WO2021139175A1 - Electric power operation ticket character recognition method based on convolutional neural network - Google Patents
Electric power operation ticket character recognition method based on convolutional neural network Download PDFInfo
- Publication number
- WO2021139175A1 WO2021139175A1 PCT/CN2020/111550 CN2020111550W WO2021139175A1 WO 2021139175 A1 WO2021139175 A1 WO 2021139175A1 CN 2020111550 W CN2020111550 W CN 2020111550W WO 2021139175 A1 WO2021139175 A1 WO 2021139175A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- neural network
- convolutional neural
- feature
- handwriting
- power operation
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
Definitions
- the invention relates to a character recognition method, in particular to a power operation ticket character recognition method based on a convolutional neural network.
- the traditional text recognition method based on convolutional neural network directly uses the CNN method to train the handwritten Chinese character picture sample set, and then obtains the text classification model.
- This type of method only uses the CNN method to learn the image spatial feature information, and the training is efficient and the design is simple.
- the CNN method has limited samples in the training process and cannot learn all possible handwritten font features. Its network structure is simple, the image representation ability is limited, and training is easy to overfit. The recognition performance of such methods needs to be further improved.
- a text recognition method based on improved convolutional neural networks has also appeared, using image deformation network GTN, affine deformation AD, elastic deformation ED and other text deformation methods to enrich the handwritten font sample set and overcome the sample Limitations:
- Use fractional pooling method FMP training method DropSample, relaxed convolutional neural network R-CNN and other methods to optimize the performance of convolutional neural network, and overcome the problems of training over-fitting and simple spatial feature expression.
- This type of method only considers sample set amplification, network structure optimization, etc., and ignores the deeper and fine-grained handwriting features of handwritten fonts, such as handwriting direction changes, which makes the accuracy of power operation ticket image text recognition low.
- the technical problem to be solved and the technical task proposed by the present invention are to perfect and improve the existing technical scheme, and provide a method for recognizing power operation ticket text based on convolutional neural network, so as to improve the accuracy of text recognition.
- the present invention adopts the following technical solutions.
- a method for recognizing power operation ticket characters based on convolutional neural network which is characterized in that it comprises the following steps:
- step 1) for a sample image p i (p i ⁇ c, 1 ⁇ i ⁇ N ) to give the training set, where N is the total number of training set sample image (c) includes the acquiring of M ⁇ p i The M numerical matrix A i and the corresponding clear image numerical matrix B i .
- step 2 construct a three-layer convolutional neural network model C 0 in step 2), which only contains the convolutional layer, select the activation function ReLU, set the step size to 1, and do not fill the convolution operation with 0, the network structure is:
- step 3 define the loss function in step 3).
- the purpose of the loss function is to obtain the minimum F norm.
- the calculation formula is as follows:
- ⁇ ⁇ W j ,b j ⁇ ;
- W j ⁇ W j k :1 ⁇ k ⁇ n j ⁇ , W j k is the convolution matrix of the jth layer of the convolutional network, b j is the deviation value, n j is the number of convolution kernels in the jth layer of the convolutional network.
- step 5 is non-linear mapping function F ⁇ (p), c in the training set is calculated for each sample image to enhance the image corresponding to p i of P i, to give a new training set C (P i ⁇ C,: as a preferred technology 1 ⁇ i ⁇ N).
- Step 6 includes the steps:
- the imaginary stroke feature extracts the degree of change in the direction of the handwriting. For each adjacent pixel, the calculation formula is as follows:
- the path signature feature extracts the handwriting curvature value, the starting and ending interval of the given handwriting is [s, t], and the curvature feature is defined as follows:
- the 8-direction feature can perfectly fit the strokes of Chinese characters, such as horizontal and vertical strokes, and can be used to extract handwriting direction information; assuming a two-dimensional coordinate, the 8-direction features are from 0°, 45°, 90°, 135°, 180°225°, 270°, 315° calculate the handwriting gradient size; given the starting and ending coordinates (x 1 , y 1 ) and (x 2 , y 2 ) of a handwriting, the gradient calculation formula is as follows:
- the size of the convolution kernel of the first layer of the convolutional network of the integrated convolutional neural network model C 1 is set to 3 ⁇ 3, the number of convolution kernels is 80, and the number is increased by 80 in sequence;
- the size of the convolution kernel of the second to sixth layers of the convolutional network is set to 2 ⁇ 2; select the activation function ReLU, select the zero-complement convolution operation, and the step size is 1; the size of the pooling layer matrix is 2 ⁇ 2; integrated convolution
- the structure of neural network model C 1 is:
- n the number of handwriting feature dimensions
- the integrated convolutional neural network model is used to train the power operation ticket text recognition model, which effectively improves the accuracy of text recognition.
- FIG. 1 is a flowchart of the present invention.
- the present invention includes the steps:
- Step 1 for a sample image p i (p i ⁇ c, 1 ⁇ i ⁇ N ), where N is the total number of samples of training images comprises a set of c, obtaining M ⁇ M matrix of values of A i p i, and the corresponding clear Image value matrix B i .
- Step 2 Construct a three-layer convolutional neural network model C 0 with only the convolutional layer, select the activation function ReLU, set the step size to 1, and do not fill the convolution operation with 0.
- the network structure is shown in Table 1:
- Step 3 Define the loss function. Compared with B i , the purpose of the loss function is to obtain the minimum F norm.
- the calculation formula is as follows:
- Step 4 Traverse all the sample images in the training set c, and train the output image enhancement calculation function, that is, the non-linear mapping function F ⁇ (p).
- Step 1 using a nonlinear mapping function F ⁇ (p), calculated in the training set c p i enhanced image corresponding to each sample P i, to give a new training set C (P i ⁇ C, 1 ⁇ i ⁇ N) .
- the imaginary stroke feature extracts the degree of change in the direction of the handwriting. For each adjacent pixel, the calculation formula is as follows:
- Different pixel value comparison dcd the calculated virtual matrix P i of the stroke.
- the path signature feature extracts the handwriting curvature value.
