Handwritten Manchu recognition system, method and equipment for historical Manchu archive images

Abstract

The invention discloses a handwritten Manchu recognition system, method and equipment for a Manchu historical archive image, and belongs to the field of minority document image detection and recognition. The handwritten Manchu identification method based on the segmentation method has accumulative errors. In order to solve the technical problem, the handwritten Manchu recognition method for the Manchu historical archive image is designed and comprises the steps that anisotropic Gaussian filtering is carried out on the Manchu historical archive image, projection is carried out on a horizontal axis, and the number of Manchu columns contained in the Manchu historical archive image is obtained; constructing a Manchu file image feature coding network based on deformable convolution, and obtaining a feature tensor of the Manchu file image; constructing a Manchu file image decoding network based on column feature fusion; training the constructed network model; and identifying by using the trained model and outputting an identification result in a Manchu Latin transcription form. According to the method, accumulative errors caused by handwritten Manchu recognition based on a segmentation method are overcome, the steps of handwritten Manchu recognition are simplified, and the recognition rate of handwritten Manchu is effectively increased.

Description

Technical Field

[0001] The present invention relates to the field of detection and recognition of minority document images, and in particular to a handwritten Manchu recognition system, method and device for Manchu historical archive images. Background Art

[0002] Manchu historical archives are crucial resources for studying Ming and Qing history and Manchu culture. However, due to repeated use, these materials have suffered varying degrees of damage and contamination. In addition to preserving these historical materials, protecting Manchu culture is imperative. Digitizing Manchu historical archives effectively utilizes modern technologies such as computer technology, database technology, and artificial intelligence to preserve these documents and provide diverse research tools. Optical character recognition (OCR) is one of the core technologies used in digitizing historical archives.

[0003] The vast majority of Manchu historical archives are handwritten, and there are problems such as free writing form, adhesion between columns, close distance between words, and breaks between words, which affect the segmentation of Manchu historical archive images. The Manchu recognition method based on segmentation is to first divide the document into columns and then into words, which will produce cumulative errors and seriously affect the Manchu recognition accuracy of Manchu historical archive images.

[0004] The present invention is aimed at the recognition of handwritten Manchu characters in Manchu historical archive images. It uses deep learning technology to realize the recognition of handwritten Manchu characters in non-segmented whole-page Manchu historical archive images. It is one of the key supporting technologies for the digitization of Manchu historical archives. Summary of the Invention

[0005] The purpose of the present invention is to overcome the problem of cumulative error in handwritten Manchu recognition based on the segmentation method, as well as the problem of complicated handwritten Manchu recognition steps and low recognition rate of handwritten Manchu.

[0006] The above purpose is achieved through the following technical solutions:

[0007] A method for recognizing handwritten Manchu characters in Manchu historical archive images is implemented by the following steps:

[0008] Step 1: Perform anisotropic Gaussian filtering on the Manchu historical archive image and project it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image;

[0009] Step 2: Construct a Manchu document image feature encoding network based on deformable convolution and obtain the feature tensor of the Manchu document image;

[0010] Step 3: Construct a Manchu document image decoding network based on column feature fusion;

[0011] Step 4: Train the network model constructed in steps 2 and 3;

[0012] Step 5: Use the Manchu document image feature encoding network and the Manchu document image decoding network trained in step 4 to recognize the Manchu characters in the Manchu document image, and output the recognition results in the form of Manchu Latin transcription.

[0013] Furthermore, the step of performing anisotropic Gaussian filtering on the Manchu historical archive image and projecting it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image is as follows:

[0014] Step 11: Separate the blue single channel component from the original RGB color Manchu historical archive image as a grayscale image;

[0015] Step 1 and 2: Convolve the obtained grayscale image with the anisotropic Gaussian kernel function to obtain the text column response image I(x, y); where the anisotropic Gaussian kernel function is: σ x and σ y Take 3 and 30 respectively, where x and y represent the coordinates on the image;

[0016] Step 13: The anisotropic Gaussian filter response of the Manchu historical archive image obtained in step 12 is accumulated in the Y-axis direction and normalized. The normalized result is Where h is the height of the image;

[0017] Step 14: Perform median filtering on the obtained value of H(x) and calculate the number of its peaks as the number of Manchu columns L contained in the Manchu historical archive image.

