Network image text recognition method and system based on multi-scale feature fusion
By using U-net network to extract multi-scale features and combining full convolution network and rotation module to perform affine transformation of text feature maps, the poor effect of complex background and multi-direction text recognition in the prior art is solved. Problem, achieving more efficient and accurate text recognition.
Patent Information
- Application Number
- CN202111506810.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-10
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2041-12-10
AI Technical Summary
The existing text recognition methods are not effective when processing network images with complex backgrounds and diverse text directions, and a single detection network and recognition network cannot effectively share visual clues, resulting in high time cost and low recognition efficiency.
U-net multi-scale feature extraction network based on residual network is adopted, combined with the full convolutional network output text prediction box and rotation module for affine transformation, realize the axial alignment of the text feature map, and integrate the text detection and recognition modules. End-to-end identification.
It improves the robustness and accuracy of the recognition of texts of different scales and directions, reduces the amount of data required for training, and improves the efficiency and accuracy of the system.
Smart Images

Figure CN114359906B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of text recognition, and in particular to a network image text recognition method and system based on multi-scale feature fusion. Background Art
[0002] Text recognition is a technology that uses computers to automatically recognize characters. It is an important field of computer vision task application. In the field of computer vision, text recognition has always been an important task because text is an important clue that provides richer and more accurate information for computer vision tasks. At present, text recognition technology has been widely used in various scenarios in life, such as text verification code application, document analysis and mathematical formula recognition.
[0003] At present, text recognition methods are divided into two categories. The first category is to divide it into two independent subtasks: text detection and text recognition. Deep learning-based methods dominate both subtasks. In text detection, convolutional neural networks are usually used to extract feature maps from scene images, and then different decoders are used to decode text regions. Text recognition, on the other hand, performs sequence prediction based on predicted text regions. However, this type of method requires a lot of time cost when recognizing images with many text regions. At the same time, this type of method ignores the relevance of visual cues shared in detection and recognition. A single detection network cannot be supervised by labels from text recognition, and similarly, a single recognition network cannot be supervised by labels from the detection network.
[0004] The second type of method is the end-to-end text recognition method, which considers the mutual supervision of text detection and recognition at the same time. Compared with previous methods, this type of method learns more general features through convolutional neural networks. These features are shared between text detection and text recognition, and the supervision of the two tasks is complementary. However, the current end-to-end methods for text recognition in mixed Chinese and English scenes mainly focus on horizontal text with simple backgrounds and relatively regular layouts, and are not applicable to network image text with complex backgrounds and diverse text directions. Summary of the invention
[0005] In order to overcome the deficiencies of the above technologies, the present invention provides a network image text recognition method based on multi-scale feature fusion, and also provides a network image text recognition system based on multi-scale feature fusion.
[0006] In the present invention, a U-net multi-scale feature extraction network based on a residual network is first used to extract the text feature map of the original image, and then a fully convolutional network is used to output the text prediction box and the horizontal angle of the text prediction box. The rotation module then performs an affine transformation on the text feature map according to the horizontal angle of the text prediction box, and then a text recognition module is used to transcribe the text label. Finally, the text detection module is combined to complete the end-to-end recognition task of the network image.
[0007] The technical solution adopted by the present invention to overcome the technical problems is:
[0008] A network image text recognition method based on multi-scale feature fusion, comprising at least the following steps:
[0009] Through the U-net network structure of the residual network, the feature map of all text information of each original network image in the training data set is extracted;
[0010] Based on the feature map of all text information extracted by the residual network, the full convolutional network is used to output the text prediction box and the horizontal angle of the text prediction box;
[0011] The feature map of all text information extracted by the residual network is combined with the horizontal angle of the text prediction box, and the affine transformation is performed on the directed feature area of the feature map to obtain the axially aligned feature map;
[0012] The axially aligned feature maps are transcribed with text labels, and combined with the text detection results output by the fully convolutional network, the detection and recognition results of all texts in the original image are output.
[0013] Furthermore, through the U-net network structure of the residual network, the feature map of all text information of each original network image in the training data set is extracted, including:
[0014] The U-net network structure is used for the original network image, and the resolution features of 1 / 32 to 1 / 4 of the original image are fused in the bottom-up path. The fusion process is expressed by formulas (1) and (2):
[0015]
[0016]
[0017] The above equations (1) and (2) describe the upsampling process, where t i is the feature tensor output by each upsampling level, u i is the fused feature map, f i is the feature tensor of the normal convolution, [t i-1 ;f i ] indicates that the upsampling result of the previous level is connected and fused with the features of this level.