- the starting and ending interval of the given handwriting is [s, t], and the curvature feature is defined as follows:
- the k value can take any value, and it should not be too large under normal circumstances, otherwise it will cause the computational complexity to increase exponentially, but it will not be possible to obtain more effective handwriting features.
- the path P i calculated signature matrix For each adjacent pixel, the path P i calculated signature matrix.
- the 8-direction feature can perfectly fit the strokes such as horizontal and vertical strokes of Chinese characters, and can be used to extract handwriting direction information. Assuming that a two-dimensional coordinate is given, the 8-direction features are calculated from 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° respectively. Given the starting and ending coordinates (x 1 , y 1 ) and (x 2 , y 2 ) of a handwriting, the gradient calculation formula is as follows:
- Step 3 Construct an integrated convolutional network model C 1 , which contains a 6-layer convolutional network.
- the next layer of the first 5 layers of convolutional network is configured with a pooling layer, and the next layer of the 6th layer of convolutional network is configured with a fully connected layer .
- the size of the convolution kernel of the first layer of convolutional network is set to 3 ⁇ 3, the number of convolution kernels is 80, and the number of convolution kernels is increased by 80; the size of the convolution kernel of the second to sixth layers of convolutional network is set to 2 ⁇ 2; choose The activation function ReLU is selected to complement 0 convolution operation, and the step size is 1.
- the size of the pooling layer matrix is 2 ⁇ 2.
- the model structure is shown in Table 2, where n represents the number of handwriting feature dimensions.
- Step 4 Traverse all the sample images in the training set C, combine the imaginary stroke features, path signature features and 8-direction features to train and output a text recognition model.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Character Discrimination (AREA)
Abstract
Disclosed is an electric power operation ticket character recognition method based on a convolutional neural network, relating to a character recognition method. At present, character recognition of an electric power operation ticket is unclear. The method in the present invention comprises the steps of: constructing a convolutional neural network model with only three convolutional layers, no pooling layer and no fully connected layer, performing training to obtain a nonlinear mapping function, and improving a peak signal-to-noise ratio of an image; using a handwriting feature calculation method to respectively obtain imaginary stroke features, path signature features and 8-directional features of image characters of an electric power operation ticket; and constructing an integrated convolutional neural network model with six convolutional layers, five pooling layers and one fully connected layer, and in combination with the imaginary stroke features, the path signature features and the 8-directional features, performing training to obtain a character recognition model. According to the present technical solution, the excellent spatial feature learning ability of a convolutional neural network is used, and an image enhancement method based on the convolutional neural network and a character recognition method based on the convolutional neural network are used, such that the accuracy of image character recognition of an electric power operation ticket is improved.
Description
本发明涉及一种文字识别方法,尤其涉及一种基于卷积神经网络的电力操作票文字识别方法。The invention relates to a character recognition method, in particular to a power operation ticket character recognition method based on a convolutional neural network.
传统基于卷积神经网络的文字识别方法直接使用CNN方法训练手写汉字图片样本集,进而得到文字分类模型。此类方法仅使用CNN方法学习图像空间特征信息,训练高效、设计简单。但是,CNN方法在训练的过程中样本有限,无法学习所有可能存在的手写字体特征,其网络结构简单,图像表示能力有限,训练易过拟合。此类方法的识别性能有待进一步提高。The traditional text recognition method based on convolutional neural network directly uses the CNN method to train the handwritten Chinese character picture sample set, and then obtains the text classification model. This type of method only uses the CNN method to learn the image spatial feature information, and the training is efficient and the design is simple. However, the CNN method has limited samples in the training process and cannot learn all possible handwritten font features. Its network structure is simple, the image representation ability is limited, and training is easy to overfit. The recognition performance of such methods needs to be further improved.
为进一步提高文字识别的准确度,目前也出现了基于改进卷积神经网络的文字识别方法,使用图像变形网络GTN、仿射变形AD、弹性变形ED等文字变形方法丰富手写字体样本集,克服样本局限性;使用分数池化方法FMP、训练方法DropSample、松弛卷积神经网络R-CNN等方法优化卷积神经网络性能,克服训练过拟合、空间特征表达简单等问题。此类方法仅考虑样本集扩增、网络结构优化等方面,忽略了手写字体更深层次、细粒度的笔迹特征,如笔迹方向变化等,使得电力操作票图像文字识别的准确度低。In order to further improve the accuracy of text recognition, a text recognition method based on improved convolutional neural networks has also appeared, using image deformation network GTN, affine deformation AD, elastic deformation ED and other text deformation methods to enrich the handwritten font sample set and overcome the sample Limitations: Use fractional pooling method FMP, training method DropSample, relaxed convolutional neural network R-CNN and other methods to optimize the performance of convolutional neural network, and overcome the problems of training over-fitting and simple spatial feature expression. This type of method only considers sample set amplification, network structure optimization, etc., and ignores the deeper and fine-grained handwriting features of handwritten fonts, such as handwriting direction changes, which makes the accuracy of power operation ticket image text recognition low.
发明内容Summary of the invention
本发明要解决的技术问题和提出的技术任务是对现有技术方案进行完善与改进,提供一种基于卷积神经网络的电力操作票文字识别方法,以提高文字识别准确度目的。为此,本发明采取以下技术方案。The technical problem to be solved and the technical task proposed by the present invention are to perfect and improve the existing technical scheme, and provide a method for recognizing power operation ticket text based on convolutional neural network, so as to improve the accuracy of text recognition. To this end, the present invention adopts the following technical solutions.