[0018] Furthermore, the steps of constructing a deformable convolution-based full-text document image feature encoding network and obtaining a feature tensor of the full-text document image described in step 2 are specifically as follows:

[0019] Step 21: Construct a full-text document image feature encoder based on deformable convolution;

[0020] Step 2. Use the constructed full-text document image feature encoder to obtain the feature tensor of the full-text document image; in the full-text document image feature encoder, the input image passes through a deformable convolution module to obtain a w×h×16 feature tensor, which is expressed as the feature tensor f w×h×16 ;

[0021] Among them, the deformable convolution module includes deformable convolution DConv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU set in sequence;

[0022] Step 2 and 3: Feature tensor f w×h×16 After several convolution modules, the feature tensor w×h×256 is obtained, which is expressed as the feature tensor f w×h×256 ;

[0023] Among them, the convolution module includes deformable convolution Conv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, and activation function LeakyReLU, which are set in sequence.

[0024] Furthermore, the steps of constructing the Manchu document image decoding network based on column feature fusion described in step 3 are as follows:

[0025] The Manchu document image decoding network includes a column feature fusion module and a BiLSTM decoding module, and calculates the loss through the CTC algorithm represents the cumulative CTC error of the recognition results from the first column to the Lth column, k represents the current column, L represents the total number of columns, that is, the result L calculated in step 1; where p k is the predicted data, y k is the label data; CTC's English full name is Connectionist Temporal Classification; among them,

[0026] The input of the column feature fusion module includes: the full document image feature f obtained by the feature extraction network w×h×256 , the hidden layer state parameters of the BiLSTM decoding module And the output result α of the feature fusion module in the previous step t-1，i ,Right now

[0027]

[0028] W f 、W a 、W h represents a learnable parameter; α t-1，k represents the column feature fusion vector calculated at the previous moment, k represents the kth column; t represents the current moment;

[0029] Among them, α t，k Calculated by the following formula:

[0030] e t，k =W s ·S t，k ,

[0031]

[0032] Furthermore, the steps of training the network model constructed in steps 2 and 3 described in step 4 are as follows:

[0033] A dataset of handwritten Manchu column images was constructed, and the models of steps 2 and 3 were trained. During the training of the model in step 2, the obtained parameters were used as a pre-trained model for the recognition of the entire Manchu historical archive image. A hybrid Dropout strategy was adopted during the training process, that is, Dropout and Dropout2d were randomly used.

[0034] A handwritten Manchu character recognition system for Manchu historical archive images, the system comprising:

[0035] The Manchu column number prediction module is used to perform anisotropic Gaussian filtering on the Manchu historical archive image and project it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image;

[0036] A feature tensor acquisition module for Manchurian document images, which is used to construct a Manchurian document image feature encoding network based on deformable convolution and obtain the feature tensor of the Manchurian document image;

[0037] Manchurian document image decoding module, used to build a Manchurian document image decoding network based on column feature fusion;

[0038] Model training module, used to train the network model constructed in steps 2 and 3;

[0039] The output module is used to use the trained Manchu document image feature encoding network and Manchu document image decoding network to recognize Manchu characters in Manchu document images, and output the recognition results in the form of Manchu Latin transcription.

[0040] Furthermore, the Manchu column number prediction module is used to perform anisotropic Gaussian filtering on the Manchu historical archive image and project it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image, as follows:

[0041] The blue single channel component is separated from the original RGB color Manchu historical archive image as a grayscale image. Then, the obtained grayscale image is convolved with the anisotropic Gaussian kernel function to obtain the text column response image I(x, y). The anisotropic Gaussian kernel function is: σ x and σ y Take 3 and 30 respectively, where x and y represent the coordinates on the image. Then, the anisotropic Gaussian filter response of the Manchu historical archive image is accumulated in the Y-axis direction and normalized. The normalized result is Where h is the height of the image. Then, based on the obtained value of H(x), median filtering is performed on it, and the number of peaks is calculated as the number of Manchu columns L contained in the Manchu historical archive image.

