A global interactive image semantic parsing method based on attention mechanism
By introducing attention mechanism and bidirectional recurrent neural network into the image semantic understanding model, the problems of poor semantic analytical logic and slow convergence speed in traditional methods are solved, and higher attention to detail and recognition accuracy are achieved.
Patent Information
- Application Number
- CN202011251599.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-11-11
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2040-11-11
AI Technical Summary
Traditional image semantics understanding methods rely on syntax templates and reference image text databases, resulting in poor logic of generated semantic analysis, slow convergence speed, and difficulty in accurately describing image detail properties.
Adopt global interactive image semantic understanding model based on attention mechanism, semantic analysis of image information through bidirectional recurrent neural network, and introduce attention mechanism in the generation process to pay attention to the global information and different regional information of the image in real time to ensure semantic coherence and logic.
It improves the logic and attention to detail of image semantics, improves the accuracy of identifying content and the accuracy of generating text, reduces the risk of overfitting the model, and achieves a faster convergence speed.
Smart Images

Figure CN114549887B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a semantic parsing method of a global interactive image semantic understanding model based on an attention mechanism, which belongs to the technical fields of target detection and semantic analysis algorithms, and is applied to the fields of content-based image retrieval, medical image analysis, blind guide assistance, early childhood education, etc. That is, given a picture, a text description of the picture is generated. Background Art
[0002] Image semantic understanding is based on image recognition and integrates interdisciplinary research in computer science, psychology, linguistics and other disciplines. It has also made important contributions to the study of cross-modal interaction between images and texts. Image semantic understanding technology wants to rationally or emotionally understand the target image as a whole and generate a natural language description that conforms to human habits. It not only needs to extract and identify the scenes, objects and attributes contained in the target image, but also analyze the relationship between each object and attribute, including the action, shape, psychology and emotion of each object, and generate a text description of the image based on this information. Therefore, this is a very complex and challenging task.
[0003] At present, image semantic understanding technology has been widely used in many fields such as image retrieval, medical image analysis, assisted guide, early childhood education, news automation, space military security, etc., and has played an increasingly important role in various fields.
[0004] Traditional methods for image semantic understanding are mainly based on templates and transfer generation methods. The limitation of these methods is that the overall model is too dependent on a certain grammatical template or reference image text database, ignoring the process of the language model to flexibly parse the image and generate new text. Therefore, the output results of the model are not satisfactory. In recent years, with the application of encoder-decoder based neural network models in the field of image semantic understanding, this task has made rapid progress and achievements.
[0005] The invention mainly innovates on how to effectively improve and enhance the image semantic understanding effect of the neural network model based on the encoder-decoder, focusing on solving the problem of using the bidirectional recurrent neural network model for image semantic parsing; regularizing the image data and text data on the basis of the bidirectional recurrent neural network model, and introducing the image information into the bidirectional recurrent neural network in a global interactive manner; introducing the attention mechanism on the basis of global interaction. This model has the characteristics of strong logic of the image semantic understanding results, high accuracy in identifying image content, and accurate description of image detail attributes. Summary of the invention
[0006] Aiming at the problems of slow model convergence speed and poor logicality of generated semantic analysis in the baseline model (Neural Image Caption Generator, NIC), a representative model of the encoding and decoding structure, the present invention proposes to feed global image information into the semantic analysis model in real time during the semantic analysis stage of the image semantic task to guide semantic generation, and adopt a bidirectional recurrent neural network model to perform semantic analysis on the image, and introduce an attention mechanism in this process, so as to obtain a more logical and detailed image semantic description. The overall structure of the global interactive image semantic understanding model based on the attention mechanism is as follows: Figure 1 shown.
[0007] The technical solution adopted by the present invention is as follows: a global interactive image semantic understanding model based on an attention mechanism is used for image semantic generation, that is, a bidirectional recurrent neural network model is used to perform semantic analysis on image information in the process of generating semantic text corresponding to the image, so that the model can pay attention to the previous and next context information in real time during the semantic analysis process to ensure semantic coherence and logic; in the semantic analysis process, real-time attention is paid to the global information of the image to guide semantic generation; and an attention mechanism is introduced in the semantic analysis process, so that different areas of the image information are paid attention to at different times in the semantic analysis process, so that the image semantic understanding model is more suitable for processing image detail information.
