A deep learning-based dot matrix character recognition method

Abstract

This invention belongs to the field of machine vision, specifically a deep learning-based dot matrix character recognition method, comprising the following steps: 1) acquiring the font and size of characters from an inkjet printer, creating a dot matrix sequence, and generating dot matrix character images in a newly created background image according to the dot matrix sequence; 2) adding random elastic deformation to each character image in the dot matrix character images to obtain a dataset containing multiple character images; 3) fusing each character image as a foreground image with the background image to obtain a fused image; 4) using the fused image to establish an image dataset, constructing a network training model, and training the network model based on the image dataset, inputting the image to be recognized into the network training model for text detection and recognition to obtain the character. This invention establishes a new network structure by inserting a Spatial Variation Network (STN) into a Spatial Convolutional Recurrent Neural Network (CRNN), enabling accurate recognition of dot matrix characters.

Description

Technical Field

[0001] This invention belongs to the field of machine vision, specifically a dot matrix character recognition method based on deep learning. Background Technology

[0002] Product information for food products is often printed on packaging in the form of dot matrix characters. To ensure food safety, it is essential to guarantee the accuracy of production information such as origin, batch number, and production date. To improve product quality and reduce market complaints, food factories often manually inspect printed information, which suffers from low accuracy, inefficiency, high labor costs, and lack of data tracking. Therefore, researching a fast and accurate dot matrix character recognition method can improve product quality and reduce enterprise costs.

[0003] Currently, dot matrix character recognition methods are mainly divided into three categories: template matching-based methods, machine learning-based methods, and deep learning-based methods. Template matching and support vector machine-based methods first perform grayscale template matching and feature template matching. If the matching results differ, an SVM (Support Vector Machine) is used for judgment to obtain the final result. However, this method performs poorly in complex backgrounds. Machine learning-based methods utilize backpropagation (BP) neural networks to recognize characters, improving the accuracy of character recognition in complex backgrounds to some extent. However, BP neural networks suffer from slow learning speed and limited generalization ability. Deep learning-based methods use convolutional neural networks for character recognition, further improving the algorithm's recognition ability. However, they still suffer from poor character segmentation, complex processes, long processing times, and low accuracy.

[0004] Due to factors such as wrinkles, seams, and printer performance, dot matrix characters on food packaging may experience localized deformation, blurring, tilting, or wavy shapes. These characteristics of dot matrix characters further increase the difficulty of character recognition, resulting in low accuracy and poor applicability of some commonly used traditional character recognition methods, which cannot meet user needs. Summary of the Invention

[0005] To address the shortcomings of the aforementioned technologies, this invention proposes a deep learning-based dot matrix character recognition method. This method can quickly and accurately recognize dot matrix characters against complex background patterns, overcoming problems such as complex background patterns, large differences in character size, character deformation, and different fonts, achieving fast and high-precision dot matrix character recognition. Simultaneously, a dot matrix character generation method is proposed, which can automatically generate a large amount of character data, expanding the dataset for network training and enabling the rapid application of the deep learning-based dot matrix character recognition method.

[0006] The technical solution adopted by this invention to achieve the above objectives is: a dot matrix character recognition method based on deep learning, comprising the following steps:

[0007] 1) Obtain the font and size of the characters for the inkjet printer, create a dot matrix sequence using HALCON software, and generate dot matrix character images in a newly created background image according to the dot matrix sequence;

[0008] 2) Add random elastic deformation to each character image in the dot matrix character image to obtain a dataset containing multiple character images;

[0009] 3) Fuse each character image as a foreground image with the background image to obtain the fused image;

[0010] 4) Use the fused images to build an image dataset, construct a network training model, and train the network model based on the image dataset. Input the image to be recognized into the network training model to perform text detection and recognition to obtain characters.

[0011] Step 2) specifically includes:

[0012] 2-1) For each character in the image, add random elastic deformation. By controlling the factor β and using a random function, establish a random displacement field (Pr, Pc), that is:

[0013] Pr=(rand(1)-0.5)*2*β (1)

[0014] Pc=(rand(1)-0.5)*2*β (2)

[0015] Where rand(1) is a 0-1 random function, Pr is the random offset of the pixel in the row direction, and Pc is the random offset of the pixel in the column direction.