[0042] Furthermore, the feature tensor acquisition module of the Manchurian document image is used to construct a Manchurian document image feature encoding network based on deformable convolution, and obtain the feature tensor of the Manchurian document image, as follows:

[0043] A full-text document image feature encoder is constructed based on deformable convolution. Then, the feature tensor of the full-text document image is obtained using the constructed full-text document image feature encoder. In the full-text document image feature encoder, the input image passes through a deformable convolution module to obtain a w×h×16 feature tensor, which is expressed as the feature tensor f w×h×16 ; Afterwards, the feature tensor f w×h×16 After several convolution modules, the feature tensor w×h×256 is obtained, which is expressed as the feature tensor f w×h×256 ; Among them, the deformable convolution module includes the deformable convolution DConvv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU set in sequence; the convolution module includes the deformable convolution Conv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU set in sequence;

[0044] The Manchurian document image decoding module is used to construct a Manchurian document image decoding network based on column feature fusion, as follows:

[0045] The Manchu document image decoding network includes a column feature fusion module and a BiLSTM decoding module, and calculates the loss through the CTC algorithm represents the cumulative CTC error of the recognition results from the first column to the Lth column, k represents the current column, L represents the total number of columns, that is, the result L calculated in step 1; where p k is the predicted data, y k is the label data; CTC's English full name is Connectionist Temporal Classification; among them,

[0046] The input of the column feature fusion module includes: the full document image feature f obtained by the feature extraction network w×h×256 , the hidden layer state parameters of the BiLSTM decoding module And the output result α of the feature fusion module in the previous step t-1，i ,Right now

[0047]

[0048] W f 、W a 、W h represents a learnable parameter; α t-1，k represents the column feature fusion vector calculated at the previous moment, k represents the kth column; t represents the current moment;

[0049] Among them, α t，k Calculated by the following formula:

[0050] e t，k =W s ·S t，k ,

[0051]

[0052] Furthermore, the steps of training the constructed network model are as follows:

[0053] A handwritten Manchu column image dataset was constructed, and the Manchu document image feature encoding network based on deformable convolution and the Manchu document image decoding network model based on column feature fusion were trained. During the training of the Manchu document image feature encoding network model based on deformable convolution, the obtained parameters were used as a pre-training model for the recognition of the entire Manchu historical archive image. A hybrid Dropout strategy was adopted during the training process, that is, Dropout and Dropout2d were randomly used.

[0054] A device for recognizing handwritten Manchu characters in Manchu historical archive images, comprising:

[0055] processor;

[0056] a memory storing executable instructions for the processor;

[0057] The processor is configured to execute the steps of the method for handwritten Manchu recognition for Manchu historical archive images by executing the executable instructions.

[0058] A computer-readable storage medium for storing a program, wherein when the program is executed, the steps of the method for handwritten Manchu recognition for Manchu historical archive images are implemented.

[0059] The beneficial effects of the present invention are:

[0060] The principle of Manchu recognition based on segmentation is to first segment the document into columns and then into words. However, there are errors in the segmentation process, such as Figure 1 This will seriously affect the accuracy of subsequent recognition results. This invention, targeting Manchu historical archival images, overcomes the cumulative errors associated with segmentation-based handwritten Manchu recognition. It employs deep learning technology to achieve unsegmented, full-page handwritten Manchu recognition, simplifying the handwritten Manchu recognition process and effectively improving the recognition rate. This represents one of the key technologies supporting the digitization of Manchu historical archives. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a diagram illustrating a situation in which errors exist in the existing handwritten Manchu word segmentation involved in the present invention;

[0062] Figure 2 It is an example of the Manchu historical archive image of the present invention;

[0063] Figure 3 is a flow chart of the method of the present invention;