[0018] Furthermore, based on the feature map of all text information extracted by the residual network, the full convolutional network is used to output the text prediction box and the horizontal angle of the text prediction box, specifically including:
[0019] A fully convolutional network is used for text detection. The convolution output results include two parts:
[0020] In the first part, we complete the output of the classification task and generate a text score map. The output of the classification task has one channel, which is used to calculate the probability that each pixel belongs to a positive sample.
[0021] The second part is to output the text prediction box. The output of the text prediction box has 5 channels, 4 of which output the distance between the pixel and the top, bottom, left and right of the text prediction box, and the other channel outputs the horizontal angle of the text prediction box.
[0022] Each positive sample is processed by thresholding and non-maximum suppression algorithm to obtain the final text detection result. The thresholding and non-maximum suppression algorithm is expressed by formula (3):
[0023]
[0024] In the above formula, B is the text prediction box with the largest current score, N t is the suppression threshold, s i is a fraction, b i Predict other text boxes for the same text line, iou equals B and b i The sum of the areas of B and b i The ratio of the areas of the union of .
[0025] Furthermore, the convolution kernel size of the fully convolutional network is 1×1 and the output channel is 6.
[0026] Furthermore, the feature map of all text information extracted by the residual network is combined with the horizontal angle of the text prediction box, and the affine transformation rotation operation is performed on the directed feature area of the feature map to obtain the axially aligned feature map, which specifically includes:
[0027] Step 1: First, calculate the displacement deviation between the pixel point and the original pixel point after the feature map affine transformation, and use it as the translation transformation parameter of the affine transformation matrix to align the center points before and after the transformation. The calculation is expressed by formulas (4) and (5):
[0028] Δx=lcosθ-tsinθ-x (4)
[0029] Δy=tcosθ+lsinθ-y (5)
[0030] In the above formula, Δx and Δy are the displacement deviations on the x-axis and y-axis before and after the affine transformation of a point in the feature map respectively;
[0031] Then by fixing the height h t Calculate the scaling ratio s of the feature map, for width w t Scaling is performed at the same ratio, and the formulas are shown in (6) and (7):
[0032] s=h t / (t+b) (6)
[0033] w t =s×(l+r) (7)
[0034] Finally, the text prediction box is rotated according to the horizontal angle θ, and the final affine transformation matrix is shown in formula (8):
[0035]
[0036] In the above equations (4)-(8), M represents the final affine transformation matrix, h t and w t Represent the height and width of the feature map after affine transformation, (x, y) represents the coordinates of a point in the feature map, t, b, l, r represent the distance from the point (x, y) to the top, bottom, left and right of the text prediction box, and θ represents the horizontal angle of the prediction box;
[0037] Step 2: Using the above affine transformation parameters, the final axially aligned feature map is obtained, which is expressed by formula (8):
[0038]
[0039] In the above formula, is the coordinate of a point before affine transformation, Represents the coordinates of a point after affine transformation.
[0040] The present invention also discloses a network image text recognition system based on multi-scale feature fusion, which at least comprises:
[0041] The feature extraction module is configured to extract a feature map of all text information of each network original image in the training data set through the U-net network structure of the residual network;
[0042] A text detection module is configured to output a text prediction box and a horizontal angle of the text prediction box based on a feature map of all text information extracted by a residual network using a full convolutional network;
[0043] A rotation module is configured to combine the feature map of all text information extracted by the residual network with the horizontal angle of the text prediction box, perform an affine transformation rotation operation on the directional feature area of the feature map, and obtain an axially aligned feature map;
[0044] The text recognition module is configured to transcribe text labels for the axially aligned feature maps, and output the detection and recognition results of all texts in the original image in conjunction with the text detection results output by the fully convolutional network.
[0045] Furthermore, the original network image space size is 512×512.
[0046] Furthermore, the U-net network structure adopts SE_ResneXt50 as the basic network.
[0047] Furthermore, the text recognition module includes a sequential convolutional layer, a pooling layer, a bidirectional long short-term memory network, a fully connected layer and a CTC decoder.
