CN116704519A - Character recognition method, character recognition device, electronic equipment and storage medium - Google Patents

Abstract

The invention discloses a character recognition method, a character recognition device, electronic equipment and a storage medium. The character recognition method comprises the following steps: the method comprises the steps of obtaining a character recognition model, wherein the character recognition model comprises a convolution layer, a circulation layer and a transcription layer, the convolution layer is formed by stacking depth-separable convolution blocks, and the convolution layer is used for extracting a characteristic sequence of an input image; the circulating layer is used for determining a predicted result of the input image according to the characteristic sequence and calculating the deviation between the predicted result and the true value to obtain a probability matrix; the transcription layer is used for outputting the recognition result of the input image according to the probability matrix; and inputting the image containing the characters to be identified into the character identification model to obtain a character identification result. According to the technical scheme, the probability matrix is calculated by using the light text recognition model, and the recognition result is output according to the probability matrix, so that the real-time performance of reasoning is improved, and the text recognition efficiency is improved.

Description

Character recognition method, device, electronic device and storage medium

technical field

Embodiments of the present invention relate to the field of image processing, and in particular, to a character recognition method, device, electronic equipment, and storage medium.

Background technique

Traditional Optical Character Recognition (OCR) technology often requires multiple processing of images, such as denoising, binarization, character segmentation, etc. These processing processes are not only time-consuming, but also easily affected by factors such as light and noise. affect the recognition accuracy. As the computing power of mobile chips continues to increase, it becomes possible to deploy some relatively complex deep learning models that require high computing power on the mobile end, and the demand for deep learning reasoning on mobile devices is also increasing.

The model used in real-time text recognition is a model with a large amount of parameters and a large amount of calculation. It usually contains hundreds of millions of parameters, and the size of the model ranges from tens or hundreds of megabytes, which will take up Larger memory resources and computing resources, this model is suitable for cloud service inference with Graphics Processing Unit (GPU), because it has stronger computing power support, and there is a response on the mobile terminal with limited computing resources The risk of long time or system crash, slow reasoning speed, and low efficiency of text recognition make it difficult to meet the real-time response requirements of mobile terminal deployment.

Contents of the invention

The invention provides a character recognition method, device, electronic equipment and storage medium to improve the real-time performance of reasoning and improve the efficiency of character recognition.

In a first aspect, an embodiment of the present invention provides a text recognition method, including:

Obtain a text recognition model, the text recognition model includes a convolutional layer, a loop layer and a transcription layer, the convolutional layer is formed by stacking separable convolutional blocks in depth, and the convolutional layer is used to extract the feature sequence of the input image The cyclic layer is used to determine the prediction result of the input image according to the feature sequence and calculate the deviation between the prediction result and the true value to obtain a probability matrix; the transcription layer is used to output the recognition result of the input image according to the probability matrix;

The image containing the character to be recognized is input to the character recognition model to obtain a character recognition result.

Optionally, each of the depthwise separable convolution blocks includes short-circuit connections;

If the depthwise separable convolution block is located in the downsampling layer, the short-circuit connection is a 1×1 convolution;

Otherwise, the short-circuit connection is the feature map input by the upper layer.

Optionally, the character recognition model is used based on a greedy retrieval algorithm. For each slice, the character corresponding to the highest probability is selected according to the probability matrix corresponding to the slice as the recognition result corresponding to the slice.

Optionally, input the image containing the text to be recognized into the text recognition model, including:

The images containing the characters to be recognized are detected segment by segment, and each time an image with a set length is detected, the currently detected image with a set length is input to the character recognition model.

Optionally, the text recognition model is also used for: if the output of the text recognition model for images with different set lengths includes overlapping areas, then weighting the corresponding positions in the corresponding probability matrix for the overlapping areas mean, to obtain the probability matrix of the coincident region.

Optionally, obtain the text recognition model, including:

Build a text recognition model and train the text recognition model;

Converting the text recognition model into an open standard file format and merging partial layers through simplified tools;

The recognition model is converted into an inference framework format using a conversion tool.

Optionally, the method further includes: scaling the image containing the text to be recognized to a preset size according to the principle of proportionality.