[0064] Figure 4 The present invention relates to a diagram of obtaining the number of Manchu columns contained in a Manchu document image;

[0065] Figure 5 It is a feature fusion heat map involved in the present invention;

[0066] Figure 6 Schematic diagram of the network model of the handwritten Manchu character recognition system of the present invention;

[0067] Figure 7 This is a diagram of the pre-training error of a single column model involved in the present invention;

[0068] Figure 8 It is a full-page recognition model training error diagram involved in the present invention;

[0069] Figure 9 This is a diagram of the pre-training loss of a single column model involved in the present invention

[0070] Figure 10 The full-page recognition model involved in the present invention uses a pre-trained model training loss;

[0071] Figure 11 The method of the present invention is used to Figure 2 Compare the diagrams during the identification process;

[0072] Figure 12 The method of the present invention is used to Figure 2 After recognition, the recognition results are output. DETAILED DESCRIPTION

[0073] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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 making creative efforts shall fall within the scope of protection of the present invention.

[0074] Preferred embodiments of the present invention:

[0075] See also Figure 2-Figure 12 , the present invention provides a technical solution:

[0076] A handwritten Manchu recognition method for Manchu historical archive images, such as Figure 3 As shown, the method is implemented by the following steps:

[0077] Step 1: For example Figure 2 The Manchu historical archive image shown in the figure is anisotropically Gaussian filtered and projected on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image, as shown in the figure. Figure 4 As shown;

[0078] Step 2: Construct a Manchu document image feature encoding network based on deformable convolution and obtain the feature tensor of the Manchu document image;

[0079] Step 3: Construct a Manchu document image decoding network based on column feature fusion;

[0080] Step 4: training the network model constructed in steps 2 and 3;

[0081] Step 5: Use the Manchu document image feature encoding network and the Manchu document image decoding network trained in step 4 to recognize the Manchu characters in the Manchu document image, and output the recognition results in the form of Manchu Latin transcription. Specific implementation method 2:

[0083] The present embodiment is a method for recognizing handwritten Manchu characters in Manchu historical archive images. The difference from the first embodiment is that the steps of performing anisotropic Gaussian filtering on the Manchu historical archive image and projecting it on the horizontal axis to obtain the number of Manchu character columns contained in the Manchu archive image are as follows:

[0084] Step 11: Separate the blue single channel component from the original RGB color Manchu historical archive image as a grayscale image;

[0085] Step 1 and 2: Convolve the obtained grayscale image with the anisotropic Gaussian kernel function to obtain the text column response image I(x, y); where the anisotropic Gaussian kernel function is: σ x and σ y Take 3 and 30 respectively, where x and y represent the coordinates on the image;

[0086] Step 13: The anisotropic Gaussian filter response of the Manchu historical archive image obtained in step 12 is accumulated in the Y-axis direction and normalized. The normalized result is Where h is the height of the image;

[0087] Step 14: Perform median filtering on the obtained value of H(x) and calculate the number of its peaks as the number of Manchu columns L contained in the Manchu historical archive image. Specific implementation method three:

[0089] The present embodiment is a method for recognizing handwritten Manchu characters for Manchu historical archival images. The difference from the second embodiment is that the steps of constructing a deformable convolution-based Manchu document image feature encoding network and obtaining a feature tensor of the Manchu document image in step 2 are as follows:

[0090] Step 21: Construct a full-text document image feature encoder based on deformable convolution;

[0091] Step 2. Use the constructed full-text document image feature encoder to obtain the feature tensor of the full-text document image; in the full-text document image feature encoder, the input image passes through a deformable convolution module to obtain a w×h×16 feature tensor, which is expressed as the feature tensor f w×h×16 ;

[0092] Among them, the deformable convolution module includes deformable convolution DConv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU set in sequence;

[0093] Step 2 and 3: Feature tensor f w×h×16 After several convolution modules, the feature tensor w×h×256 is obtained, which is expressed as the feature tensor f w×h×256 ;

[0094] Among them, the convolution module includes deformable convolution Conv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, and activation function LeakyReLU, which are set in sequence. Specific implementation method four:

[0096] The present embodiment of a handwritten Manchu character recognition method for Manchu historical archival images differs from the third embodiment in that the steps of constructing a Manchu character archival image decoding network based on column feature fusion described in step 3 are as follows:

[0097] The Manchu document image decoding network includes a column feature fusion module and a BiLSTM decoding module, and calculates the loss through the CTC algorithm represents the cumulative CTC error of the recognition results from the first column to the Lth column, k represents the current column, L represents the total number of columns, that is, the result L calculated in step 1; where p k is the predicted data, y k is the label data; CTC's English full name is Connectionist Temporal Classification, which is a text recognition algorithm;

[0098] The input of the column feature fusion module includes: the full document image feature f obtained by the feature extraction network w×h×256 , the hidden layer state parameters of the BiLSTM decoding module And the output result α of the feature fusion module in the previous step t-1，i ,Right now

[0099]

[0100] W f 、W a 、W h represents a learnable parameter; α t-1，k represents the column feature fusion vector calculated at the previous moment, k represents the kth column; t represents the current moment;

[0101] Among them, α t，k Calculated by the following formula:

[0102] e t，k =W s ·S t，k ,

[0103] Specific implementation method five:

[0105] The present embodiment of the present invention is a method for recognizing handwritten Manchu characters in Manchu historical archive images. The difference from the fourth embodiment is that the steps of training the network model constructed in steps 2 and 3 in step 4 are as follows:

[0106] A dataset of handwritten Manchu column images was constructed, and the models of steps 2 and 3 were trained. During the training of the model in step 2, the obtained parameters were used as a pre-trained model for the recognition of the entire Manchu historical archive image, accelerating the entire network training process. During the training process, a hybrid Dropout strategy was adopted, that is, Dropout and dropout2d (a function in pyTorch) were randomly used. Specific implementation method six:

[0108] The present embodiment provides a handwritten Manchu character recognition system for Manchu historical archive images, which is applied to the above-mentioned handwritten Manchu character recognition method for Manchu historical archive images. The system includes:

[0109] The Manchu column number prediction module is used to perform anisotropic Gaussian filtering on the Manchu historical archive image and project it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image;

[0110] A feature tensor acquisition module for Manchurian document images, which is used to construct a Manchurian document image feature encoding network based on deformable convolution and obtain the feature tensor of the Manchurian document image;

[0111] The Manchurian document image decoding module is used to construct a Manchurian document image decoding network based on column feature fusion;

[0112] Model training module, used to train the network model constructed in steps 2 and 3;

[0113] The output module is used to use the trained Manchu document image feature encoding network and Manchu document image decoding network to recognize Manchu characters in Manchu document images, and output the recognition results in the form of Manchu Latin transcription. Specific implementation method seven:

[0115] The present embodiment is a handwritten Manchu character recognition system for Manchu historical archival images. The difference from the sixth embodiment is that the Manchu character column number prediction module is used to perform anisotropic Gaussian filtering on the Manchu historical archival image and project it on the horizontal axis to obtain the number of Manchu character columns contained in the Manchu historical archival image, as follows:

[0116] The blue single channel component is separated from the original RGB color Manchu historical archive image as a grayscale image. Then, the obtained grayscale image is convolved with the anisotropic Gaussian kernel function to obtain the text column response image I(x, y). The anisotropic Gaussian kernel function is: σ x and σ y Take 3 and 30 respectively, where x and y represent the coordinates on the image. Then, the anisotropic Gaussian filter response of the Manchu historical archive image is accumulated in the Y-axis direction and normalized. The normalized result is Where h is the height of the image. Then, based on the obtained value of H(x), median filtering is performed on it, and the number of peaks is calculated as the number of Manchu columns L contained in the Manchu historical archive image. Specific implementation method eight:

[0118] The present embodiment is a handwritten Manchu character recognition system for Manchu historical archival images. The difference from the seventh embodiment is that the feature tensor acquisition module of the Manchu document image is used to construct a Manchu document image feature encoding network based on deformable convolution and obtain the feature tensor of the Manchu document image. The details are as follows:

[0119] A full-text document image feature encoder is constructed based on deformable convolution. Then, the feature tensor of the full-text document image is obtained using the constructed full-text document image feature encoder. In the full-text document image feature encoder, the input image passes through a deformable convolution module to obtain a w×h×16 feature tensor, which is expressed as the feature tensor f w×h×16 ; Afterwards, the feature tensor f w×h×16 After several convolution modules, the feature tensor w×h×256 is obtained, which is expressed as the feature tensor f w×h×256 ; Among them, the deformable convolution module includes the deformable convolution DConv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU, etc. set in sequence; the convolution module includes the deformable convolution Conv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU, etc. set in sequence;

[0120] The Manchurian document image decoding module is used to construct a Manchurian document image decoding network based on column feature fusion, as follows:

[0121] The Manchu document image decoding network includes a column feature fusion module and a BiLSTM decoding module, and calculates the loss through the CTC algorithm represents the cumulative CTC error of the recognition results from the first column to the Lth column, k represents the current column, L represents the total number of columns, that is, the result L calculated in step 1; where p k is the predicted data, y k It is the label data; CTC's English full name is Connectionist Temporal Classification, which is a common algorithm;

[0122] The input of the column feature fusion module includes: the full document image feature f obtained by the feature extraction network w×h×256 , the hidden layer state parameters of the BiLSTM decoding module And the output result α of the feature fusion module in the previous step t-1，i ,Right now

[0123]

[0124] W f 、W a 、W h represents a learnable parameter; α t-1，k represents the column feature fusion vector calculated at the previous moment, k represents the kth column; t represents the current moment;

[0125] Among them, α t，k Calculated by the following formula:

[0126] e t，k =W s ·S t，k ,

[0127] Specific implementation method nine:

[0129] The present embodiment is a handwritten Manchu character recognition system for Manchu historical archival images. The difference from the eighth embodiment is that the steps of training the constructed network model are as follows:

[0130] A handwritten Manchu column image dataset was constructed, and the Manchu document image feature encoding network based on deformable convolution and the Manchu document image decoding network model based on column feature fusion were trained. In the process of training the Manchu document image feature encoding network model based on deformable convolution, the obtained parameters were used as a pre-training model in the recognition of the entire Manchu historical archive image, accelerating the entire network training process. A hybrid Dropout strategy was adopted during the training process, that is, Dropout and dropout2d (function in pyTorch) were randomly used. Specific implementation method ten:

[0132] The present embodiment provides a device for recognizing handwritten Manchu characters in Manchu historical archive images, comprising:

[0133] processor;

[0134] a memory storing executable instructions for the processor;

[0135] The processor is configured to execute the steps of the method for handwritten Manchu recognition for Manchu historical archive images by executing the executable instructions. Specific implementation method eleven:

[0137] A computer-readable storage medium of this embodiment is used to store a program, and when the program is executed, the steps of the handwritten Manchu recognition method for Manchu historical archive images are implemented.

[0138] The embodiments disclosed in the present invention are preferred embodiments, but are not limited to them. Ordinary technicians in this field can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. As long as they do not deviate from the spirit of the present invention, they are all within the scope of protection of the present invention.

Claims

1. A method for recognizing handwritten Manchu characters in Manchu historical archive images, characterized by: The method is implemented by the following steps: Step 1: Perform anisotropic Gaussian filtering on the Manchu historical archive image and project it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image; Step 2: Construct a Manchu document image feature encoding network based on deformable convolution and obtain the feature tensor of the Manchu document image; Step 3: Construct a Manchu document image decoding network based on column feature fusion; Step 4: Train the network model constructed in steps 2 and 3; Step 5: Use the Manchu document image feature encoding network and the Manchu document image decoding network trained in step 4 to recognize the Manchu characters in the Manchu document image, and output the recognition results in the form of Manchu Latin transcription.