一种基于卷积神经网络的电力操作票文字识别方法,其特征在于包括以下步骤:A method for recognizing power operation ticket characters based on convolutional neural network, which is characterized in that it comprises the following steps:
1)获取样本图像,得到训练集;1) Obtain sample images and get the training set;
2)构建仅具有3层卷积层,无池化层,无全连接层的卷积神经网络模型C
0;
2) Construct a convolutional neural network model C 0 with only 3 convolutional layers, no pooling layer, and no fully connected layer;
3)定义C
0的损失函数,
3) Define the loss function of C 0,
4)训练得到非线性映射函数F
λ(p),遍历训练集c中所有样本图像,训练输出图像增强计算函数,即非线性映射函数F
λ(p);
4) Train to obtain the nonlinear mapping function F λ (p), traverse all the sample images in the training set c, and train the output image enhancement calculation function, namely the nonlinear mapping function F λ (p);
5)基于非线性映射函数F
λ(p),计算输出图像p的高峰值信噪比图像;
5) Calculate the high-peak signal-to-noise ratio image of the output image p based on the nonlinear mapping function F λ (p);
6)使用笔迹特征计算方法,计算高峰值信噪比图像的假想笔画特征、路径签名特征与8方向特征;6) Use the handwriting feature calculation method to calculate the imaginary stroke feature, path signature feature and 8-direction feature of the high-peak signal-to-noise ratio image;
7)构建具有6层卷积层、5层池化层和1层全连接层的集成卷积神经网络模型C
1;
7) Construct an integrated convolutional neural network model C 1 with 6 layers of convolutional layers, 5 layers of pooling layers and 1 layer of fully connected layers;
8)遍历训练集中所有样本图像,结合假想笔画特征、路径签名特征与8个方向特征,训练得到电力操作票文字识别模型;8) Traverse all the sample images in the training set, combine the imaginary stroke features, path signature features and 8 direction features to train to obtain the power operation ticket text recognition model;
9)获取需要识别的电力操作票,通过电力操作票文字识别模型进行文字识别。9) Obtain the power operation ticket that needs to be recognized, and perform text recognition through the power operation ticket text recognition model.
作为优选技术手段:在步骤1)中,针对样本图像p
i(p
i∈c,1≤i≤N)得到训练集,其中N是训练集c包含的样本图像总数,获取p
i的M×M数值矩阵A
i,及其对应的清晰图像数值矩阵B
i。
Preferred techniques: in step 1), for a sample image p i (p i ∈c, 1≤i≤N ) to give the training set, where N is the total number of training set sample image (c) includes the acquiring of M × p i The M numerical matrix A i and the corresponding clear image numerical matrix B i .
作为优选技术手段:在步骤2)中构建三层卷积神经网络模型C
0,仅含卷积层,选择激活函数ReLU,步长设置为1,不对卷积运算填充0, 网络结构为:
As a preferred technical means: construct a three-layer convolutional neural network model C 0 in step 2), which only contains the convolutional layer, select the activation function ReLU, set the step size to 1, and do not fill the convolution operation with 0, the network structure is:
作为优选技术手段:在步骤3)中定义损失函数,损失函数目的是获得最小F范数,计算公式如下:As a preferred technical means: define the loss function in step 3). The purpose of the loss function is to obtain the minimum F norm. The calculation formula is as follows:
其中λ={W
j,b
j};其中W
j={W
j
k:1≤k≤n
j},W
j
k为卷积网络第j层的卷积矩阵,b
j为偏差值,n
j为卷积网络第j层的卷积核个数。
Where λ = {W j ,b j }; where W j ={W j k :1≤k≤n j }, W j k is the convolution matrix of the jth layer of the convolutional network, b j is the deviation value, n j is the number of convolution kernels in the jth layer of the convolutional network.
作为优选技术手段:在步骤5)中使用非线性映射函数F
λ(p),计算训练集c中每一个样本图像p
i对应的增强图像P
i,得到新训练集C(P
i∈C,1≤i≤N)。
In step 5) is non-linear mapping function F λ (p), c in the training set is calculated for each sample image to enhance the image corresponding to p i of P i, to give a new training set C (P i ∈C,: as a preferred technology 1≤i≤N).
作为优选技术手段:步骤6)包括步骤:As a preferred technical means: Step 6) includes the steps:
a)假想笔画特征提取笔迹方向变化程度,针对每一相邻像素点,计算公式如下:a) The imaginary stroke feature extracts the degree of change in the direction of the handwriting. For each adjacent pixel, the calculation formula is as follows:
其中,θ为不同笔画之间相连构成的夹角度数(180≤θ≤180),l为笔画长度,
w=1/8;比较不同像素点dcd的值,计算得到P
i的假想笔画矩阵;
Among them, θ is the number of angles between different strokes (180≤θ≤180), and l is the length of the stroke, w = 1/8; value comparing different pixels dcd, the calculated virtual matrix P i of the stroke;
b)路径签名特征提取笔迹曲率值,给定笔迹起止区间为[s,t],其曲率特征定义如下:b) The path signature feature extracts the handwriting curvature value, the starting and ending interval of the given handwriting is [s, t], and the curvature feature is defined as follows:
若k=0,则0重积分特征计算结果为1,表示笔迹的二值图像特征;若k=1,则1重积分特征表示笔迹的位移特征;若k=2,则2重积分特征表示笔迹的曲率特征;k值可取任意值,正常情况下不宜取值太大,否则会导致计算复杂度指数级增加,却不能获取更多有效笔迹特征;针对每一相邻像素点,计算得到P
i的路径签名矩阵;
If k=0, the zero-fold integral feature calculation result is 1, which represents the binary image feature of the handwriting; if k=1, the 1-fold integral feature represents the displacement feature of the handwriting; if k=2, the double-integrated feature represents The curvature characteristic of the handwriting; the value of k can take any value, under normal circumstances it should not be too large, otherwise it will cause the computational complexity to increase exponentially, but it will not be able to obtain more effective handwriting features; for each adjacent pixel, P is calculated i 's path signature matrix;
c)8方向特征能够出色地拟合汉字的横竖撇捺等笔画,可用于提取笔迹方向信息;假设给定一个二维坐标,8方向特征分别从0°,45°,90°,135°,180°225°,270°,315°计算笔迹梯度大小;给定一段笔迹的起止坐标(x
1,y
1)与(x
2,y
2),梯度计算公式如下:
c) The 8-direction feature can perfectly fit the strokes of Chinese characters, such as horizontal and vertical strokes, and can be used to extract handwriting direction information; assuming a two-dimensional coordinate, the 8-direction features are from 0°, 45°, 90°, 135°, 180°225°, 270°, 315° calculate the handwriting gradient size; given the starting and ending coordinates (x 1 , y 1 ) and (x 2 , y 2 ) of a handwriting, the gradient calculation formula is as follows:
其中d
x=|x
2-x
1|,d
y=|y
2-y
1|,
针对每一相邻像素点,计算得到P
i的8方向矩阵。
Where d x =|x 2 -x 1 |, d y = |y 2 -y 1 |, For each adjacent pixel, calculated matrix P i of the 8 directions.