[0016] 2-2) The image is smoothed using Gaussian filtering to obtain the final displacement field, i.e.:

[0017]

[0018] Among them, G σ (x) represents the final displacement field, σ represents the standard deviation, x represents the x-coordinate of the pixel after adding random elastic deformation, and y represents the y-coordinate of the pixel after adding random elastic deformation.

[0019] 2-3) After filling the missing pixels in the neighborhood of the elastically deformed image using bilinear interpolation, set the same offset for each connected component in the character image, i.e.:

[0020] Pr n =max(Pr), Pr∈RegionR (4)

[0021] Pc n =max(Pc), Pc∈RegionC (5)

[0022] Where RegionR and RegionC represent the offsets of all pixels within the nth black dot region; max(Pr) represents taking the maximum value; Pr n The maximum value of the row direction offset of all pixels within the nth black dot region is used as the row direction offset of the nth black dot; Pc n The maximum value of the column direction offset of all pixels within the nth black dot region is used as the column direction offset of the nth black dot.

[0023] 2-4) For character images that are tilted forward or backward, random forward or backward tilt deformation is added using the following formula:

[0024] Pc=(Height-R)*β (6)

[0025] Where Pc is the column offset, Height is the image height, and R is the row vector.

[0026] Step 3) specifically includes:

[0027] Based on the spatial relationship between pixels and the center of the black dot region, an alpha channel image is created. The closer to the center of the black dot region, the larger α is, and the larger the proportion of foreground image pixels. The alpha channel image is then used to fuse multiple character images.

[0028] g dst =g c *α+g b *(1-α)+δ (7)

[0029] Among them, g dst For the fused image, g c For a character image, α is the weighting coefficient of the character image, and g b δ represents the background image, and δ represents the offset.

[0030] The construction of the network training model includes the following steps:

[0031] When generating each fused image, a corresponding character bounding box is generated to form an image dataset. Using the image dataset, DBNET is used for training to obtain the trained DBNET network model. The DBNET network model is used to detect the input image to obtain the position of the character bounding box, and the character bounding box region is cropped as the character bounding box image.

[0032] A dataset consisting of character bounding box images and labeled characters is fed into the recognition network for training, resulting in a network training model.

[0033] Step 4): Based on the trained model, perform text detection and recognition on the image to obtain characters, including the following steps:

[0034] The image to be recognized is input into the DBNET network model to obtain segmented character bounding box images containing characters;

[0035] Input the character frame image into the recognition network and obtain the characters recognized by the recognition network.

[0036] The identification network is constructed through the following steps:

[0037] The recognition network mainly consists of sequentially connected CNNs and RNNs. A spatial variation network (STN) is inserted between the input image and the CNN to correct the input image, or it is inserted into the CNN to correct the feature map. The final output is an image or feature map with random and specific deformations removed. The RNN predicts each frame of the feature sequence and converts it into the final label character sequence, which is the final recognition dot matrix character.

[0038] The present invention has the following beneficial effects and advantages:

[0039] 1. This invention uses DBNET for text detection, which not only has a faster detection speed, but can also detect dot matrix character text instances of arbitrary shapes;

[0040] 2. This invention establishes a new network structure by inserting the Spatial Variation Network (STN) into the Spatial Convolutional Recurrent Neural Network (CRNN), which enables accurate recognition of dot matrix characters.

[0041] 3. This invention proposes a novel method for synthesizing dot matrix characters, which closely resemble images acquired during actual production processes. This allows for rapid dataset expansion, saving annotation time and facilitating network training, thus enabling the rapid application of deep learning-based dot matrix character recognition methods. An improved CRNN dot matrix character recognition method is proposed. First, DBNET is used for text detection. Then, combining the advantages of CRNN and STN, STN is inserted into the CRNN to complete character recognition. This method can quickly and accurately recognize dot matrix characters against complex pattern backgrounds, achieving an accuracy rate of up to 99.5%. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the overall method of the present invention;

[0043] Figure 2a The dot matrix character image generated by this invention;

[0044] Figure 2b The random vector field established for this invention;

[0045] Figure 2c According to the present invention Figure 2b Character images subjected to random elastic deformation by a random vector field;

[0046] Figure 2d The back-tendency field established for this invention;

