identified is normalized, or to the operation that described character image sample to be identified is normalized.

10. a character recognition device, is characterized in that, comprising:

Extraction module, is configured to adopt convolutional neural networks to extract feature to character image to be identified, obtains character features value;

Sort module, is configured to adopt MQDF sorter to carry out classification process to the described character features value that described extraction module obtains, obtains the word that described character image to be identified is corresponding.

11. character recognition devices according to claim 10, is characterized in that, described device also comprises:

Training module, is configured to carry out model training to described convolutional neural networks.

12. character recognition devices according to claim 11, is characterized in that, described training module comprises:

First process submodule, the convolutional layer being configured to character image sample to be identified and label to be input to described convolutional neural networks processes, and obtains convolution eigenwert;

Second process submodule, the full articulamentum being configured to the described convolution eigenwert that described first process submodule obtains to be input to described convolutional neural networks processes, and obtains full connection features value;

3rd process submodule, the loss function layer being configured to the described full connection features value that described second process submodule obtains to be input to described convolutional neural networks carries out backpropagation process, adjust the parameter of convolution kernel in convolutional layer, until the loss function in described loss function layer converges to convergence threshold, otherwise repeat to carry out model training to described convolutional neural networks.

13. character recognition devices according to claim 12, is characterized in that, described first process submodule comprises:

First processing unit, is configured to described character image sample to be identified and described label to be input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert process, obtains the first convolution eigenwert;

N-th processing unit, be configured to described (n-1)th convolution eigenwert to be input in the n-th convolutional layer of described convolutional neural networks process, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

14. character recognition devices according to claim 13, is characterized in that, described n-th processing unit comprises:

Process of convolution subelement, be configured to the n-th true convolutional layer described (n-1)th convolution eigenwert being input to described n-th convolutional layer, multiple convolution kernels in described n-th true convolutional layer carry out convolution to described (n-1)th convolution eigenwert, obtain the (n-1)th true convolution eigenwert;

Pondization process subelement, the described (n-1)th true convolution eigenwert being configured to described process of convolution subelement to obtain is input in the n-th pond layer carries out pondization process, obtains the n-th convolution eigenwert.

15. character recognition devices according to claim 12, is characterized in that, described 3rd process submodule comprises:

Computing unit, the loss function layer be configured to described full connection features value is input to described convolutional neural networks carries out loss function calculating, and obtain loss function value, described loss function is the difference functions of classification value and label;

Adjustment unit, be configured to when the described loss function value that described computing unit calculates be less than last train the described loss function value obtained time, backpropagation is carried out to adjust the parameter of convolution kernel in described convolutional layer to described convolutional neural networks, and repeat to carry out model training to described convolutional neural networks, until the described loss function value obtained is less than loss function threshold value and restrains.

16. character recognition devices according to claim 10, is characterized in that, described extraction module comprises:

Process of convolution submodule, the convolutional layer be configured to described character image to be identified is input to described convolutional neural networks processes, and obtains convolution eigenwert;

Full connection handling submodule, the full articulamentum that the described convolution eigenwert being configured to be obtained by described process of convolution submodule is input to described convolutional neural networks processes, and obtains full connection features value;

Function process submodule, the function layer being configured to the described full connection features value that described full connection handling submodule obtains to be input to described convolutional neural networks carries out function process, obtains described character features value.

17. character recognition devices according to claim 16, is characterized in that, described process of convolution submodule comprises:

First processing unit, is configured to described character image to be identified to be input in the first volume lamination of described convolutional neural networks as the 0th convolution eigenwert process, obtains the first convolution eigenwert;

N-th processing unit, be configured to described (n-1)th convolution eigenwert to be input in the n-th convolutional layer of described convolutional neural networks process, obtain the n-th convolution eigenwert, n be greater than or equal to 2 integer, described n-th convolution eigenwert be described convolutional neural networks convolutional layer export described convolution eigenwert.

18. character recognition devices according to any one of claim 10-12, it is characterized in that, described device also comprises:

Normalization module, is configured to the operation be normalized described character image to be identified, or to the operation that described character image sample to be identified is normalized.

19. 1 kinds of character recognition devices, is characterized in that, comprising:

Processor;

For the storer of storage of processor executable instruction;

Wherein, described processor is configured to:

Adopt convolutional neural networks to extract feature to character image to be identified, obtain character features value;

Adopt MQDF sorter to carry out classification process to described character features value, obtain the word that described character image to be identified is corresponding.