In the second aspect, the embodiment of the present invention also provides a character recognition device, including:

The acquisition module is used to acquire a text recognition model, the text recognition model includes a convolutional layer, a loop layer and a transcription layer, the convolutional layer is formed by stacking separable convolutional blocks in depth, and the convolutional layer is used to extract The feature sequence of the input image; the recurrent layer is used to determine the predicted result of the input image according to the feature sequence and calculate the deviation between the predicted result and the true value to obtain a probability matrix; the transcription layer is used to output the input according to the probability matrix Image recognition results;

The recognition module is used to input the image containing the character to be recognized into the character recognition model to obtain the character recognition result.

In a third aspect, an embodiment of the present invention provides an electronic device, including:

one or more processors;

storage means for storing one or more programs;

When the one or more programs are executed by the one or more processors, the one or more processors are made to implement the character recognition method as described in the first aspect.

In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the character recognition method as described in the first aspect is implemented.

The embodiment of the present invention provides a character recognition method, device, electronic equipment and storage medium. The text recognition method includes: obtaining a text recognition model, the text recognition model includes a convolutional layer, a loop layer and a transcription layer, the convolutional layer is formed by stacking separable convolutional blocks in depth, and the convolutional layer is used for Extracting the feature sequence of the input image; the recurrent layer is used to determine the prediction result of the input image according to the feature sequence and calculate the deviation between the prediction result and the true value to obtain a probability matrix; the transcription layer is used to output according to the probability matrix Inputting the recognition result of the image; inputting the image containing the character to be recognized into the character recognition model to obtain the character recognition result. In the above technical solution, a lightweight character recognition model is used to calculate a probability matrix and output a recognition result according to the probability matrix, which improves the real-time performance of reasoning and improves the efficiency of character recognition.

Description of drawings

The above and other features, advantages and aspects of the various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic and that elements and elements are not necessarily drawn to scale.

FIG. 1 is a flow chart of a character recognition method provided by Embodiment 1 of the present invention;

Fig. 2 is a schematic diagram of a short-circuit connection provided by an embodiment;

FIG. 3 is a flow chart of a character recognition method provided in Embodiment 2 of the present invention;

FIG. 4 is a schematic diagram of an output including overlapping regions provided by an embodiment;

Fig. 5 is a schematic diagram of a probability matrix for determining overlapping regions provided by an embodiment;

FIG. 6 is a schematic structural diagram of a character recognition device provided in Embodiment 3 of the present invention;

FIG. 7 is a schematic structural diagram of an electronic device provided by Embodiment 4 of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, but not to limit the present invention. In addition, the embodiments and the features in the embodiments of the present invention can be combined with each other under the condition of no conflict. In addition, it should be noted that, for the convenience of description, only some structures related to the present invention are shown in the drawings but not all structures.

Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processing, many of the steps may be performed in parallel, concurrently, or simultaneously. Additionally, the order of steps may be rearranged. The process may be terminated when its operations are complete, but may also have additional steps not included in the figure. The processing may correspond to a method, function, procedure, subroutine, subroutine, or the like.

Embodiment one

FIG. 1 is a flow chart of a character recognition method provided by Embodiment 1 of the present invention. This embodiment is applicable to character recognition, for example, it can be applied to mobile devices such as scanning pens that can automatically collect data. Specifically, the character recognition method can be executed by a character recognition device, and the character recognition device can be realized by means of software and/or hardware, and integrated into an electronic device. Further, the electronic device may refer to a desktop computer, a server, a notebook computer, or a scanning device, etc.

As shown in Figure 1, the method specifically includes the following steps:

S110. Acquire a character recognition model.

Among them, the text recognition model mainly refers to a pre-trained deep learning model, including a convolutional layer, a recurrent layer, and a transcription layer. The convolutional layer is formed by stacking depth-separable convolutional blocks and is used to extract the feature sequence of the input image; the recurrent layer The layer is used to determine the prediction result of the input image according to the feature sequence and calculate the deviation between the prediction result and the true value to obtain a probability matrix; the transcription layer is used to output the recognition result of the input image according to the probability matrix.