[0047] Figure 2e The character image of this invention is subjected to backtiling deformation according to the 2d backtiling field;

[0048] Figure 3a These are images collected during actual production processes in embodiments of the present invention;

[0049] Figure 3b For the present invention and Figure 3a Images with the same characters on different backgrounds;

[0050] Figure 3c For the present invention and Figure 3b Images with the same background but different characters;

[0051] Figure 4 This is a flowchart of the DBNET text detection process of the present invention;

[0052] Figure 5 This is the DBNET text detection result of the present invention;

[0053] Figure 6 This invention proposes a deep learning network for recognition results. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0055] The relationships between the modules of this invention are as follows: Figure 1 As shown, the proposed framework mainly consists of three steps. The first step is to generate a dataset, the second step is to segment characters, and the third step is to recognize characters.

[0056] The present invention specifically includes the following steps:

[0057] 1) Obtain the font and size of the characters for the inkjet printer, create a dot matrix sequence using HALCON software, and generate dot matrix character images in a newly created background image according to the dot matrix sequence;

[0058] 2) Add random elastic deformation to each character image in the dot matrix character image to obtain a dataset containing multiple character images;

[0059] 3) Fuse each character image as a foreground image with the background image to obtain the fused image;

[0060] 4) Use the fused images to build an image dataset, construct a network training model, and train the network model based on the image dataset. Input the image to be recognized into the network training model to perform text detection and recognition to obtain characters.

[0061] As in steps 1) to 3), the specific steps are as follows:

[0062] Step 1.1: Generate a standard dot matrix character image based on the font and size of the characters to be recognized, such as... Figure 2a As shown.

[0063] Step 1.2: To better approximate the deformation caused by unevenness in the packaging, random elastic deformation is added. A random displacement field is created according to formulas 1 and 2. The control factor β is multiplied by the random displacement field to obtain the displacement field (Pr, Pc), which is used to control the deformation intensity. Formula 3 uses Gaussian filtering to smooth the image, obtaining the final displacement field. The larger σ is, the smoother the resulting image. After elastic deformation, some locations in the image may lack appropriate pixel values. These pixels are then filled using bilinear interpolation to ensure the image matches in the neighborhood.

[0064] Pr=(rand(1)-0.5)*2*β (1)

[0065] Pc=(rand(1)-0.5)*2*β (2)

[0066]

[0067] Where rand(1) is a function that generates a random number between 0 and 1, Pr is the random offset of the pixel in the row direction, Pc is the random offset of the pixel in the column direction, and G... σ (x) represents the final displacement field, σ represents the standard deviation, x represents the x-coordinate of the pixel after adding random elastic deformation, and y represents the y-coordinate of the pixel after adding random elastic deformation.

[0068] The above method for elastic deformation of the entire image yields excellent results for regular characters, but for bitmap characters, it causes severe distortion of the printed black dots. To address this issue, an automatic bitmap character production method based on an improved elastic deformation algorithm is proposed. This method sets the same offset for each connected component, as shown in formulas 4 and 5. RegionR and RegionC represent the offsets of all pixels within the nth black dot region. This random elastic deformation can well approximate the deformation caused by unevenness in the packaging. Figure 2b Random vector fields Figure 2c According to Figure 2b Character images subjected to random elastic deformation by a random vector field. To represent characters tilting forward or backward, tilt deformation can be added. To improve training performance, specific deformations can be added to the training set, as shown in Equation 6. Figure 2d Back-tendency field, Figure 2e It is a dot matrix character image that undergoes backtiling deformation according to a 2D backtiling field. This not only increases the amount of data but also better simulates the actual situation, making the generated dataset as close as possible to the real dataset.

[0069] Pr n =max(Pr), Pr∈RegionR (4)

[0070] Pc n =max(Pc), Pc∈RegionC (5)

[0071] Pc=(Height-R)*β (6)

[0072] Where RegionR and RegionC represent the offsets of all pixels within the nth black dot region; max(Pr) represents taking the maximum value; Pr n The maximum value of the row direction offset of all pixels within the nth black dot region is used as the row direction offset of the nth black dot; Pc n The maximum value of the column direction offset of all pixels within the nth black dot region is used as the column direction offset of the nth black dot. Pc is the column direction offset, Height is the image height, and R is the row vector.