[0048] Furthermore, there are 7 sequential convolutional layers, and the convolution kernel size of the sequential convolutional layer is 3×3; there are 2 pooling layers, and the convolution kernel size of the pooling layer is 2×1.
[0049] The beneficial effects of the present invention are:
[0050] 1. Compared with the prior art, the present invention uses the U-net feature extraction network to fuse feature maps of different scales during feature extraction, thereby realizing the recognition of texts of different scales in network images and improving the robustness of the system.
[0051] 2. Compared with the prior art, the present invention uses a rotation module based on affine transformation to obtain an axially aligned feature map, thereby achieving detection and recognition effects on texts in different directions and improving the accuracy of the system.
[0052] 3. Compared with the prior art, the present invention integrates the text detection module and the text recognition module into one system, realizes the complementarity of the detection module and the recognition module, and reduces the data required for training. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 It is an overall framework diagram of a network image text recognition method and system based on multi-scale feature fusion according to an embodiment of the present invention;
[0054] Figure 2 This is a network image text feature extraction diagram based on multi-scale feature fusion of residual network in an embodiment of the present invention. DETAILED DESCRIPTION
[0055] In order to facilitate those skilled in the art to better understand the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The following is only exemplary and does not limit the protection scope of the present invention.
[0056] The present invention discloses a network image text recognition system based on multi-scale feature fusion, which at least comprises:
[0057] The feature extraction module is configured to extract a feature map of all text information of each network original image in the training data set through the U-net network structure of the residual network;
[0058] A text detection module is configured to output a text prediction box and a horizontal angle of the text prediction box based on a feature map of all text information extracted by a residual network using a full convolutional network;
[0059] A rotation module is configured to combine the feature map of all text information extracted by the residual network with the horizontal angle of the text prediction box, perform an affine transformation rotation operation on the directional feature area of the feature map, and obtain an axially aligned feature map;
[0060] The text recognition module is configured to transcribe text labels for the axially aligned feature maps, and output the detection and recognition results of all texts in the original image in conjunction with the text detection results output by the fully convolutional network.
[0061] As preferred in this embodiment, the network original image space size is 512×512; the U-net network structure uses SE_ResneXt50 as the basic network.
[0062] The method of using network image text recognition system based on multi-scale feature fusion, such as Figure 1 As shown, at least the steps include:
[0063] Step S1: extract the feature map of all text information of each original network image in the training data set through the U-net network structure of the residual network (i.e., through the feature extraction module).
[0064] by Figure 2 The end-to-end network image text feature extraction diagram based on residual network is shown as an example. The feature map resolution is increased from bottom to top. The bottom feature map is first upsampled to double the original resolution. Then it is connected with the features of the previous layer. A 1×1 convolution layer merges multiple channels into one channel. Then a 3×3 convolution layer is used to fuse high semantics and high resolution information. Finally, the fused feature tensor is passed to the next layer until the top layer is finally merged through a 3×3 convolution layer to output the feature map of the network image text. The fusion process is expressed by formulas (1) and (2):
[0065]
[0066]
[0067] The above equations (1) and (2) describe the upsampling process, where t i is the feature tensor output by each upsampling level, u i is the fused feature map, f i is the feature tensor of the normal convolution, [t i-1 ;f i ] indicates that the upsampling result of the previous level is connected and fused with the features of this level.
[0068] Step S2: Based on the feature map of the text information extracted by the residual network in step S1, a fully convolutional network is used to output the text prediction box and the horizontal angle of the text prediction box.
[0069] Specifically, the text detection module uses a text detector of a full convolutional network to obtain a text prediction box and a horizontal angle of the text prediction box through multiple channels. In this embodiment, the convolution kernel size of the full convolutional network is preferably 1×1 and the output channel is 6.
[0070] A fully convolutional network is used for text detection. The convolution output results include two parts:
[0071] In the first part, we complete the output of the classification task and generate a text score map. The output of the classification task has one channel, which is used to calculate the probability that each pixel belongs to a positive sample.
[0072] The second part is to output the text prediction box. The output of the text prediction box has 5 channels, 4 of which output the distance between the pixel and the top, bottom, left and right of the text prediction box respectively, and the other channel outputs the horizontal angle of the text prediction box.
