CN112308058B - Method for recognizing handwritten characters - Google Patents

Abstract

The invention discloses a method for recognizing handwritten characters, which includes steps 1) making standard images corresponding to handwritten characters; step 2) building a U-shaped fully convolutional neural network, initializing network parameters, using handwritten character images as training data, and taking steps 1) Use the standard images of each character produced as the desired output to train the network; Step 3) Build a classifier, use the standard images of each character produced in Step 1) as training data, and train the classifier so that it can accurately classify the standard images; Step 4) Cascade U‑Net and classifier to complete the recognition of handwritten characters. The handwritten character recognition method provided by the present invention first processes the handwritten image into a standard printed image that is easy to recognize, and then the classifier classifies the standard printed image, which can effectively improve the accuracy of handwritten recognition.

Description

A method for recognizing handwritten characters

Technical field

The invention belongs to the field of computer vision, and particularly relates to a handwritten character recognition method, which is mainly used for optical character recognition.

Background technique

Handwritten character recognition is a basic research content in the field of computer vision. Its main task is to recognize text information obtained from media such as paper, photos, or touch screens. With the development of machine learning, most optical character recognition uses artificial neural network methods and has achieved good recognition results. However, compared with the recognition of standard printed characters, since handwritten characters vary from person to person and have many changes, they are difficult to be recognized correctly by machines, and some are even difficult to be recognized correctly by humans.

In order to improve the recognition accuracy of handwritten characters, the present invention proposes a recognition method that first processes complex and changeable handwritten images into standard printed images that are easy to recognize, and then uses a simple classifier to classify the standard printed images. This invention uses the network structure of U-Net to realize the operation of processing handwritten images into standard printed images. U-Net was proposed by Ronneberger in 2015 as a U-shaped full convolution containing multiple down-sampling and up-sampling operations. Neural networks are often used for semantic segmentation of images. The classifier does not need to learn a large number of handwritten images. It only needs to learn to correctly classify standard printed characters. Therefore, a simple classifier can be selected, which greatly reduces the model complexity of the classifier and improves the recognition speed of the classifier. Finally, by cascading U-Net with a classifier, efficient recognition of handwriting images is achieved.

Contents of the invention

The purpose of the present invention is to provide a handwritten character recognition method, thereby solving the current problem of insufficient accuracy in handwritten character recognition.

In order to achieve the above goals, the technical methods adopted are as follows:

Step 1), create a printed standard image corresponding to the handwritten characters. For different handwritten images of the same character, create a standard printed version of the corresponding character as a standard image, and then perform operations such as grayscale and binarization on the standard image to facilitate U-Net learning;

Step 2), build a U-shaped fully convolutional neural network, initialize the network parameters, and use the standard image of each character produced in step 1) as the desired output to train U-Net, so as to input different handwritten images of the same character into U-Net. Output the standard image of the character;

Step 3), build a simple classifier, use the single standard image of each character produced in step 1) as training data, and train the classifier so that it can accurately classify the standard image;

Step 4), cascade U-Net and classifier. The original image is input into U-Net to obtain the processed standard image, which is then sent to the classifier to obtain the classification result.

The handwritten character recognition method provided by the present invention first processes the handwritten image into a standard printed image that is easy to recognize, and then the classifier classifies the standard printed image, which can effectively improve the accuracy of handwritten recognition.

Description of the drawings

Figure 1 is a schematic diagram of making a standard image, that is, for different handwritten versions of the same character, a standard printed version is made as a standard image of the corresponding character;

Figure 2 is a schematic diagram of the U-shaped fully convolutional neural network structure;

Figure 3 is a schematic flow chart of handwritten character recognition according to the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the specific embodiments described here are only used to explain the present invention and do not limit the present invention.

According to an embodiment of the present invention, the handwriting samples are derived from the MNIST data set.

Step 1: Refer to the schematic diagram in Figure 1. For each type of character in the data set, make a standard printed image material of that type of character, grayscale it into a single-channel image, then binarize it, and finally process its image. Size so that it matches the image size in the dataset. Through the above operations, standard images of various types of characters are generated.