[0073] To achieve the goal of creating a dataset that closely resembles actual production conditions, it is also necessary to fuse the foreground and background images together. Since the character images are automatically generated by an algorithm, an alpha channel image is created during the character image generation process based on the spatial relationship between pixels and the nearest printed center. Image fusion calculations are then performed using the alpha channel image, as shown in Formula 7. The generated image is highly similar to the real image, with random deformation and backward tilting added to the raster image, achieving excellent results. Figure 2b Random vector fields Figure 2c According to Figure 2b A dot matrix character image subjected to random elastic deformation by a random vector field. To represent characters tilting forward or backward, a tilt deformation can be added. Figure 2d Back-tendency field, Figure 2e It is a dot matrix character image that undergoes backtiling deformation according to a 2D backtiling field. This not only increases the amount of data but also better simulates the actual situation, making the generated dataset as close as possible to the real dataset.

[0074] gdst =g c *α+g b *(1-α)+δ (7)

[0075] Among them, g dst For the fused image, g c For a character image, α is the weighting coefficient of the character image, and g b δ represents the background image, and δ represents the offset.

[0076] Step 1.3: After preparing the character image, i.e., the foreground image, an image of the food packaging is acquired as the background image. Then, the foreground and background images are merged together. Since the character image is automatically generated by an algorithm, an alpha channel image is created during the character image generation process based on the spatial relationship between pixels and the nearest printed circle center. The alpha channel image is then used for image fusion calculation. The fused image result is shown in Figure 3. Figure 3a These are images collected during the actual production process; Figure 3b is a result of this process. Figure 3a Images with the same characters on different backgrounds, generated Figure 3c Is and Figure 3b Images with the same background but different characters. As shown in Figure 3, the generated images are highly similar to the real images, achieving excellent results. Therefore, the model trained on the generated dataset can be perfectly applied to real-world production processes.

[0077] In step 2, DBNET is used for text detection, specifically:

[0078] Building a network training model includes the following steps:

[0079] When generating each fused image, a corresponding character bounding box is generated to form an image dataset. Using the image dataset, DBNET is used for training to obtain the trained DBNET network model. The DBNET network model is used to detect the input image to obtain the position of the character bounding box, and the character bounding box region is cropped as the character bounding box image.

[0080] A dataset consisting of character bounding box images and labeled characters is fed into the recognition network for training, resulting in a trained recognition network.

[0081] The DBNET network model proposes a robust and fast scene text detector by combining a simple semantic segmentation network with the proposed DB module. The DB module aims to insert binarization operations into a segmentation network for joint optimization. However, standard binarization functions are non-differentiable. Equation 8 shows a differentiable binarization (DB) module that approximates standard binarization to make it differentiable, integrating it into training to obtain more accurate boundaries and significantly reduce post-processing time. It achieves better results on multiple standard scene text datasets, including text of various shapes, while also being much faster.

[0082]

[0083] Where k is the expansion factor, P i,j T refers to the pixels in the probability map. i,j This refers to the pixels in the threshold map, B i,j It refers to the pixels in the binary image;

[0084] 4-1) Extract feature map F from the fused image: Input the fused image into the feature pyramid backbone network, upsample the pyramid features to the same scale, and cascade them to generate feature F.

[0085] First, the image is input into a ResNet50_vd backbone network with an FPN structure to extract feature maps. Then, the top-level features are upsampled and fused with the bottom-level features, and then upsampled to 1 / 4. Finally, all the resulting 1 / 4 feature maps are fused to obtain feature map F, as shown below. Figure 4 As shown.

[0086] 4-2) Predicting probability map P and threshold map T using feature map F. Deep learning generally involves three steps: first, preparing the dataset; second, building your own network; and third, training your model. Now that we have trained the model using the dataset, this section describes the applicable process: input image, feature extraction, and the trained network predicting probability map P and threshold map T.

[0087] Then, the feature map F is used to predict the probability map P and the threshold map T. Finally, an approximate binary mapping (B) is calculated using P and T. During the training phase, P, T, and B are trained under supervised conditions, with P and B using the same supervisory signal. During the inference phase, the text box can be obtained using either P or B. Figure 5 The results of this paper's algorithm show that it overcomes the interference of complex background patterns and text, and accurately detects the specific location of the character to be identified.