作为优选技术手段:在步骤7)中,集成卷积神经网络模型C
1的第1层卷积网络的卷积核大小设置为3×3,卷积核个数为80,且依次递增80;第2~6层卷积网络的卷积核大小设置为2×2;选择激活函数ReLU,选择补0卷积运算,步长取1;池化层矩阵大小均为2×2;集成卷积神经网络模型C
1结构为:
As a preferred technical means: in step 7), the size of the convolution kernel of the first layer of the convolutional network of the integrated convolutional neural network model C 1 is set to 3×3, the number of convolution kernels is 80, and the number is increased by 80 in sequence; The size of the convolution kernel of the second to sixth layers of the convolutional network is set to 2×2; select the activation function ReLU, select the zero-complement convolution operation, and the step size is 1; the size of the pooling layer matrix is 2×2; integrated convolution The structure of neural network model C 1 is:
表2集成卷积神经网络模型C
1结构
Table 2 Integrated convolutional neural network model C 1 structure
其中n表示笔迹特征维度个数。Where n represents the number of handwriting feature dimensions.
1)使用自定义三层卷积神经网络训练得到非线性映射函数,输出高峰值信噪比值图像,解决输入图像低清晰度问题,有利于后续文字的准确识别;1) Use a custom three-layer convolutional neural network to train to obtain a non-linear mapping function, output a high-peak signal-to-noise ratio image, solve the problem of low definition of the input image, and facilitate the accurate recognition of subsequent text;
2)基于笔迹特征方法,分别计算手写字体的假想笔画特征、路径签名特征与8方向特征,解决卷积神经网络图像空间特征学习的局限性,准确地表达了手写字体笔迹信息;2) Based on the handwriting feature method, separately calculate the imaginary stroke feature, path signature feature and 8-direction feature of handwritten fonts, solve the limitations of convolutional neural network image spatial feature learning, and accurately express the handwriting information of handwritten fonts;
3)通过融合笔迹特征,使用集成卷积神经网络模型训练得到电力操作票文字识别模型,有效地提升了文字识别的准确度。3) Through the fusion of handwriting features, the integrated convolutional neural network model is used to train the power operation ticket text recognition model, which effectively improves the accuracy of text recognition.
图1是本发明的流程图。Figure 1 is a flowchart of the present invention.
以下结合说明书附图对本发明的技术方案做进一步的详细说明。The technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings of the specification.
如图1所示,本发明包括步骤:As shown in Figure 1, the present invention includes the steps:
1)获取样本图像,得到训练集;1) Obtain sample images and get the training set;
2)构建仅具有3层卷积层,无池化层,无全连接层的卷积神经网络模型C
0;
2) Construct a convolutional neural network model C 0 with only 3 convolutional layers, no pooling layer, and no fully connected layer;
3)定义C
0的损失函数,
3) Define the loss function of C 0,
4)训练得到非线性映射函数F
λ(p),遍历训练集c中所有样本图像,训练输出图像增强计算函数,即非线性映射函数F
λ(p);
4) Train to obtain the nonlinear mapping function F λ (p), traverse all the sample images in the training set c, and train the output image enhancement calculation function, namely the nonlinear mapping function F λ (p);
5)基于非线性映射函数F
λ(p),计算输出图像p的高峰值信噪比图像;
5) Calculate the high-peak signal-to-noise ratio image of the output image p based on the nonlinear mapping function F λ (p);
6)使用笔迹特征计算方法,计算高峰值信噪比图像的假想笔画特征、路径签名特征与8方向特征;6) Use the handwriting feature calculation method to calculate the imaginary stroke feature, path signature feature and 8-direction feature of the high-peak signal-to-noise ratio image;
7)构建具有6层卷积层、5层池化层和1层全连接层的集成卷积神经网络模型C
1;
7) Construct an integrated convolutional neural network model C 1 with 6 layers of convolutional layers, 5 layers of pooling layers and 1 layer of fully connected layers;
8)遍历训练集中所有样本图像,结合假想笔画特征、路径签名特征与8个方向特征,训练得到电力操作票文字识别模型;8) Traverse all the sample images in the training set, combine the imaginary stroke features, path signature features and 8 direction features to train to obtain the power operation ticket text recognition model;
9)获取需要识别的电力操作票,通过电力操作票文字识别模型进行文字识别。9) Obtain the power operation ticket that needs to be recognized, and perform text recognition through the power operation ticket text recognition model.
以上步骤分为图像增强处理、文字识别处理两个阶段。The above steps are divided into two stages: image enhancement processing and character recognition processing.
以下就各阶段进行具体的说明:The following is a detailed description of each stage:
图像增强处理阶段Image enhancement processing stage
图像增强处理的步骤如下:The steps of image enhancement processing are as follows:
步骤1,针对样本图像p
i(p
i∈c,1≤i≤N),其中N是训练集c包含的样本图像总数,获取p
i的M×M数值矩阵A
i,及其对应的清晰图像数值矩阵B
i。
Step 1, for a sample image p i (p i ∈c, 1≤i≤N ), where N is the total number of samples of training images comprises a set of c, obtaining M × M matrix of values of A i p i, and the corresponding clear Image value matrix B i .