[0073] Finally, each positive sample is processed by thresholding and non-maximum suppression algorithm to obtain the final text detection result. The thresholding and non-maximum suppression algorithm is expressed by formula (3):
[0074]
[0075] In the above formula, B is the text prediction box with the largest current score, N t is the suppression threshold, s i is a fraction, b i Predict other text boxes for the same text line, iou equals B and b i The sum of the areas of B and b i The ratio of the areas of the union of .
[0076] Step S3: The feature map of the text information extracted by the feature extraction module and the horizontal angle of the text prediction box output by the text detection module are used as inputs of the rotation module, and an affine transformation rotation operation is performed on the directed feature area of the feature map to obtain an axially aligned feature map. The operation of the rotation module specifically includes two steps:
[0077] Step 1: First, calculate the displacement deviation between the pixel point and the original pixel point after the feature map affine transformation, and use it as the translation transformation parameter of the affine transformation matrix to align the center points before and after the transformation. The calculation is expressed by formulas (4) and (5):
[0078] Δx=lcosθ-tsinθ-x (4)
[0079] Δy=tcosθ+lsinθ-y (5)
[0080] Then by fixing the height h t Calculate the scaling ratio s of the feature map, for width w t Scaling is performed at the same ratio, and the formulas are shown in (6) and (7):
[0081] s=h t / (t+b) (6)
[0082] w t =s×(l+r) (7)
[0083] Finally, the text prediction box is rotated according to the horizontal angle θ, and the final affine transformation matrix is shown in formula (8):
[0084]
[0085] In the above equations (4)-(8), M represents the final affine transformation matrix, h t and w t Represent the height and width of the feature map after affine transformation, (x, y) represents the coordinates of a point in the feature map, t, b, l, r represent the distance from the point (x, y) to the top, bottom, left and right of the text prediction box, θ represents the horizontal angle of the prediction box, Δx and Δy represent the displacement deviation of a point in the feature map on the x-axis and y-axis before and after affine transformation;
[0086] Step 2: Using the above affine transformation parameters, the final horizontal feature map is obtained, which is expressed by formula (8):
[0087]
[0088] In the above formula, is the coordinate of a point before affine transformation, Represents the coordinates of a point after affine transformation.
[0089] Step S4: transcribe the text labels for the axially aligned feature maps, and output the detection and recognition results of all the texts in the original image in conjunction with the text detection results output by the full convolutional network.
[0090] As a preferred embodiment of the present invention, the text recognition module includes a sequential convolution layer, a pooling layer, a bidirectional long short-term memory network, a fully connected layer and a CTC decoder. The number of the sequential convolution layers is 7, and the convolution kernel size of the sequential convolution layer is 3×3. The number of the pooling layers is 2, and the convolution kernel size of the pooling layer is 2×1. As shown in Table 1. In Table 1, conv_bn_relu means batch normalization and ReLU activation function processing after convolution. The purpose of height-max-pool is to reduce the feature dimension along the height axis. bi-directional_lstm means bidirectional LSTM, and fully-connect means a fully connected layer. First, the spatial features are input into multiple convolution layers and pooling layers along the height axis for dimensionality reduction to extract higher-level features; then the extracted high-level feature map L∈R c×H×W Converted to timestamp as feature sequence l1,l2,…,l W ∈R c×H , and input into a recurrent neural network (RNN) for encoding. A bidirectional LSTM with 256 (D) output channels in each direction is used in the RNN to capture the range dependency of the input sequence features. Then the hidden states h1,h2,…,h are calculated at each time step in both directions. W ∈R D The sum is input into the fully connected layer, and the fully connected layer outputs each state x W ∈R |S| The distribution on the character category set S, and finally the CTC decoder transcribes the frame-by-frame classification scores into a text label sequence. The character category set S includes 3000 commonly used Chinese characters, 52 uppercase and lowercase English characters, 10 numbers, and special symbols "*!~@#$%^&(){}[]".
[0091] Table 1 Text recognition module structure
[0092]
[0093]
[0094] As shown in Table 2, the method of the present invention is compared with the existing FOTS method, NTAI method, and CCFlab method on the network image text dataset MTWI, using the accuracy of network image text recognition and the amount of additional data required by the method as evaluation indicators, respectively. The higher the accuracy, the less additional data is used, and the better the performance of the method. It can be seen that the method of the present invention achieves the best results in both indicators on the network image text dataset.