[0088] 4-3) Obtain the binary mapping B based on the probability map P and the threshold map T, as shown in Formula 8. When P i,j >Ti,j At that time, B i,j ≈1, indicating that the pixel is considered the target; when P i,j <T i,j At that time, B i,j If the value is approximately 0, the pixel is considered to be part of the background.

[0089] In step 4), the Spatial Variation Network (STN) is inserted into the Spatial Convolutional Recurrent Neural Network (CRNN) to establish a new network structure, specifically as follows:

[0090] The recognition network mainly consists of sequentially connected CNNs and RNNs. A spatial variation network (STN) is inserted between the input image and the CNN to correct the input image, or it is inserted into the CNN to correct the feature map. The final output is an image or feature map with random and specific deformations removed. The RNN predicts each frame of the feature sequence and transforms it into the final label character sequence, which is the final recognition dot matrix character.

[0091] The CRNN character recognition network integrates the advantages of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). Compared with traditional methods and other CNN and RNN-based algorithms, CRNN exhibits superior performance and has been widely applied in text recognition, demonstrating outstanding performance on many datasets. The key technologies of CRNN consist of three parts: extracting image convolutional features using CNNs; further extracting text sequence features based on the convolutional features using a deep bidirectional LSTM network; and using Connectionist Temporal Classification (CTC) to address the problem of character misalignment during training.

[0092] A regular character, as a whole, has all its regions forming a connected domain, making it easier to treat as a single character during the learning process. However, dot matrix characters are composed of multiple isolated dots, which can easily lead to missegmentation and misrecognition, especially when there is tilting or elastic deformation. Therefore, this invention utilizes the advantages of STN (Single-Dimensional Network) to first perform image correction and then extract features to recognize characters.

[0093] The STN spatial transformation network has the following advantages: it can be inserted into any existing convolutional structure; it allows unsupervised training, enabling the neural network to actively transform the feature map spatially based on its own characteristics. The spatial transformation network allows the model to learn spatial transformation invariance (including translation, scaling, rotation, and distortion), improving neural network performance. The STN spatial transformation network mainly consists of three parts: parameter prediction, which generates transformation parameters θ based on the characteristics of the feature map; coordinate mapping, which calculates the correspondence between pixel coordinates using the transformation parameters obtained in the previous step; and pixel sampling calculation, which uses bilinear interpolation to calculate the pixel value of each coordinate point in the output feature map.

[0094] To clearly explain this network, assume that spatial transformations only involve affine transformations. In linear algebra, affine transformations such as translation H(T), rotation H(θ), and scaling H(S) of an image can all be represented using matrix operations, as shown in Equations 9 and 10. The matrix H is obtained after calculation, where ε={ω 11 ,ω 12 ,ω 13 ,ω 21 ,ω 22 ,ω 23 The six variables in} are parameters used to map the relationship between the coordinate points of the input and output graphs.

[0095]

[0096]

[0097] Formula 11 represents the Grid generator process, which calculates the correspondence between pixel coordinates.

[0098]

[0099] in, The input feature map contains the coordinates of the pixels. This is used to output the coordinates of the pixels in the feature map.

[0100] Formula 12 represents the Sample process. It is the pixel value of channel c at position (n, m) of the input feature map. It is the output feature map at the location The pixel value of channel c is used. The output feature map V is calculated using bilinear interpolation.

[0101]

[0102]

[0103] Where i is the pixel number and c is the channel number. It is the pixel value of channel c at position (n,m) of the input feature map. It is the output feature map at the location The pixel value of channel c corresponds to the input feature map. The pixel values ​​of channel c, since their coordinates are all integers, are not present in the input feature map. For the corresponding points, the bilinear interpolation method is used to search for points (n, m) that meet the conditions, and then the final pixel value is calculated.

[0104] Combining the characteristics of STN and CRNN, this invention inserts STN into CRNN. ​​The input image is deformed by STN, the backbone network extracts features, and the features are fed into the recurrent layer. Bidirectional RNN (BLSTM) is used to predict the feature sequence, learn each feature vector in the sequence, and output the predicted label distribution. Finally, the results are fed into the transcription layer, and CTC loss is used to transform the series of label distributions obtained from the recurrent layer into the final label sequence.