In this embodiment, a lightweight character recognition model is constructed using depthwise separable convolution. Depthwise separable convolution is a lightweight convolution operation that can reduce the number of model parameters and calculations. It is mainly composed of channel-by-channel convolution (Depthwise Convolution) and pointwise convolution (Pointwise Convolution), which can maintain While improving the performance of the model, it can significantly reduce the use of computing and storage resources, and can be applied to mobile devices with limited computing and storage resources.

The backbone network (Backbone) of the text recognition model can be stacked by depth separable convolutional blocks (Block) to achieve the purpose of feature extraction. Exemplarily, for an input image whose input width and height channels are (100, 32, 3) dimensions, after the convolutional layer is down-sampled, the image width and height are respectively 1/8 and 1/32 of the original, and the obtained features The feature sequence of (w/8, Batch, 512) is obtained after the image is serialized, where w represents the width of the original image, and Batch represents the number of batches, and its value is 512. The obtained feature sequence is sent to the recurrent layer, and finally the result of (w/4, batch, class_num) is output. This result uses a loss function (such as Connectionist Temporal Classification Loss, referred to as CTC Loss) to match the image recognition result with the true value and The difference between the predicted result and the true value is calculated by using the forward and backward derivation methods, and the obtained loss (Loss) can be further used to backpropagate the parameters of the training model.

In this embodiment, the text recognition model can be implemented based on the TNN reasoning framework. TNN is a lightweight cross-platform deep learning reasoning framework. It uses a highly optimized computing engine to achieve efficient model reasoning and supports central processing unit (Central Processing Unit, CPU), graphics processing unit (Graphics Processing Unit, GPU) and artificial intelligence (AI) accelerators and other hardware platforms, and has low memory footprint and computing resource consumption; in addition, TNN supports a variety of deep learning frameworks and model formats, enabling rapid model deployment and service development and management; with small code size and binary size, it is suitable for use in resource-constrained environments such as mobile devices and embedded systems; and supports multiple operating systems and programming languages, and can be implemented on different hardware platforms and system environments model reasoning.

S120. Input an image containing a character to be recognized into the character recognition model to obtain a character recognition result.

Specifically, after the character recognition model is trained and tested, it has a reliable ability to recognize characters. On this basis, the image containing the character to be recognized is input to the character recognition model to obtain a character recognition result.

Optionally, the method further includes: scaling the image containing the text to be recognized to a preset size according to the principle of proportionality. Exemplarily, the width and height of the detected image may be stretched according to the principle of equal proportion, and the image may be scaled to a size of (32, w, 3).

In practical applications, text recognition models can be deployed in hardware with scanning capabilities. The final function of recognition can be completed by multiple models, which can involve the detection model, which can also be built by a similar Backbone. Together with three lightweight output heads, it can effectively speed up the reasoning speed of the model.

The character recognition method in this embodiment can improve the speed of real-time reasoning from the perspective of model optimization, and further optimize the reasoning speed by using a reasoning framework suitable for mobile terminals, so as to achieve the purpose of real-time reasoning. Traditional OCR technology often requires multiple processing of images, such as denoising, binarization, character segmentation, etc. These processing processes are not only time-consuming, but also easily affected by factors such as light and noise, thereby affecting the recognition accuracy. However, the text recognition model in this embodiment can use the deep learning principle to automatically learn the features of the image, and directly output the recognition result, avoiding multiple steps in the traditional OCR process, having higher recognition accuracy and faster speed, and improving It not only improves the efficiency and accuracy of text recognition, but also meets the real-time response requirements of mobile terminal deployment. In addition, the real-time text recognition technology based on deep vision can also support real-time recognition, multi-language recognition and other functions to provide users with a better experience.

Optionally, each depth-separable convolution block contains a short-circuit connection; if the depth-separable convolution block is located in the downsampling layer, the short-circuit connection is a 1×1 convolution; otherwise, the short-circuit connection is the feature map of the upper layer input .