[0095] Table 2 Performance comparison between the method of the present invention and the existing end-to-end network image text recognition method (%)
[0096] method Accuracy Additional data volume FOTS 66.1 8034 NTAI 68.8 1.1 million CCFlab 56.3 30,000 Method of the present invention 68.9 8034
[0097] The above only describes the basic principles and preferred embodiments of the present invention. Those skilled in the art may make many changes and improvements based on the above description, and these changes and improvements should fall within the protection scope of the present invention.
Claims
1. A network image text recognition method based on multi-scale feature fusion, characterized in that: At least the steps include: Through the U-net network structure of the residual network, the feature map of all text information of each original network image in the training data set is extracted, including: The U-net network structure is used for the original network image, and the resolution features of 1 / 32 to 1 / 4 of the original image are fused in the bottom-up path. The fusion process is expressed by formulas (1) and (2): The above equations (1) and (2) describe the upsampling process, where t i is the feature tensor output by each upsampling level, u i is the fused feature map, f i is the feature tensor of the normal convolution, [t i-1 ;f i ] indicates that the upsampling result of the previous level is connected and fused with the features of this level; Based on the feature map of all text information extracted by the residual network, the full convolutional network is used to output the text prediction box and the horizontal angle of the text prediction box, including: A fully convolutional network is used for text detection. The convolution output results include two parts: In the first part, we complete the output of the classification task and generate a text score map. The output of the classification task has one channel, which is used to calculate the probability that each pixel belongs to a positive sample. The second part is to output the text prediction box. The output of the text prediction box has 5 channels, 4 of which output the distance between the pixel and the top, bottom, left and right of the text prediction box, and the other channel outputs the horizontal angle of the text prediction box. Each positive sample is processed by thresholding and non-maximum suppression algorithm to obtain the final text detection result. The thresholding and non-maximum suppression algorithm is expressed by formula (3): In the above formula, B is the text prediction box with the largest current score, N t is the suppression threshold, s i is a fraction, b i Predict other text boxes for the same text line, iou equals B and b i The sum of the areas of B and b i The ratio of the areas of the union of ; The feature map of all text information extracted by the residual network is combined with the horizontal angle of the text prediction box, and the affine transformation rotation operation is performed on the directed feature area of the feature map to obtain the axially aligned feature map, which includes: Step 1: First, calculate the displacement deviation between the pixel point and the original pixel point after the feature map affine transformation, and use it as the translation transformation parameter of the affine transformation matrix to align the center points before and after the transformation. The calculation is expressed by formulas (4) and (5): Δx=lcosθ-tsinθ-x (4) Δy=tcosθ+lsinθ-y (5) In the above formula, Δx and Δy are the displacement deviations on the x-axis and y-axis before and after the affine transformation of a point in the feature map respectively; Then by fixing the height h t Calculate the scaling ratio s of the feature map, for width w t Scaling is performed at the same ratio, and the formulas are shown in (6) and (7): s=h t / (t+b) (6) w t =s×(l+r) (7) Finally, the text prediction box is rotated according to the horizontal angle θ, and the final affine transformation matrix is shown in formula (8): In the above equations (4)-(8), M represents the final affine transformation matrix, h t and w t Represent the height and width of the feature map after affine transformation, (x, y) represents the coordinates of a point in the feature map, t, b, l, r represent the distance from the point (x, y) to the top, bottom, left and right of the text prediction box, and θ represents the horizontal angle of the prediction box; Step 2: Using the above affine transformation parameters, the final axially aligned feature map is obtained, which is expressed by formula (8): In the above formula, is the coordinate of a point before affine transformation, Represents the coordinates of a point after affine transformation; The axially aligned feature maps are transcribed with text labels, and combined with the text detection results output by the fully convolutional network, the detection and recognition results of all texts in the original image are output.
2. The network image text recognition method based on multi-scale feature fusion according to claim 1 is characterized in that: The convolution kernel size of the fully convolutional network is 1×1 and the output channel is 6.