步骤2,构建三层卷积神经网络模型C
0,仅含卷积层,选择激活函数ReLU,步长设置为1,不对卷积运算填充0,网络结构如表1所示:
Step 2. Construct a three-layer convolutional neural network model C 0 with only the convolutional layer, select the activation function ReLU, set the step size to 1, and do not fill the convolution operation with 0. The network structure is shown in Table 1:
表1自定义卷积神经网络模型C
0结构
Table 1 Custom Convolutional Neural Network Model C 0 Structure
步骤3,定义损失函数,与B
i相比较,损失函数目的是获得最小F范数,计算公式如下:
Step 3. Define the loss function. Compared with B i , the purpose of the loss function is to obtain the minimum F norm. The calculation formula is as follows:
其中λ={W
j,b
j}。其中W
j={W
j
k:1≤k≤n
j},W
j
k为卷积网络第j层的卷积矩阵,b
j为偏差值,n
j为卷积网络第j层的卷积核个数。
Where λ={W j ,b j }. Where W j = {W j k :1≤k≤n j }, W j k is the convolution matrix of the jth layer of the convolutional network, b j is the deviation value, and n j is the convolution of the jth layer of the convolutional network The number of cores.
步骤4,遍历训练集c中所有样本图像,训练输出图像增强计算函数,即非线性映射函数F
λ(p)。
Step 4: Traverse all the sample images in the training set c, and train the output image enhancement calculation function, that is, the non-linear mapping function F λ (p).
文字识别处理阶段Word recognition processing stage
文字识别处理的步骤如下:The steps of word recognition processing are as follows:
步骤1,使用非线性映射函数F
λ(p),计算训练集c中每一个样本图像p
i对应的增强图像P
i,得到新训练集C(P
i∈C,1≤i≤N)。
Step 1 using a nonlinear mapping function F λ (p), calculated in the training set c p i enhanced image corresponding to each sample P i, to give a new training set C (P i ∈C, 1≤i≤N) .
步骤2,基于笔迹特征计算方法,分别计算P
i的假想笔画特征、路径签名特征与8方向特征。具体计算方法如下:
2, wherein the method is calculated based on handwriting, P i are calculated hypothetical stroke features, and wherein the signature path 8 wherein the step direction. The specific calculation method is as follows:
d)假想笔画特征提取笔迹方向变化程度,针对每一相邻像素点,计算公式如下:d) The imaginary stroke feature extracts the degree of change in the direction of the handwriting. For each adjacent pixel, the calculation formula is as follows:
其中,θ为不同笔画之间相连构成的夹角度数(180≤θ≤180),l为笔画长度,
w=1/8。比较不同像素点dcd的值,计算得到P
i的假想笔画矩阵。
Among them, θ is the number of angles between different strokes (180≤θ≤180), and l is the length of the stroke, w=1/8. Different pixel value comparison dcd, the calculated virtual matrix P i of the stroke.
e)路径签名特征提取笔迹曲率值,给定笔迹起止区间为[s,t],其曲率特征定义如下:e) The path signature feature extracts the handwriting curvature value. The starting and ending interval of the given handwriting is [s, t], and the curvature feature is defined as follows:
若k=0,则0重积分特征计算结果为1,表示笔迹的二值图像特征;若k=1,则1重积分特征表示笔迹的位移特征;若k=2,则2重积分特征表示笔迹的曲率特征。k值可取任意值,正常情况下不宜取值太大,否则会导致计算复杂度指数级增加,却不能获取更多有效笔迹特征。针对每一相邻像素点,计算得到P
i的路径签名矩阵。
If k=0, the zero-fold integral feature calculation result is 1, which represents the binary image feature of the handwriting; if k=1, the 1-fold integral feature represents the displacement feature of the handwriting; if k=2, the double-integrated feature represents The curvature characteristics of the handwriting. The k value can take any value, and it should not be too large under normal circumstances, otherwise it will cause the computational complexity to increase exponentially, but it will not be possible to obtain more effective handwriting features. For each adjacent pixel, the path P i calculated signature matrix.
f)8方向特征能够出色地拟合汉字的横竖撇捺等笔画,可用于提取笔迹方向信息。假设给定一个二维坐标,8方向特征分别从0°,45°,90°,135°,180°225°,270°,315°计算笔迹梯度大小。给定一段笔迹的起止坐标(x
1,y
1)与(x
2,y
2),梯度计算公式如下:
f) The 8-direction feature can perfectly fit the strokes such as horizontal and vertical strokes of Chinese characters, and can be used to extract handwriting direction information. Assuming that a two-dimensional coordinate is given, the 8-direction features are calculated from 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° respectively. Given the starting and ending coordinates (x 1 , y 1 ) and (x 2 , y 2 ) of a handwriting, the gradient calculation formula is as follows:
其中d
x=|x
2-x
1|,d
y=|y
2-y
1|,
针对每一相邻像素点,计算得到P
i的8方向矩阵。
Where d x =|x 2 -x 1 |, d y = |y 2 -y 1 |, For each adjacent pixel, calculated matrix P i of the 8 directions.
步骤3,构建集成卷积网络模型C
1,该模型包含6层卷积网络,前5层卷积网络下一层均配置池化层,第6层卷积网络的下一层配置全连接层。第1层卷积网络的卷积核大小设置为3×3,卷积核个数为80,且依次递增80;第2~6层卷积网络的卷积核大小设置为2×2;选择激活函数ReLU,选择补0卷积运算,步长取1。池化层矩阵大小均为2×2。模型结构如表2所示,其中n表示笔迹特征维度个数。
Step 3. Construct an integrated convolutional network model C 1 , which contains a 6-layer convolutional network. The next layer of the first 5 layers of convolutional network is configured with a pooling layer, and the next layer of the 6th layer of convolutional network is configured with a fully connected layer . The size of the convolution kernel of the first layer of convolutional network is set to 3×3, the number of convolution kernels is 80, and the number of convolution kernels is increased by 80; the size of the convolution kernel of the second to sixth layers of convolutional network is set to 2×2; choose The activation function ReLU is selected to complement 0 convolution operation, and the step size is 1. The size of the pooling layer matrix is 2×2. The model structure is shown in Table 2, where n represents the number of handwriting feature dimensions.
