White gold body Tibetan ancient book similar character recognition method based on deep learning

Abstract

The invention discloses a method for identifying similar characters of an ancient book of Wujin Tibetan based on deep learning. According to the method, similar characters in the Wujin Tibetan ancient books can be identified by using a deep neural network technology. By establishing the deep convolutional neural network, the system can automatically learn the characteristics of the characters and perform accurate recognition. Meanwhile, model parameters are trained by utilizing the constructed similar character samples of the Tibetan and ancient books. According to the method, feature extraction is carried out on a Tibetan character image segmented in advance, gradient 8-direction features, 8 * 8 grid density features and character peripheral features are extracted, and the features are fused together. Through PCA dimensionality reduction and k-means clustering, samples are marked through unsupervised clustering, and similar character set samples are constructed; and finally, effective feature learning and recognition are realized through a deep neural network (DNN).

Description

Technical Field

[0001] The present invention relates to deep learning-based Tibetan text image recognition, and in particular to a deep learning-based method for recognizing similar characters in ancient Tibetan books in Ujin style. Background Art

[0002] Ancient Tibetan books contain modern Tibetan and Sanskrit Tibetan, with a total of more than 7,000 characters. These characters are highly similar, and there are many similar characters, such as " "," "and" "," "," "," "and" ", etc., which brings challenges to text recognition. Ancient Tibetan books are usually made of woodcuts. Due to the different depths of hand-carved marks, the ink is unevenly applied during later printing, which may cause missing character strokes. In addition, character strokes may also be missing during the image processing of ancient books, which increases the difficulty of recognition.

[0003] Currently, there are few studies on the recognition of similar characters in ancient Tibetan book images. Due to the non-reproducibility of ancient Tibetan books, character samples can only be extracted from book images. The project team has implemented the preprocessing, binarization and character segmentation of ancient Tibetan book images. Due to the requirements of Tibetan Buddhism in book printing, most Tibetan scriptures use the Ujin font. The characteristic of the Ujin font is that the top stroke of each letter is horizontal and vertical, and the baseline is on a straight line when the characters are arranged. The baseline helps to divide the characters into the vowel part above the baseline and the part below the baseline. There are fewer types of characters above the baseline, only a dozen or so, and fewer similar characters. Summary of the invention

[0004] The purpose of the present invention is to provide a method for recognizing similar characters in ancient Tibetan books in Ujin style based on deep learning, so as to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for recognizing similar characters in ancient Tibetan books in Wujin style based on deep learning, comprising the following steps:

[0007] 1: Tibetan estimated text preprocessing

[0008] The preprocessing process includes filtering and enhancing the whole image, as well as tilt and distortion correction, and then segmenting it into different regions, and marking the types of text, image, table or border in the region, and finally outputting different regions such as text, image and table border;

[0009] 2: Construct similar character set samples

[0010] Get real ancient Tibetan character samples from ancient Tibetan document images. By segmenting the image after binarization, we can get ancient Tibetan character image samples. Commonly used ancient Tibetan text segmentation algorithms include connected domain clustering, baseline segmentation, and background-based graph segmentation. Through feature extraction and clustering analysis, similar character sample sets are formed through classification and labeling.

[0011] 3: Feature extraction

[0012] (1) Gradient 8-directional features

[0013] First, the image of Tibetan ancient book characters is normalized to 136×50 size. To ensure the image has minimal distortion, bicubic interpolation is used in the deformation process. Then, a 4×2 equally divided grid is used to divide the original image into 8 small grids according to size. The gradient features of the character pixels in each small grid are calculated. The gradient is decomposed into 8 directions according to the Bai ZL method to form an 8-dimensional gradient direction feature. The 8 small grid features are combined to obtain a 64-dimensional gradient direction feature.

[0014] (2) 8×8 grid density characteristics

[0015] First, the Tibetan ancient book character image is deformed to 64×64 size. To ensure the image has minimal distortion, bicubic interpolation is used in the deformation process. Then, an 8×8 equally divided grid is used to divide the original image into 64 small grids according to the size. The percentage of character pixels in each small grid to the total character pixels is calculated to obtain a 64-dimensional feature.

[0016] (3) Character peripheral features

[0017] Using the grid divided during extraction in (2), continue to extract the pixel peripheral features of each small grid in four directions: from top to bottom, from bottom to top, from left to right, and from right to left. Combine the features in the four directions to form a 1D feature. The total number of features in 64 small grids is 64.

[0018] Finally, after integrating the above three features, the feature dimension is 192 in total. Through principal component analysis, the dimension is reduced to 80-dimensional features. K-means clustering is used, and the correspondence between each character file name and the distance between each class centroid is recorded. The characters are sorted within the class according to the distance. The first K characters that are classified into the same class and have a close distance are classified as similar characters to form a similar character set.