Specifically, each Block not only contains depthwise separable convolution modules, but also contains short-circuit connections, and the way of short-circuit connections can be determined according to whether the Block is a downsampling layer. Exemplarily, FIG. 2 is a schematic diagram of a short-circuit connection provided by an embodiment. As shown on the right side of Figure 2, when the layer where the Block is located is the downsampling layer, the short-circuit connection is 1×1 (also can be expressed as 1*1) convolution; as shown on the left side of Figure 2, the short-circuit connection is the upper layer The input feature map. Adding a short-circuit connection branch in each block can not only make up for the gradient disappearance and gradient explosion problems that may be caused by the deep network, but also solve the problem of insufficient feature extraction by deep separable convolution.

Optionally, the character recognition model is used based on a greedy retrieval algorithm. For each slice, the character corresponding to the highest probability is selected according to the probability matrix corresponding to the slice as the recognition result corresponding to the slice. In this embodiment, the predicted output of the text recognition model during the recognition process is a probability matrix. Different from the training process, this process can use a greedy search method to obtain the highest probability of each slice (Slice) as the prediction result.

Embodiment two

FIG. 3 is a flow chart of a character recognition method provided by Embodiment 2 of the present invention. This embodiment embodies the character recognition process on the basis of the above embodiments. It should be noted that for technical details not exhaustively described in this embodiment, reference may be made to any of the foregoing embodiments.

Specifically, as shown in Figure 3, the method specifically includes the following steps:

S210. Construct a character recognition model and train the character recognition model.

S220, converting the character recognition model into an open standard file format and merging partial layers through a simplified tool.

In this embodiment, the trained text recognition model can be exported as an open standard file format (Open Neural Network Exchange, ONNX) using the Python machine learning library (such as Pytorch); in addition, some layers can be merged using simplified tools, and It can be understood as removing redundant layers. Specifically, the text recognition model can be built according to a model structure similar to CRNN. The redundant layer is, for example, the middle 3*3 convolutional layer. These layers will increase the model’s performance to a certain extent. The amount of parameters affects the inference speed of the model. After testing, removing these layers will not significantly reduce the accuracy of the model.

S230. Use a conversion tool to convert the recognition model into an inference framework format.

Specifically, the converted text recognition model in ONNX format can be converted into the inference framework (TNN) format using the TNN conversion tool, which helps the inference framework to self-load the model and do further quantitative work.

S240. Detect the image containing the text to be recognized segment by segment.

S250. Is an image with a set length detected? If yes, execute S260, otherwise return to execute S240.

S260. Input the currently detected image with a set length into the character recognition model to obtain a corresponding character recognition result.

S270. Does the output of the text recognition model for images with different set lengths include overlapping regions? If yes, execute S280, otherwise return to execute S240.

S280. For the overlapping area, weight and average each corresponding position in the corresponding probability matrix to obtain the probability matrix of the overlapping area.

In this embodiment, considering the real-time display of recognition results, the whole process can be carried out in a real-time splicing manner. The images input to the character recognition model are detected segment by segment, and each time they are spliced to a certain length, they are sent to the character recognition model for recognition. , the length of an image sent to the text recognition model each time can be the same or different. On this basis, the process screenshots of detection and recognition can be performed simultaneously. Moreover, the output obtained by the text recognition model for each input image may have overlapping areas. For the overlapping area, the corresponding positions in the corresponding probability matrix can be weighted and averaged to obtain the final probability matrix of the overlapping area.

S290. Obtain an identification result of the overlapping area according to the probability matrix.

Specifically, according to the final probability matrix, the character corresponding to the maximum probability can be used as the recognition result corresponding to the slice

Fig. 4 is a schematic diagram of an output including overlapping regions provided by an embodiment. Since a character may be divided into two parts, the image sent to the text recognition model may contain the second half of the image sent to the text recognition model last time. After the text recognition model is recognized, between two adjacent results There may be some character repetition. In this embodiment, redundant parts may be removed by merging character strings. As shown in Figure 4, each small box in this picture represents the predicted probability matrix of each slice after each picture has been recognized. This slice corresponds to the original input image with a width of 8 and a height of 32 pixels. The upper part is the prediction result of the previous image, and the lower frame sequence is the prediction result of the next image. It can be seen that the adjacent two images overlap. For the overlapping part, this problem is solved by weighting and averaging the corresponding positions of the probability matrix of the overlapping area.