2. The method for handwritten Manchu character recognition for Manchu historical archive images according to claim 1 is characterized in that: The steps of performing anisotropic Gaussian filtering on the Manchu historical archive image and projecting it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image as described in step 1 are as follows: Step 11: Separate the blue single channel component from the original RGB color Manchu historical archive image as a grayscale image; Step 1 and 2: Convolve the obtained grayscale image with the anisotropic Gaussian kernel function to obtain the text column response image I(x,y); where the anisotropic Gaussian kernel function is: σ x and σ y Take 3 and 30 respectively, where x and y represent the coordinates on the image; Step 13: The anisotropic Gaussian filter response of the Manchu historical archive image obtained in step 12 is accumulated in the Y-axis direction and normalized. The normalized result is Where h is the height of the image; Step 14: Perform median filtering on the obtained value of H(x) and calculate the number of its peaks as the number of Manchu columns L contained in the Manchu historical archive image.

3. The method for handwritten Manchu character recognition for Manchu historical archive images according to claim 2 is characterized in that: The steps of constructing a deformable convolution-based full-text document image feature encoding network and obtaining the feature tensor of the full-text document image described in step 2 are as follows: Step 21: Construct a full-text document image feature encoder based on deformable convolution; Step 2. Use the constructed full-text document image feature encoder to obtain the feature tensor of the full-text document image; in the full-text document image feature encoder, the input image passes through a deformable convolution module to obtain a w×h×16 feature tensor, which is expressed as the feature tensor f w×h×16 ; Among them, the deformable convolution module includes deformable convolution DConv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU set in sequence; Step 2 and 3: Feature tensor f w×h×16 After several convolution modules, the feature tensor w×h×256 is obtained, which is expressed as the feature tensor f w×h×256 ; Among them, the convolution module includes deformable convolution Conv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, and activation function LeakyReLU, which are set in sequence.

4. The method for recognizing handwritten Manchu characters in Manchu historical archive images according to claim 3 is characterized in that: The steps for constructing a Manchu document image decoding network based on column feature fusion described in step 3 are as follows: The Manchu document image decoding network includes a column feature fusion module and a BiLSTM decoding module, and calculates the loss through the CTC algorithm represents the cumulative CTC error of the recognition results from the first column to the Lth column, k represents the current column, L represents the total number of columns, that is, the result L calculated in step 1; where p k is the predicted data, y k is the label data; CTC's English full name is Connectionist Temporal Classification; among them, The input of the column feature fusion module includes: the full document image feature f obtained by the feature extraction network w×h×256 , the hidden layer state parameters of the BiLSTM decoding module And the output result α of the feature fusion module in the previous step t-1，i ,Right now W f 、W a 、W h represents a learnable parameter; α t-1，k represents the column feature fusion vector calculated at the previous moment, k represents the kth column; t represents the current moment; Among them, α t，k Calculated by the following formula: e t，k =W s ·S t，k , 5. The method for recognizing handwritten Manchu characters in Manchu historical archive images according to claim 4 is characterized in that: The steps for training the network model constructed in steps 2 and 3 described in step 4 are as follows: A dataset of handwritten Manchu column images was constructed, and the models of steps 2 and 3 were trained. During the training of the model in step 2, the obtained parameters were used as a pre-trained model for the recognition of the entire Manchu historical archive image. A hybrid Dropout strategy was adopted during the training process, that is, Dropout and Dropout2d were randomly used.

6. A handwritten Manchu character recognition system for Manchu historical archive images, applied to the handwritten Manchu character recognition method for Manchu historical archive images according to any one of claims 1 to 5, characterized in that: The system comprises: The Manchu column number prediction module is used to perform anisotropic Gaussian filtering on the Manchu historical archive image and project it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image; A feature tensor acquisition module for Manchurian document images, which is used to construct a Manchurian document image feature encoding network based on deformable convolution and obtain the feature tensor of the Manchurian document image; Manchurian document image decoding module, used to build a Manchurian document image decoding network based on column feature fusion; Model training module, used to train the network model constructed in steps 2 and 3; The output module is used to use the trained Manchu document image feature encoding network and Manchu document image decoding network to recognize Manchu characters in Manchu document images, and output the recognition results in the form of Manchu Latin transcription.