Step 2.1: Build a U-shaped fully convolutional neural network. Referring to the schematic diagram in Figure 2, the U-shaped fully convolutional neural network contains a total of 13 convolutional layers (Convolutional Layer), 2 down-sampling layers (Down-sampling Layer), 2 up-sampling layers (Up-sampling Layer), and Two cross-layer connections.

Except for the last convolution layer, the other convolution kernel sizes (Kernel Size) are 3×3, the sliding stride (Stride) is 1, the edge padding (Padding) is 1, and the activation function is ReLU. The formula is as follows:

f(x)=max(0,x)

Where x is the input value of the node, and f(x) is the output value of the node input after passing through the activation function.

Batch Normalization (Batch Normalization) is added between the convolution layer and the activation function. The formula is as follows:

where x is the result after convolution, μ and σ are the mean and variance of the data respectively, and γ and β are two learnable parameters that control the scaling and translation of the data respectively.

The downsampling layer adopts the Max Pooling method, the window size is 2×2, and the sliding step size is 2. The upsampling layer uses bilinear interpolation to amplify the feature map, with an amplification ratio of 2. Cross-layer connection uses concatenation to connect the feature maps of the front layer and the back layer.

Step 2.2: Train U-Net. Initialize the network parameters, send the handwritten image data as input to U-Net, and use the standard image produced in step 1 as the desired output. The loss function uses binary cross entropy (Binary CrossEntropy), and the formula is as follows:

in is the probability that the model predicts that the sample is a positive example, y is the sample label, if the sample is a positive example, the value is 1, otherwise the value is 0.

Step 3.1: Build a simple classifier. In this example, a Support Vector Machine (SVM) is used as the classifier, a radial basis function kernel (RBF kernel) is used, the regularization coefficient is 1, and the hyperparameter is 0.5.

Step 3.2: Train the SVM classifier. Only a single standard image of each character is used as the input of the SVM for training, so that the SVM can accurately classify the standard image of each character.

Step 4: Cascade U-Net with the classifier, see the schematic diagram in Figure 3. The handwritten image is sent to U-Net, which is processed into a standard image by U-Net, and then the standard image is sent to the classifier to obtain the classification result.

Claims (1)

1. A handwritten character recognition method, characterized by: including: Step 1), create a printed standard image corresponding to the handwritten characters; Step 2), build a U-shaped fully convolutional neural network, and initialize the network parameters; use handwritten character images as training data, and use the printed standard image corresponding to each character produced in step 1) as the expected output of the network; Step 3), build a classifier, and train the classifier so that it can accurately classify the standard images of each character produced in step 1); Step 4), cascade U-Net and classifier to complete the recognition of handwritten characters; In the described step 1), a printed standard image corresponding to the handwritten characters is produced, the printed standard image is grayscaled into a single-channel image, and then binarized, and finally its image size is processed to make it consistent with the handwritten characters. Character images are consistent; The described step 2) builds a U-Net network structure, using a U-shaped fully convolutional neural network as the basic network structure. The number of upsampling layers and downsampling layers in the network is consistent, so that the input data size is consistent with the output data size, so as to Handwritten character images are used as training data, and the printing standard body produced in step 1) is used as the expected output of the network. The loss function uses binary cross-entropy, and the formula is as follows: in is the probability that the model predicts that the sample is a positive example, y is the sample label, if the sample is a positive example, the value is 1, otherwise the value is 0; Described step 3) builds a classifier, and the classifier learns the single standard image of each character produced in step 1); The described step 4) cascades U-Net and classifier. U-Net and classifier are trained in steps 2) and 3). The parameters of both will no longer change. The handwritten character image first passes through U-Net. It is processed into a standard image, and then the classifier classifies the standard image to complete the recognition of handwritten characters; The U-shaped fully convolutional neural network contains a total of 13 convolutional layers, 2 downsampling layers, 2 upsampling layers, and two cross-layer connections; Except for the last convolution layer, the other convolution kernel sizes (Kernel Size) are all 3×3, the sliding step size is 1, the edge filling is 1, and the activation function is ReLU. The formula is as follows: f(x)＝max(0,x) where x is the input value of the node, and f(x) is the output value of the node input after passing through the activation function; Batch normalization is added between the convolutional layer and the activation function, and the formula is as follows: where x is the result after convolution, μ and σ are the mean and variance of the data respectively, and γ and β are two learnable parameters that control the scaling and translation of the data respectively.