Fig. 5 is a schematic diagram of a probability matrix for determining overlapping regions provided by an embodiment. As shown in Figure 5, the last Slice of img1 is at the center of img2, and the first Slice of img2 is the center of img1. These two positions are the easiest to predict and make mistakes, but the center position of each other will be greatly reduced. Risk of error. Therefore, the first Slice of img2 selects the prediction result corresponding to the center of img1, and the last Slice result of img1 selects the prediction result of the middle Slice in the lower part. In the actual splicing process, KEYI selects according to the splicing effect.

A text recognition method provided by Embodiment 2 of the present invention uses a lightweight framework to build a text recognition model for text recognition, and can recognize up to 10,000 characters (including Chinese and English). Through the transformation of the model and the deployment of the mobile reasoning framework TNN, the precise text recognition task can be completed under the condition that the accuracy of the model remains unchanged; the text recognition model can be deployed on mobile devices with limited resources without the assistance of cloud servers. It can be used offline to achieve real-time recognition, and can be combined with the mobile terminal reasoning framework to optimize memory and operators. It can be adapted to various low-computing power development boards. Even with a CPU with limited resources, it can achieve real-time results. Even if you don’t The function of real-time display can also be completed when quantization is used. In order to improve the user experience, the mobile terminal device can use the image segment or block method for recognition reasoning, which can timely feedback the recognition results, and efficiently and accurately combine the recognition results to effectively improve the correctness of the recognition stitching results. In addition, the text recognition model also supports real-time recognition, multi-language recognition and other functions. The text recognition method adopts the text recognition model based on the Torch framework design, and the processing of the prediction results does not use the method of searching for the path with the highest probability after transcription in the CTC Loss decoding process as the recognition result, but uses the greedy retrieval method. The Slice probability matrix only takes the character corresponding to the maximum probability as the recognition result of the Slice, thus effectively improving the retrieval efficiency and reasoning speed. In addition, this method can identify images of variable length in time and complete the splicing of the results, which can be fed back to the display screen in time, which can not only reduce the error rate of model reasoning, but also improve the user experience.

Embodiment three

FIG. 6 is a schematic structural diagram of a character recognition device provided by Embodiment 3 of the present invention. The text recognition device provided in this embodiment includes:

The obtaining module 310 is used to obtain a text recognition model, the text recognition model includes a convolutional layer, a loop layer and a transcription layer, the convolutional layer is formed by stacking separable convolutional blocks in depth, and the convolutional layer is used for Extracting the feature sequence of the input image; the recurrent layer is used to determine the prediction result of the input image according to the feature sequence and calculate the deviation between the prediction result and the true value to obtain a probability matrix; the transcription layer is used to output according to the probability matrix The recognition result of the input image;

The recognition module 320 is configured to input an image containing a character to be recognized into the character recognition model to obtain a character recognition result.

A character recognition device provided in Embodiment 3 of the present invention obtains a character recognition model through an acquisition module. The character recognition model includes a convolution layer, a loop layer, and a transcription layer, and the convolution layer is stacked by depth-separable convolution blocks. Formed, the convolution layer is used to extract the feature sequence of the input image; the loop layer is used to determine the prediction result of the input image according to the feature sequence and calculate the deviation between the prediction result and the true value to obtain a probability matrix; The transcription layer is used to output the recognition result of the input image according to the probability matrix; the recognition module inputs the image containing the character to be recognized to the character recognition model to obtain the character recognition result. On this basis, the lightweight text recognition model is used to calculate the probability matrix and output the recognition results according to the probability matrix, which improves the real-time reasoning and text recognition efficiency.

On the basis of the above embodiments, each of the depthwise separable convolution blocks includes a short-circuit connection;

If the depthwise separable convolution block is located in the downsampling layer, the short-circuit connection is a 1×1 convolution;

Otherwise, the short-circuit connection is the feature map input by the upper layer.

On the basis of the above embodiments, the character recognition model is used based on a greedy search algorithm. For each slice, the character corresponding to the highest probability is selected according to the probability matrix corresponding to the slice as the recognition result corresponding to the slice.