表2集成卷积神经网络模型C
1结构
Table 2 Integrated convolutional neural network model C 1 structure
步骤4,遍历训练集C中所有样本图像,结合假想笔画特征、路径签名特征与8方向特征,训练输出文字识别模型。Step 4: Traverse all the sample images in the training set C, combine the imaginary stroke features, path signature features and 8-direction features to train and output a text recognition model.
以上图1所示的一种基于卷积神经网络的电力操作票文字识别方法是本发明的具体实施例,已经体现出本发明实质性特点和进步,可根据实际的使用需要,在本发明的启示下,对其进行形状、结构等方面的等同修改,均在本方案的保护范围之列。The above-mentioned method for recognizing power operation ticket characters based on convolutional neural network shown in Figure 1 is a specific embodiment of the present invention, which has embodied the substantive features and progress of the present invention, and can be used in accordance with actual needs of the present invention. Under the enlightenment, equivalent modifications to the shape, structure, etc., are all within the scope of protection of this scheme.
Claims (7)
- 一种基于卷积神经网络的电力操作票文字识别方法,其特征在于包括以下步骤:A method for recognizing power operation ticket characters based on convolutional neural network, which is characterized in that it comprises the following steps:1)获取样本图像,得到训练集;1) Obtain sample images and get the training set;2)构建仅具有3层卷积层,无池化层,无全连接层的卷积神经网络模型C 0; 2) Construct a convolutional neural network model C 0 with only 3 convolutional layers, no pooling layer, and no fully connected layer;3)定义C 0的损失函数; 3) Define the loss function of C 0;4)训练得到非线性映射函数F λ(p),遍历训练集c中所有样本图像,训练输出图像增强计算函数,即非线性映射函数F λ(p); 4) Train to obtain the nonlinear mapping function F λ (p), traverse all the sample images in the training set c, and train the output image enhancement calculation function, namely the nonlinear mapping function F λ (p);5)基于非线性映射函数F λ(p),计算输出图像p的高峰值信噪比图像; 5) Calculate the high-peak signal-to-noise ratio image of the output image p based on the nonlinear mapping function F λ (p);6)使用笔迹特征计算方法,计算高峰值信噪比图像的假想笔画特征、路径签名特征与8方向特征;6) Use the handwriting feature calculation method to calculate the imaginary stroke feature, path signature feature and 8-direction feature of the high-peak signal-to-noise ratio image;7)构建具有6层卷积层、5层池化层和1层全连接层的集成卷积神经网络模型C 1; 7) Construct an integrated convolutional neural network model C 1 with 6 layers of convolutional layers, 5 layers of pooling layers and 1 layer of fully connected layers;8)遍历训练集中所有样本图像,结合假想笔画特征、路径签名特征与8个方向特征,训练得到电力操作票文字识别模型;8) Traverse all the sample images in the training set, combine the imaginary stroke features, path signature features and 8 direction features to train to obtain the power operation ticket text recognition model;9)获取需要识别的电力操作票,通过电力操作票文字识别模型进行文字识别。9) Obtain the power operation ticket that needs to be recognized, and perform text recognition through the power operation ticket text recognition model.
- 根据权利要求1所述的一种基于卷积神经网络的电力操作票文字识别方法,其特征在于:在步骤1)中,针对样本图像p i(p i∈c,1≤i≤N)得到训练集,其中N是训练集c包含的样本图像总数,获取p i的M×M数值矩阵A i,及其对应的清晰图像数值矩阵B i。 In step 1), to obtain a sample image for the p i (p i ∈c, 1≤i≤N ): according to one of claim 1 to claim power generating operation character recognition method based on convolutional neural network, wherein training set, where N is the total number of samples of training images comprises a set of c, obtaining M × M matrix of values a i p i, the sharp image and its corresponding numerical matrix B i.
- 根据权利要求2所述的一种基于卷积神经网络的电力操作票文字 识别方法,其特征在于:在步骤2)中构建三层卷积神经网络模型C 0,仅含卷积层,选择激活函数ReLU,步长设置为1,不对卷积运算填充0,网络结构为: The method for recognizing power operation ticket text based on convolutional neural network according to claim 2, characterized in that: in step 2), a three-layer convolutional neural network model C 0 is constructed, which contains only the convolutional layer, and selects activation Function ReLU, the step size is set to 1, and the convolution operation is not filled with 0, the network structure is:
- 根据权利要求3所述的一种基于卷积神经网络的电力操作票文字识别方法,其特征在于:在步骤3)中定义损失函数,损失函数目的是获得最小F范数,计算公式如下:The method for recognizing power operation ticket characters based on convolutional neural network according to claim 3, characterized in that: in step 3), a loss function is defined, the purpose of the loss function is to obtain the minimum F norm, and the calculation formula is as follows:
- 根据权利要求4所述的一种基于卷积神经网络的电力操作票文字识别方法,其特征在于:在步骤5)中使用非线性映射函数F λ(p),计算训练集c中每一个样本图像p i对应的增强图像P i,得到新训练集C(P i∈C,1≤i≤N)。 The method for recognizing power operation ticket characters based on convolutional neural network according to claim 4, characterized in that: in step 5), a non-linear mapping function F λ (p) is used to calculate each sample in the training set c p i enhanced image corresponding to P i, to give a new training set C (P i ∈C, 1≤i≤N) .