[0105] In summary, this method boasts advantages such as high recognition speed and high accuracy, and performs exceptionally well in handling complex problems such as intricate background patterns and inconsistent character sizes and fonts. The main steps include: proposing a novel method for synthesizing dot matrix characters, which can manually synthesize large datasets and labeled files without the need for annotations or existing datasets, allowing for rapid application to new products and addressing the problem of dataset scarcity; employing DBNET for dot matrix character text detection, which not only runs quickly but can also detect dot matrix character text instances of arbitrary shapes; and establishing a new network structure by inserting the Spatial Variation Network (STN) into the Spatial Convolutional Recurrent Neural Network (CRNN), enabling accurate recognition of dot matrix characters while maintaining high efficiency.

[0106] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, extensions, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A dot matrix character recognition method based on deep learning, characterized in that, Includes the following steps: 1) Obtain the font and size of the characters for the inkjet printer, create a dot matrix sequence using HALCON software, and generate dot matrix character images in a newly created background image according to the dot matrix sequence; 2) Add random elastic deformation to each character image in the dot matrix character image to obtain a dataset containing multiple character images; Step 2) specifically refers to: 2-1) For each character in the image, add random elastic deformation, controlled by a factor. Establish a random displacement field with a random function. ,Right now: ； ； in, To generate a random function between 0 and 1, This represents the random offset of the pixel in the row direction. This represents the random offset of the pixel along the column direction. 2-2) The image is smoothed using Gaussian filtering to obtain the final displacement field, i.e.: ； in, For the final displacement field, Standard deviation, To add the x-coordinate of the pixel after random elastic deformation, Add the ordinate of the pixel after random elastic deformation; 2-3) After filling the missing pixels in the neighborhood of the elastically deformed image using bilinear interpolation, set the same offset for each connected component in the character image, i.e.: ； C； in, , C represents the offset of all pixels within the nth black dot region; This indicates taking the maximum value; The maximum value of the row direction offset of all pixels within the nth black dot region is used as the row direction offset of the nth black dot. The maximum value of the column direction offset of all pixels within the nth black dot region is used as the column direction offset of the nth black dot. 2-4) For character images that are tilted forward or backward, random forward or backward tilt deformation is added using the following formula: ； in, This is the column direction offset. Image height, It is a row vector; 3) Fuse each character image as a foreground image with the background image to obtain the fused image; Step 3) specifically involves: Based on the spatial relationship between pixels and the center of the black dot region, an alpha channel image is created. The closer to the center of the black dot region, the better. The larger the value, the larger the proportion of foreground image pixels. Image fusion of multiple character images is performed using the alpha channel image, i.e.: ； in, For the merged image, For character images, These are the weighting coefficients for the character image. Background image, This is the offset; 4) Use the fused images to build an image dataset, construct a network training model, and train the network model based on the image dataset. Input the image to be recognized into the network training model to perform text detection and recognition to obtain characters.

2. The dot matrix character recognition method based on deep learning according to claim 1, characterized in that, The construction of the network training model includes the following steps: When generating each fused image, a corresponding character bounding box is generated to form an image dataset. Using the image dataset, DBNET is used for training to obtain the trained DBNET network model. The DBNET network model is used to detect the input image to obtain the position of the character bounding box, and the character bounding box region is cropped as the character bounding box image. A dataset consisting of character bounding box images and labeled characters is fed into the recognition network for training, resulting in a network training model.

3. The dot matrix character recognition method based on deep learning according to claim 1, characterized in that, The process of performing text detection and recognition on the image to be recognized to obtain characters includes the following steps: The image to be recognized is input into the DBNET network model to obtain segmented character bounding box images containing characters; Input the character frame image into the recognition network and obtain the characters recognized by the recognition network.

4. The dot matrix character recognition method based on deep learning according to claim 2, characterized in that, The identification network is constructed through the following steps: The recognition network mainly consists of sequentially connected CNNs and RNNs. A spatial variation network (STN) is inserted between the input image and the CNN to correct the input image, or it is inserted into the CNN to correct the feature map. The final output is an image or feature map with random and specific deformations removed. The RNN predicts each frame of the feature sequence and transforms it into the final label character sequence, which is the final recognition dot matrix character.

Images