On the basis of the foregoing embodiments, the recognition module includes:

The input unit is used to detect the image containing the character to be recognized segment by segment, and input the currently detected image of the set length to the character recognition model whenever an image of the set length is detected.

On the basis of the above-mentioned embodiments, the character recognition model is further configured to: if the output of the character recognition model for images with different set lengths includes overlapping areas, then for the overlapping areas, each of the corresponding probability matrices The weighted mean value corresponding to the position is obtained to obtain the probability matrix of the overlapping area.

On the basis of the foregoing embodiments, the acquisition module 310 includes:

A construction unit for constructing a text recognition model and training the text recognition model;

A simplification unit is used for converting the character recognition model into an open standard file format and merging partial layers through a simplification tool;

A format conversion unit, configured to use a conversion tool to convert the recognition model into an inference framework format.

On the basis of the foregoing embodiments, the device also includes:

The scaling module is used to scale the image containing the text to be recognized to a preset size according to the principle of equal proportion.

The character recognition device provided in the third embodiment of the present invention can be used to execute the character recognition method provided in any of the above embodiments, and has corresponding functions and beneficial effects.

Embodiment four

FIG. 7 shows a schematic structural diagram of an electronic device that can be used to implement an embodiment of the present invention.

As shown in FIG. 7 , an electronic device provided in this embodiment includes: a processor 410 and a storage device 420 . There may be one or more processors in the electronic device. A processor 410 is taken as an example in FIG. Take connection via bus as an example.

The one or more programs are executed by the one or more processors 410, so that the one or more processors implement any of the character recognition methods described in the foregoing embodiments.

The storage device 420 in the electronic device is used as a computer-readable storage medium, and can be used to store one or more programs, and the programs can be software programs, computer-executable programs and modules, such as the character recognition method in the embodiment of the present invention Corresponding program instructions/modules (for example, modules in the character recognition device shown in FIG. 6 , including: acquisition module 310 and recognition module 320). The processor 410 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the storage device 420 , that is, implements the character recognition method in the above method embodiments.

The storage device 420 mainly includes a program storage area and a data storage area, wherein the program storage area can store an operating system and an application program required by at least one function; the data storage area can store data created according to the use of the electronic device (as described above) The feature sequence, probability matrix, character recognition results, etc. in the embodiment). In addition, the storage device 420 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage devices. In some examples, the storage device 420 may further include memories that are remotely located relative to the processor 410, and these remote memories may be connected to a server through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

Moreover, when one or more programs included in the above-mentioned electronic device are executed by the one or more processors 410, the program performs the following operations: obtain a character recognition model, and the character recognition model includes a convolution layer, a loop layer and A transcription layer, the convolutional layer is stacked by depth-separable convolutional blocks, the convolutional layer is used to extract the feature sequence of the input image; the loop layer is used to determine the prediction result of the input image according to the feature sequence And calculate the deviation between the predicted result and the true value to obtain a probability matrix; the transcription layer is used to output the recognition result of the input image according to the probability matrix; input the image containing the text to be recognized to the text recognition model to obtain text recognition result.

The electronic device further includes: a communication device 430 , an input device 440 and an output device 450 .

The processor 410, the memory 420, the communication device 430, the input device 440 and the output device 450 in the electronic device may be connected through a bus or in other ways. In FIG. 4, connection through a bus is taken as an example.

The input device 440 can be used to receive input numbers or character information, and generate key signal input related to user settings and function control of the electronic device. The output device 450 may include a display device such as a display screen.

Communications device 430 may include a receiver and a transmitter. The communication device 430 is configured to perform information sending and receiving communication according to the control of the processor 410 .

The electronic device proposed in this embodiment and the character recognition method proposed in the above embodiment belong to the same inventive concept. The technical details not described in detail in this embodiment can be referred to any of the above embodiments, and this embodiment has the same features as the character recognition method. beneficial effect.