[0019] 4: Identify the network model

[0020] The convolutional neural network consists of alternating convolutional layers and pooling layers. Each layer contains multiple feature maps. The calculation process of the convolutional layer can be expressed by the following formula:

[0021]

[0022] in, Represents the number of layers where the convolutional layer is located. is the convolution kernel, the convolution kernel size is 5×5, is the bias, represents the activation function, which is , Represents an input feature map of the previous layer.

[0023] The sampling layer downsamples the feature map of the previous convolutional layer to obtain the same number of feature maps. The convolutional neural network uses stochastic gradient descent just like the traditional neural network training.

[0024] The input layer shows a 28×28 image of ancient Tibetan characters; the C1 layer represents the first convolutional layer, which contains 8 24×24 feature maps, and a node in each feature map is connected to a 5×5 area corresponding to the input layer; the S1 layer is a downsampling layer containing 8 12×12 feature maps, in which each node is connected to a 2×2 area of ​​the corresponding feature map in the C1 layer; followed by C2, as the second convolutional layer, with 16 feature maps, each of which is 8×8 in size. The connection between the S1 and C2 layers plays an important role in the feature extraction process; in addition, S2 represents the second sampling layer, with 16 feature maps, each of which is 4×4 in size; the final output layer contains 10 nodes, corresponding to the output category, and is fully connected to the S2 layer;

[0025] In the process of network training, error back propagation is adopted and gradient stochastic descent method is used to update the parameters. , ;

[0026] use Represents the error function, and the parameter update is shown as follows:

[0027]

[0028]

[0029] in is the control parameter of the descent rate. The selection is determined by testing methods, and the final selection is the descent rate parameter.

[0030] Furthermore, the preprocessed Tibetan ancient book images need to be binarized.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) Deep convolutional neural networks can automatically learn effective features and perform recognition at the pixel level, avoiding the loss of similar word details caused by manual feature selection and extraction, thereby improving the recognition rate;

[0033] (2) Increasing the number of training samples will make deep convolutional neural networks perform well in reducing the false recognition rate. Increasing the number of training samples has a significant impact on improving the recognition accuracy of deep neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The present invention introduces the Ujin Tibetan character recognition process;

[0035] Figure 2 The figure is a schematic diagram of the process of constructing a similar character set of the Tibetan Ujin style according to the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Aiming at the problem of similar character recognition in ancient Tibetan books, the present invention applies deep neural networks (DNN) to the recognition of similar characters in ancient Tibetan books in Wujin style. First, unsupervised clustering is used to label samples and construct similar character set samples; then, the deep convolutional neural network is used to automatically learn the features of similar characters in ancient Tibetan books in Wujin style and recognize them, and the similar character samples of ancient Tibetan books constructed in this paper are used to train the model parameters. Through simulation experiments, the effects of different methods on the recognition rate are compared. Compared with the traditional support vector machine (SVM) and naive Bayes classifier (NBC) methods based on gradient features, the use of deep learning to distinguish similar characters has more advantages, which is specifically manifested in: first, the deep convolutional neural network can automatically learn effective features from the pixel level and recognize them, avoiding the loss of similar character details caused by manual feature selection and extraction, thereby improving the recognition rate; second, as the number of training samples increases, the deep convolutional neural network performs significantly in reducing the error recognition rate, and the increase in training samples has a significant effect on improving the recognition rate of the deep neural network.

[0038] The details are as follows:

[0039] 2. Tibetan estimated text preprocessing

[0040] The preprocessing process includes filtering and enhancing the whole image, as well as tilt and distortion correction, and then segmenting it into different regions and calibrating the types of text, image, table or border in the region. Finally, the different regions such as text, image and table border are output.

[0041] The pre-processed Tibetan ancient book image also needs to be binarized. The image pixel values ​​only contain black and white. In systems that use different color values, the pixel values ​​representing "black and white" are slightly different. If the image is binarized, the grayscale image is converted into an image containing only two grayscale levels, 0 and 255. In this image, a pixel value of 0 represents black, while a pixel value of 255 represents white, and the entire image contains only these two values. In addition, "black and white" can also be represented by logical values ​​to represent image pixels, where 0 still represents black, but here 0 represents false (FALSE), and 1 represents white, that is, true (TRUE). In this way, the entire image contains only two values, 0 or 1. Different value representation methods will result in slightly different subsequent calculations. For the convenience of display and calculation, the entire Tibetan ancient book binarized image uses black as the text foreground color and white as the background color. For the single character image of the Tibetan ancient book in the later feature extraction and deep network calculation, white is used as the foreground color and black is used as the background color.