3. A network image text recognition system based on multi-scale feature fusion, characterized in that: At least: The feature extraction module is configured to extract the feature map of all text information of each network original image in the training data set through the U-net network structure of the residual network, including: The U-net network structure is used for the original network image, and the resolution features of 1 / 32 to 1 / 4 of the original image are fused in the bottom-up path. The fusion process is expressed by formulas (1) and (2): The above equations (1) and (2) describe the upsampling process, where t i is the feature tensor output by each upsampling level, u i is the fused feature map, f i is the feature tensor of the normal convolution, [t i-1 ;f i ] indicates that the upsampling result of the previous level is connected and fused with the features of this level; The text detection module is configured to be used for the feature map of all text information extracted based on the residual network, and outputs the text prediction box and the horizontal angle of the text prediction box using the full convolutional network, specifically including: A fully convolutional network is used for text detection. The convolution output results include two parts: In the first part, we complete the output of the classification task and generate a text score map. The output of the classification task has one channel, which is used to calculate the probability that each pixel belongs to a positive sample. The second part is to output the text prediction box. The output of the text prediction box has 5 channels, 4 of which output the distance between the pixel and the top, bottom, left and right of the text prediction box, and the other channel outputs the horizontal angle of the text prediction box. Each positive sample is processed by thresholding and non-maximum suppression algorithm to obtain the final text detection result. The thresholding and non-maximum suppression algorithm is expressed by formula (3): In the above formula, B is the text prediction box with the largest current score, N t is the suppression threshold, s i is a fraction, b i Predict other text boxes for the same text line, iou equals B and b i The sum of the areas of B and b i The ratio of the areas of the union of ; The rotation module is configured to combine the feature map of all text information extracted by the residual network with the horizontal angle of the text prediction box, perform an affine transformation rotation operation on the directional feature area of the feature map, and obtain an axially aligned feature map, specifically including: Step 1: First, calculate the displacement deviation between the pixel point and the original pixel point after the feature map affine transformation, and use it as the translation transformation parameter of the affine transformation matrix to align the center points before and after the transformation. The calculation is expressed by formulas (4) and (5): Δx=lcosθ-tsinθ-x (4) Δy=tcosθ+lsinθ-y (5) In the above formula, Δx and Δy are the displacement deviations on the x-axis and y-axis before and after the affine transformation of a point in the feature map respectively; Then by fixing the height h t Calculate the scaling ratio s of the feature map, for width w t Scaling is performed at the same ratio, and the formulas are shown in (6) and (7): s=h t / (t+b) (6) w t =s×(l+r) (7) Finally, the text prediction box is rotated according to the horizontal angle θ, and the final affine transformation matrix is shown in formula (8): In the above equations (4)-(8), M represents the final affine transformation matrix, h t and w t Represent the height and width of the feature map after affine transformation, (x, y) represents the coordinates of a point in the feature map, t, b, l, r represent the distance from the point (x, y) to the top, bottom, left and right of the text prediction box, and θ represents the horizontal angle of the prediction box; Step 2: Using the above affine transformation parameters, the final axially aligned feature map is obtained, which is expressed by formula (8): In the above formula, is the coordinate of a point before affine transformation, Represents the coordinates of a point after affine transformation; The text recognition module is configured to transcribe text labels for the axially aligned feature maps, and output the detection and recognition results of all texts in the original image in conjunction with the text detection results output by the fully convolutional network.
4. The network image text recognition system based on multi-scale feature fusion according to claim 3 is characterized in that: The original network image space size is 512×512.
5. The network image text recognition system based on multi-scale feature fusion according to claim 3 is characterized in that: The U-net network structure uses SE_ResneXt50 as the basic network.
6. The network image text recognition system based on multi-scale feature fusion according to claim 3 is characterized in that: The text recognition module includes sequential convolutional layers, pooling layers, a bidirectional long short-term memory network, a fully connected layer and a CTC decoder.
7. The network image text recognition system based on multi-scale feature fusion according to claim 6 is characterized in that: There are 7 sequential convolutional layers, and the convolution kernel size of the sequential convolutional layers is 3×3; there are 2 pooling layers, and the convolution kernel size of the pooling layers is 2×1.
Citation Information
Patent Citations
A Text Line Recognition Method in Chinese Scenes Based on Residual Convolutional and Recurrent Neural Networks
AU2020101229A4
A method for detecting and recognizing sensitive characters in natural scene images
CN109447078A