- 根据权利要求5所述的一种基于卷积神经网络的电力操作票文字识别方法,其特征在于:步骤6)包括步骤:The method for recognizing power operation ticket characters based on convolutional neural network according to claim 5, characterized in that: step 6) comprises the steps:a)假想笔画特征提取笔迹方向变化程度,针对每一相邻像素点,计算公式如下:a) The imaginary stroke feature extracts the degree of change in the direction of the handwriting. For each adjacent pixel, the calculation formula is as follows:其中,θ为不同笔画之间相连构成的夹角度数(180≤θ≤180),l为笔画长度, w=1/8;比较不同像素点dcd的值,计算得到P i的假想笔画矩阵; Among them, θ is the number of angles between different strokes (180≤θ≤180), and l is the length of the stroke, w = 1/8; value comparing different pixels dcd, the calculated virtual matrix P i of the stroke;b)路径签名特征提取笔迹曲率值,给定笔迹起止区间为[s,t],其曲率特征定义如下:b) The path signature feature extracts the handwriting curvature value, the starting and ending interval of the given handwriting is [s, t], and the curvature feature is defined as follows:若k=0,则0重积分特征计算结果为1,表示笔迹的二值图像特征;若k=1,则1重积分特征表示笔迹的位移特征;若k=2,则2重积分特征表示笔迹的曲率特征;k值可取任意值,正常情况下不宜取值太大,否则会导致计算复杂度指数级增加,却不能获取更多有效笔迹特征;针对每一相邻像素点,计算得到P i的路径签名矩阵; If k=0, the zero-fold integral feature calculation result is 1, which represents the binary image feature of the handwriting; if k=1, the 1-fold integral feature represents the displacement feature of the handwriting; if k=2, the double-integrated feature represents The curvature characteristic of the handwriting; the value of k can take any value, under normal circumstances it should not be too large, otherwise it will cause the computational complexity to increase exponentially, but it will not be able to obtain more effective handwriting features; for each adjacent pixel, P is calculated i 's path signature matrix;c)8方向特征能够出色地拟合汉字的横竖撇捺等笔画,可用于提取笔迹方向信息;假设给定一个二维坐标,8方向特征分别从0°,45°,90°,135°,180°225°,270°,315°计算笔迹梯度大小;给定一段笔迹的起止坐标(x 1,y 1)与(x 2,y 2),梯度计算公式如下: c) The 8-direction feature can perfectly fit the strokes of Chinese characters, such as horizontal and vertical strokes, and can be used to extract handwriting direction information; assuming a two-dimensional coordinate, the 8-direction features are from 0°, 45°, 90°, 135°, 180°225°, 270°, 315° calculate the handwriting gradient size; given the starting and ending coordinates (x 1 , y 1 ) and (x 2 , y 2 ) of a handwriting, the gradient calculation formula is as follows:
- 根据权利要求6所述的一种基于卷积神经网络的电力操作票文字识别方法,其特征在于:在步骤7)中,集成卷积神经网络模型C 1的第1层卷积网络的卷积核大小设置为3×3,卷积核个数为80,且依 次递增80;第2~6层卷积网络的卷积核大小设置为2×2;选择激活函数ReLU,选择补0卷积运算,步长取1;池化层矩阵大小均为2×2;集成卷积神经网络模型C 1结构为: One of the claim 6, the character recognition method in the power generating operation convolutional neural network, wherein: in step 7), integrated Convolution Neural Network Model C convolutional network layer 1 1 The kernel size is set to 3×3, the number of convolution kernels is 80, and the number of convolution kernels is 80, and the number of convolution kernels is set to 2×2; the activation function ReLU is selected, and 0 convolution is selected. Operation, the step size is 1; the size of the pooling layer matrix is 2×2; the structure of the integrated convolutional neural network model C 1 is:表2集成卷积神经网络模型C 1结构 Table 2 Integrated convolutional neural network model C 1 structure其中n表示笔迹特征维度个数。Where n represents the number of handwriting feature dimensions.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010022127.9A CN111259880B (en) | 2020-01-09 | 2020-01-09 | Electric power operation ticket character recognition method based on convolutional neural network |
CN202010022127.9 | 2020-01-09 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021139175A1 true WO2021139175A1 (en) | 2021-07-15 |
Family
ID=70954083
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2020/111550 WO2021139175A1 (en) | 2020-01-09 | 2020-08-27 | Electric power operation ticket character recognition method based on convolutional neural network |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN111259880B (en) |
WO (1) | WO2021139175A1 (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113780306A (en) * | 2021-08-11 | 2021-12-10 | 北京工业大学 | Waste mobile phone color identification method based on deep convolutional neural network |
CN113959071A (en) * | 2021-07-21 | 2022-01-21 | 北京金茂绿建科技有限公司 | Machine learning assistance-based operation control optimization method for centralized water chilling unit air conditioning system |
CN114221334A (en) * | 2021-12-13 | 2022-03-22 | 国网江苏省电力有限公司 | Rapid state estimation method based on graph neural network |
CN114218984A (en) * | 2021-12-07 | 2022-03-22 | 桂林电子科技大学 | Radio frequency fingerprint identification method based on sample multi-view learning |
CN116187313A (en) * | 2023-04-27 | 2023-05-30 | 南方电网数字电网研究院有限公司 | Power operation ticket equipment identification and error investigation method based on natural language processing technology |
CN116824306A (en) * | 2023-08-28 | 2023-09-29 | 天津大学 | Training method of pen stone fossil image recognition model based on multi-mode metadata |
CN117272244A (en) * | 2023-11-21 | 2023-12-22 | 中国石油大学(华东) | Soft measurement modeling method integrating feature extraction and self-adaptive composition |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111259880B (en) * | 2020-01-09 | 2022-11-18 | 国网浙江省电力有限公司舟山供电公司 | Electric power operation ticket character recognition method based on convolutional neural network |
CN113159604B (en) * | 2021-04-30 | 2023-05-09 | 贵州电网有限责任公司 | Power system dispatching operation visualization method based on augmented reality |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105335754A (en) * | 2015-10-29 | 2016-02-17 | 小米科技有限责任公司 | Character recognition method and device |
CN110210468A (en) * | 2019-05-29 | 2019-09-06 | 电子科技大学 | A kind of character recognition method based on the migration of convolutional neural networks Fusion Features |
CN111259880A (en) * | 2020-01-09 | 2020-06-09 | 国网浙江省电力有限公司舟山供电公司 | Electric power operation ticket character recognition method based on convolutional neural network |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140363082A1 (en) * | 2013-06-09 | 2014-12-11 | Apple Inc. | Integrating stroke-distribution information into spatial feature extraction for automatic handwriting recognition |