7. The handwritten Manchu character recognition system for Manchu historical archive images according to claim 6 is characterized in that: The Manchu column number prediction module is used to perform anisotropic Gaussian filtering on the Manchu historical archive image and project it on the horizontal axis to obtain the number of Manchu columns contained in the Manchu archive image, as follows: The blue single channel component is separated from the original RGB color Manchu historical archive image as a grayscale image. Then, the obtained grayscale image is convolved with the anisotropic Gaussian kernel function to obtain the text column response image I(x,y). The anisotropic Gaussian kernel function is: σ x and σ y Take 3 and 30 respectively, where x and y represent the coordinates on the image. Then, the anisotropic Gaussian filter response of the Manchu historical archive image is accumulated in the Y-axis direction and normalized. The normalized result is Where h is the height of the image. Then, based on the obtained value of H(x), median filtering is performed on it, and the number of peaks is calculated as the number of Manchu columns L contained in the Manchu historical archive image.

8. The handwritten Manchu character recognition system for Manchu historical archive images according to claim 7 is characterized in that: The feature tensor acquisition module of the Manchurian document image is used to construct a Manchurian document image feature encoding network based on deformable convolution, and obtain the feature tensor of the Manchurian document image, as follows: A full-text document image feature encoder is constructed based on deformable convolution. Then, the feature tensor of the full-text document image is obtained using the constructed full-text document image feature encoder. In the full-text document image feature encoder, the input image passes through a deformable convolution module to obtain a w×h×16 feature tensor, which is expressed as the feature tensor f w×h×16 ; Afterwards, the feature tensor f w×h×16 After several convolution modules, the feature tensor w×h×256 is obtained, which is expressed as the feature tensor f w×h×256 ; Among them, the deformable convolution module includes the deformable convolution DConv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU set in sequence; the convolution module includes the deformable convolution Conv, activation function LeakyReLU, convolution Conv, activation function LeakyReLU, normalization InstanceNorm2d, convolution Conv, activation function LeakyReLU set in sequence; The Manchurian document image decoding module is used to construct a Manchurian document image decoding network based on column feature fusion, as follows: The Manchu document image decoding network includes a column feature fusion module and a BiLSTM decoding module, and calculates the loss through the CTC algorithm represents the cumulative CTC error of the recognition results from the first column to the Lth column, k represents the current column, L represents the total number of columns, that is, the result L calculated in step 1; where p k is the predicted data, y k is the label data; CTC's English full name is Connectionist Temporal Classification; among them, The input of the column feature fusion module includes: the full document image feature f obtained by the feature extraction network w×h×256 , the hidden layer state parameters of the BiLSTM decoding module And the output result α of the feature fusion module in the previous step t-1，i ,Right now W f 、W a 、W h represents a learnable parameter; α t-1，k represents the column feature fusion vector calculated at the previous moment, k represents the kth column; t represents the current moment; Among them, α t，k Calculated by the following formula: e t，k =W s ·S t，k , 9. The handwritten Manchu character recognition system for Manchu historical archive images according to claim 8 is characterized in that: The steps of training the constructed network model are as follows: A handwritten Manchu column image dataset was constructed, and the Manchu document image feature encoding network based on deformable convolution and the Manchu document image decoding network model based on column feature fusion were trained. During the training of the Manchu document image feature encoding network model based on deformable convolution, the obtained parameters were used as a pre-training model for the recognition of the entire Manchu historical archive image. A hybrid Dropout strategy was adopted during the training process, that is, Dropout and Dropout2d were randomly used.

10. A device for recognizing handwritten Manchu characters in Manchu historical archive images, characterized in that: include: processor; a memory storing executable instructions for the processor; The processor is configured to execute the steps of the handwritten Manchu recognition method for Manchu historical archival images as described in any one of claims 1 to 5 by executing the executable instructions.

Images