[0042] 2. Construct similar character set samples

[0043] To obtain real ancient Tibetan character samples, we must obtain them from ancient Tibetan document images. By segmenting the image after the previous binarization, we can obtain ancient Tibetan character image samples. Commonly used ancient Tibetan text segmentation algorithms include connected domain clustering, baseline segmentation, and background-based graph segmentation. Through feature extraction, clustering analysis, and classification and labeling, similar character sample sets are formed. The similar character set construction process is shown in the attached figure. Figure 2 .

[0044] 3. Feature extraction

[0045] (1) Gradient 8-directional features

[0046] Firstly, the image of Tibetan ancient book characters is normalized to 136×50 size. To ensure the image has minimal distortion, bicubic interpolation is used in the deformation process. Then, a 4×2 equally divided grid is used to divide the original image into 8 small grids according to the size. The gradient features of the character pixels in each small grid are calculated. The gradient is decomposed into 8 directions according to the Bai ZL method to form an 8-dimensional gradient direction feature. The 8 small grid features are combined to obtain a 64-dimensional gradient direction feature.

[0047] (2) 8×8 grid density characteristics

[0048] Firstly, the image of Tibetan ancient book characters is deformed into 64×64 size. To ensure the image has minimal distortion, bicubic interpolation is used in the deformation process. Then, an 8×8 equally divided grid is used to divide the original image into 64 small grids according to their size. The percentage of character pixels in each small grid to the total character pixels is calculated to obtain a 64-dimensional feature.

[0049] (3) Character peripheral features

[0050] Using the grid divided during extraction in (2), we continue to extract the pixel peripheral features of each small grid in four directions: from top to bottom, from bottom to top, from left to right, and from right to left. The features in the four directions are combined to form a one-dimensional feature. The 64 small grids have a total of 64 dimensions of features.

[0051] After integrating the above three features, the feature dimension is 192 in total. Through principal component analysis, the dimension is reduced to 80-dimensional features. K-means clustering is used, and the correspondence between each character file name and the distance between the centroid of each category is recorded. The characters are sorted within the class according to the distance. The first K characters that are classified into the same category and have a close distance are classified as similar characters to form a similar character set.

[0052] 4. Identify network models

[0053] Convolutional neutral network (CNN) is a neural network specifically designed to process data with a grid structure, such as image data (which can be viewed as a two-dimensional pixel grid). CNN has shown a high recognition rate in two-dimensional image recognition applications, and its network structure is highly invariant to translation, scaling, tilting, or other forms of deformation. CNN directly learns features and classifies characters from the original image, without requiring excessive preprocessing and feature pre-extraction of the original character image. It is an end-to-end recognition system, which effectively avoids the defect of losing similar character details due to manual pre-feature extraction and feature selection.

[0054] Convolutional neural networks consist of alternating convolutional layers and pooling layers, each containing multiple feature maps. In the convolutional layer, each pixel (neuron) is connected to the local area of ​​the previous layer, acting as a local feature detector that can extract basic visual features such as directional segments and corner points. This local connection method makes the network parameters fewer and easier to train. Usually, the convolutional layer is followed by a pooling layer to reduce the image resolution while giving the network a certain degree of translation, scaling, and distortion invariance. In the convolutional layer, the feature map of the previous layer is convolved with multiple sets of convolution kernels, and the feature map of the current layer is generated through an activation function. The calculation process of the convolutional layer can be expressed by the following formula:

[0055]

[0056] in, Represents the number of layers where the convolutional layer is located. is the convolution kernel, the convolution kernel size is 5×5, is the bias, represents the activation function, which is , Represents an input feature map of the previous layer.

[0057] The sampling layer downsamples the feature map of the previous convolutional layer to obtain the same number of feature maps. Convolutional neural networks use stochastic gradient descent, just like traditional neural network training.

[0058] The input layer shows a 28×28 image of ancient Tibetan characters. The C1 layer represents the first convolution layer, which contains 8 24×24 feature maps. A pixel (node ​​or neuron) in each feature map is connected to a 5×5 area corresponding to the input layer. The S1 layer is a downsampling layer containing 8 12×12 feature maps, in which each node is connected to a 2×2 area of ​​the corresponding feature map in the C1 layer. This is followed by C2, which is the second convolution layer and has 16 feature maps, each of which is 8×8 in size. The connection between the S1 and C2 layers plays an important role in the feature extraction process. In addition, S2 represents the second sampling layer, which has 16 feature maps, each of which is 4×4 in size. The final output layer contains 10 nodes, corresponding to the output category, and is fully connected to the S2 layer.

[0059] In the process of network training, error back propagation is adopted and gradient stochastic descent method is used to update the parameters. , .

[0060] use Represents the error function and the parameter update is shown in the following formula.

[0061]

[0062]

[0063] in is the control parameter of the descent rate. The selection is determined by testing methods, and the final selection is the descent rate parameter of this system.

[0064] To test different The impact on recognition rate, first fix other parameters, such as the number of cyclic training rounds is fixed to 30 times, a smaller number of rounds can save training time, but it is enough to reflect Impact on recognition rate.