On the basis of the above-mentioned embodiments, this embodiment also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a data storage device, the robot scheduling method in any of the above-mentioned embodiments of the present invention is implemented. The method includes: obtaining a text recognition model, the text recognition model includes a convolutional layer, a loop layer and a transcription layer, the convolutional layer is formed by stacking depth-separable convolutional blocks, and the convolutional layer is used to extract an input image The feature sequence; the recurrent layer is used to determine the prediction result of the input image according to the feature sequence and calculate the deviation between the predicted result and the true value to obtain a probability matrix; the transcription layer is used to output the input image according to the probability matrix Recognition result: input the image containing the character to be recognized into the character recognition model to obtain the character recognition result.

Of course, a storage medium containing computer-executable instructions provided by an embodiment of the present invention, the computer-executable instructions are not limited to the operation of the data storage method as described above, and can also execute the data storage method provided by any embodiment of the present invention Relevant operations in , and have corresponding functions and beneficial effects.

Through the above description about the implementation mode, those skilled in the art can clearly understand that the present invention can be realized by means of software and necessary general-purpose hardware, and of course it can also be realized by hardware, but in many cases the former is a better implementation mode . Based on this understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a floppy disk of a computer , read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disc, etc., including several instructions to make a computer device (which can be a personal computer, A server, or a network device, etc.) executes the data storage method described in each embodiment of the present invention.

Note that the above are only preferred embodiments of the present invention and applied technical principles. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, rearrangements and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and can also include more other equivalent embodiments without departing from the concept of the present invention, and the present invention The scope is determined by the scope of the appended claims.

Claims (10)

1. A method of text recognition, comprising:

the method comprises the steps of obtaining a character recognition model, wherein the character recognition model comprises a convolution layer, a circulation layer and a transcription layer, the convolution layer is formed by stacking depth-separable convolution blocks, and the convolution layer is used for extracting a characteristic sequence of an input image; the circulating layer is used for determining a predicted result of the input image according to the characteristic sequence and calculating the deviation between the predicted result and the true value to obtain a probability matrix; the transcription layer is used for outputting the recognition result of the input image according to the probability matrix;

and inputting the image containing the characters to be identified into the character identification model to obtain a character identification result.

2. The method of claim 1, wherein each of the depth separable convolution blocks includes a shorting connection therein;

if the depth separable convolution block is located in the downsampling layer, the short circuit connection is a convolution of 1×1;

otherwise, the short circuit connection is a characteristic diagram input by the upper layer.

3. The method of claim 1, wherein the word recognition model is configured to take, for each slice, a character corresponding to a maximum probability according to a probability matrix corresponding to the slice as a recognition result corresponding to the slice based on a greedy search algorithm.

4. The method of claim 1, wherein inputting an image containing text to be recognized into the text recognition model comprises:

detecting images containing characters to be recognized segment by segment, and inputting the currently detected images with the set length into the character recognition model every time the images with the set length are detected.

5. The method of claim 4, wherein the word recognition model is further configured to:

and if the text recognition model outputs images with different set lengths and comprises a superposition area, weighting and averaging the corresponding positions in the corresponding probability matrix of the superposition area to obtain the probability matrix of the superposition area.

6. The method of claim 1, wherein obtaining a word recognition model comprises:

constructing a character recognition model and training the character recognition model;

converting the character recognition model into an open standard file format and merging partial layers through a simplifying tool;

the recognition model is converted to an inference framework format using a conversion tool.

7. The method as recited in claim 1, further comprising:

and scaling the image containing the text to be recognized to a preset size according to the principle of equal proportion.

8. A character recognition device, comprising:

the device comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring a character recognition model, the character recognition model comprises a convolution layer, a circulation layer and a transcription layer, the convolution layer is formed by stacking depth separable convolution blocks, and the convolution layer is used for extracting a characteristic sequence of an input image; the circulating layer is used for determining a predicted result of the input image according to the characteristic sequence and calculating the deviation between the predicted result and the true value to obtain a probability matrix; the transcription layer is used for outputting the recognition result of the input image according to the probability matrix;

and the recognition module is used for inputting the image containing the characters to be recognized into the character recognition model to obtain a character recognition result.

9. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,,

the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the word recognition method of any one of claims 1-7.

10. A computer-readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements a text recognition method as claimed in any one of claims 1-7.