CN104751422B (en) * | 2015-03-12 | 2016-04-06 | 中南大学 | A kind of photocopy file and picture strengthens and binarization method |
RU2661750C1 (en) * | 2017-05-30 | 2018-07-19 | Общество с ограниченной ответственностью "Аби Продакшн" | Symbols recognition with the use of artificial intelligence |
CN109583449A (en) * | 2018-10-29 | 2019-04-05 | 深圳市华尊科技股份有限公司 | Character identifying method and Related product |
CN110008917B (en) * | 2019-04-11 | 2023-02-17 | 国网河南省电力公司电力科学研究院 | Fault detection method of fixed value single data of relay protection device based on table understanding |
-
2020
- 2020-01-09 CN CN202010022127.9A patent/CN111259880B/en active Active
- 2020-08-27 WO PCT/CN2020/111550 patent/WO2021139175A1/en active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105335754A (en) * | 2015-10-29 | 2016-02-17 | 小米科技有限责任公司 | Character recognition method and device |
CN110210468A (en) * | 2019-05-29 | 2019-09-06 | 电子科技大学 | A kind of character recognition method based on the migration of convolutional neural networks Fusion Features |
CN111259880A (en) * | 2020-01-09 | 2020-06-09 | 国网浙江省电力有限公司舟山供电公司 | Electric power operation ticket character recognition method based on convolutional neural network |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113959071A (en) * | 2021-07-21 | 2022-01-21 | 北京金茂绿建科技有限公司 | Machine learning assistance-based operation control optimization method for centralized water chilling unit air conditioning system |
CN113959071B (en) * | 2021-07-21 | 2023-05-26 | 北京金茂绿建科技有限公司 | Centralized water chilling unit air conditioning system operation control optimization method based on machine learning assistance |
CN113780306B (en) * | 2021-08-11 | 2024-04-09 | 北京工业大学 | Deep convolutional neural network-based waste mobile phone color recognition method |
CN113780306A (en) * | 2021-08-11 | 2021-12-10 | 北京工业大学 | Waste mobile phone color identification method based on deep convolutional neural network |
CN114218984B (en) * | 2021-12-07 | 2024-03-22 | 桂林电子科技大学 | Radio frequency fingerprint identification method based on sample multi-view learning |
CN114218984A (en) * | 2021-12-07 | 2022-03-22 | 桂林电子科技大学 | Radio frequency fingerprint identification method based on sample multi-view learning |
CN114221334A (en) * | 2021-12-13 | 2022-03-22 | 国网江苏省电力有限公司 | Rapid state estimation method based on graph neural network |
CN116187313B (en) * | 2023-04-27 | 2023-06-27 | 南方电网数字电网研究院有限公司 | Power operation ticket equipment identification and error investigation method based on natural language processing technology |
CN116187313A (en) * | 2023-04-27 | 2023-05-30 | 南方电网数字电网研究院有限公司 | Power operation ticket equipment identification and error investigation method based on natural language processing technology |
CN116824306B (en) * | 2023-08-28 | 2023-11-17 | 天津大学 | Training method of pen stone fossil image recognition model based on multi-mode metadata |
CN116824306A (en) * | 2023-08-28 | 2023-09-29 | 天津大学 | Training method of pen stone fossil image recognition model based on multi-mode metadata |
CN117272244A (en) * | 2023-11-21 | 2023-12-22 | 中国石油大学(华东) | Soft measurement modeling method integrating feature extraction and self-adaptive composition |
CN117272244B (en) * | 2023-11-21 | 2024-03-15 | 中国石油大学(华东) | Soft measurement modeling method integrating feature extraction and self-adaptive composition |
Also Published As
Publication number | Publication date |
---|---|
CN111259880B (en) | 2022-11-18 |
CN111259880A (en) | 2020-06-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021139175A1 (en) | Electric power operation ticket character recognition method based on convolutional neural network | |
CN109993160B (en) | Image correction and text and position identification method and system | |
Tang et al. | Text-independent writer identification via CNN features and joint Bayesian | |
CN111832546B (en) | Lightweight natural scene text recognition method | |
CN105426919B (en) | The image classification method of non-supervisory feature learning is instructed based on conspicuousness | |
CN107122712B (en) | Palm print image identification method based on CNN and bidirectional VLAD | |
CN106570456A (en) | Handwritten Chinese character recognition method based on full-convolution recursive network | |
CN111401384A (en) | Transformer equipment defect image matching method | |
CN106408038A (en) | Rotary Chinese character identifying method based on convolution neural network model | |
CN108664975B (en) | Uyghur handwritten letter recognition method and system and electronic equipment | |
CN106250811B (en) | Unconstrained face identification method based on HOG feature rarefaction representation | |
CN109800746A (en) | A kind of hand-written English document recognition methods based on CNN | |
Tan et al. | A new handwritten character segmentation method based on nonlinear clustering | |
CN109886391A (en) | A kind of neural network compression method based on the positive and negative diagonal convolution in space | |
CN111666937A (en) | Method and system for recognizing text in image | |
Wu et al. | Joint spatial and radical analysis network for distorted Chinese character recognition | |
CN106203448A (en) | A kind of scene classification method based on Nonlinear Scale Space Theory | |
Gao et al. | Progressive rectification network for irregular text recognition | |
CN113160291B (en) | Change detection method based on image registration | |
Chen | Research and application of chinese calligraphy character recognition algorithm based on image analysis | |
Qin et al. | Summary of scene text detection and recognition | |
Zheng et al. | Segmentation-free multi-font printed Manchu word recognition using deep convolutional features and data augmentation | |
CN110929652B (en) | Handwriting Chinese character recognition method based on LeNet-5 network model | |
Wei et al. | A holistic recognition approach for woodblock-print Mongolian words based on convolutional neural network | |
CN108537855A (en) | A kind of ceramic marble paper method for generating pattern and device that sketch is consistent |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 20912335 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20912335 Country of ref document: EP Kind code of ref document: A1 |