[0065] 5. Experimental data

[0066] The experimental data in this paper selects two groups of similar character sets below the Tibetan character base line for experiments, each group contains 10 Tibetan character categories. The first group is a group of similar character sets formed by superimposed characters, consisting of " ”, “ ”, “ ”, “ ”, “ ”, “ ”, “ ”, “ ”, “ "and" ", denoted by G1, with a total of 5215 experimental samples.

[0067] The second group is a similar character set consisting of pure consonant characters, consisting of " ”, “ ”, “ ”, “ ”, “ ”, “ ”, “ ”, “ ”, “ "and" ", denoted by G2, with a total of 24,700 experimental samples.

[0068] In order to compare the performance of CNN in Tibetan similar character recognition, CNN is compared with naive Bayesian discriminant classifier and support vector machine classifier. For naive Bayesian discriminant and SVM classifier, the 8-direction gradient features are first extracted to obtain a 64-dimensional feature vector for each sample, and then the classifier is trained using this feature vector. For CNN, the Tibetan character image is directly compressed into a 28×28 resolution image, which reduces the parameters of CNN and thus improves the training speed of the network.

[0069] The experiment uses 64-dimensional gradient features for naive Bayes and SVM classification. K-fold Cross Validation (K is 10) is performed on the G1 and G2 sets, that is, each similarity set is divided into ten parts by category, represented by T1, T2, T3, ..., T10, and one part is taken as the test set each time, and the remaining 9 parts are used as training sets. The Ti row represents the error rate obtained by taking Ti as the test sample and the remaining 9 parts as the training set.

[0070] On this basis, online real-time experiments are carried out, and more sample data collected can be used to further optimize the model.

[0071] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0072] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for recognizing similar characters in ancient Tibetan books in Wujin style based on deep learning, characterized in that: The following steps are involved: S1: Tibetan text estimation preprocessing The preprocessing process includes filtering and enhancing the whole image, as well as tilt and distortion correction, and then segmenting it into different regions, and marking the types of text, image, table or border in the region, and finally outputting different regions such as text, image and table border; S2: Construct similar character set samples Get real ancient Tibetan character samples from ancient Tibetan document images. By segmenting the image after binarization, we can get ancient Tibetan character image samples. Commonly used ancient Tibetan text segmentation algorithms include connected domain clustering, baseline segmentation, and background-based graph segmentation. Through feature extraction and clustering analysis, similar character sample sets are formed through classification and labeling. S3: Feature extraction (1) Gradient 8-directional features First, the image of Tibetan ancient book characters is normalized to 136×50 size. To ensure the image has minimal distortion, bicubic interpolation is used in the deformation process. Then, a 4×2 equally divided grid is used to divide the original image into 8 small grids according to size. The gradient features of the character pixels in each small grid are calculated. The gradient is decomposed into 8 directions according to the Bai ZL method to form an 8-dimensional gradient direction feature. The 8 small grid features are combined to obtain a 64-dimensional gradient direction feature. (2) 8×8 grid density characteristics First, the Tibetan ancient book character image is deformed to 64×64 size. To ensure the image has minimal distortion, bicubic interpolation is used in the deformation process. Then, an 8×8 equally divided grid is used to divide the original image into 64 small grids according to the size. The percentage of character pixels in each small grid to the total character pixels is calculated to obtain a 64-dimensional feature. (3) Character peripheral features Using the grid divided during extraction in (2), continue to extract the pixel peripheral features of each small grid in four directions: from top to bottom, from bottom to top, from left to right, and from right to left. Combine the features in the four directions to form a 1D feature. The total number of features in 64 small grids is 64. Finally, after integrating the above three features, the feature dimension is 192 in total. Through principal component analysis, the dimension is reduced to 80-dimensional features. K-means clustering is used, and the correspondence between each character file name and the distance between each class centroid is recorded. The characters are sorted within the class according to the distance. The first K characters that are classified into the same class and have a close distance are classified as similar characters to form a similar character set. S4: Identify network models The convolutional neural network consists of alternating convolutional layers and pooling layers. Each layer contains multiple feature maps. The calculation process of the convolutional layer can be expressed by the following formula: in, Represents the number of layers where the convolutional layer is located. is the convolution kernel, the convolution kernel size is 5×5, is the bias, represents the activation function, which is , Represents an input feature map of the previous layer; In the process of network training, error back propagation is adopted and gradient stochastic descent method is used to update the parameters. , ; use Represents the error function, and the parameter update is shown as follows: in is the control parameter of the descent rate. The selection is determined by testing methods, and the final selection is the descent rate parameter.

2. The method for recognizing similar characters in ancient Tibetan books in Wujin style based on deep learning according to claim 1, characterized in that: The preprocessed Tibetan ancient book images also need to be binarized.

Images