[0008] A global interactive image semantic parsing method based on an attention mechanism includes the following steps:
[0009] 1) Image feature information extraction: The image information is extracted from the input image through a convolutional neural network to generate a high-dimensional feature vector of the image information, and this feature vector is sent to the semantic parsing model as the global information of the image for decoding;
[0010] 2) Decode image feature information through semantic parsing model:
[0011] The global information of the image is fed into the forward GRU network and the backward GRU network. The forward GRU network and the backward GRU network each independently generate an update state at time t, which is used to input into the next GRU unit in each network. The outputs of the forward GRU network and the backward GRU network are directly linearly superimposed.
[0012] The global information of the image is processed by the attention model to obtain the image area c that the semantic parsing model should focus on when predicting the next text information at the current moment t , respectively sent to the forward GRU network and the backward GRU network, and superimposed with the outputs of the GRU units in the forward GRU network and the backward GRU network, and finally the image text corresponding to the input image at the current moment is obtained.
[0013] When the CNN model and the semantic parsing model are pre-trained through steps 1) to 2), in the step of decoding the image feature information through the semantic parsing model, the global information of the image is processed by the attention model to obtain image data c for highlighting the image area of interest t , c t The text information of the input image is superimposed to obtain image data containing the image text information, which is input into the forward GRU and backward GRU respectively.
[0014] When training the semantic parsing model, the input information of the model is the mean of the information of different regions of the input image. And the current word vector S t Cascaded:
[0015]
[0016] In the formula, S t The word vector representing the text information corresponding to the input image at time t;
[0017]
[0018] L represents the number of features in different regions of the image, v i Represents the information of the i-th image region.
[0019] The attention model calculates similarity based on the global information of the image and the update state at the previous moment; the update state represents the update state of the GRU unit, and the global information of the image represents the regional image features in the image; the greater the cosine similarity between the two, the more the image information of this region is currently being parsed, and the attention mechanism will make the semantic parsing model pay more attention to the image information of this region.
[0020] The weight e of the attention model ti Obtained by the following formula:
[0021]
[0022]
[0023] Nonlinear function f att Represents the cosine distance metric function between vectors, h t represents the hidden state of the GRU unit at time t, a ti Indicates the image region v assigned to the i-th image region at time t i The attention weight is used to indicate the importance of different image regions for the current semantic parsing model to predict the next text information;
[0024] The information at different positions of the image is paid attention to differently, and the following formula is used to obtain:
[0025]
[0026] Among them, c t After being processed by the attention model, the image information of different regions is accumulated according to the weights to represent the image information that needs to be paid attention to at the current time t. L represents the number of features in different regions of the image, and v i Represents the information of the i-th image region.
[0027] The present invention has the following advantages:
[0028] The generated target image has a stronger semantic description logic, a higher degree of attention to details and is more accurate; the image content is recognized with higher accuracy; the algorithm model converges quickly and is not prone to overfitting, etc. It has a very obvious effect on improving the accuracy of image semantic analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 This is a schematic diagram of the overall structure of the global interactive image semantic understanding model based on the attention mechanism.
[0030] Figure 2 It is a schematic diagram of the GRU unit structure.
[0031] Figure 3 It is a schematic diagram of the GRU unit structure of global interaction.
[0032] Figure 4 It is a schematic diagram of the bidirectional GRU recurrent neural network unfolding in time sequence.
[0033] Figure 5 This is a schematic diagram of the overall structure of the image semantic understanding model based on global interaction.
[0034] Figure 6a This is a comparison of the semantic parsing model structure before and after adding the attention mechanism. Figure 1 .
[0035] Figure 6b This is a comparison of the semantic parsing model structure before and after adding the attention mechanism. Figure 2 .
[0036] Figure 7 This is a schematic diagram of the model of the semantic parsing stage after the attention mechanism is added.
[0037] Figure 8 This is an example diagram of the image semantic understanding model test based on the global interaction of the attention mechanism. Specific implementation plan
[0038] The present invention is further described in detail below.
[0039] The present invention constructs an attention mechanism model and method in the process of semantic parsing of machine vision data, and applies it to the generation of image semantic titles. The specific steps are: 1) Select a target image feature extraction model to extract and encode features of image data; 2) Build a globally interactive bidirectional recurrent neural network to parse image features; 3) Perform standard regularization on the extracted image feature information, and send it to the semantic parsing model in real time in the form of global information for model training; 4) Add an attention mechanism to the semantic parsing process so that the model focuses on the information of different image regions during the semantic parsing process; 5) Perform semantic parsing on new target images through the trained model. The image semantic understanding model established by the present invention and the image semantics generated by the algorithm have the characteristics of strong logic and rich semantics, fast model convergence speed, high semantic parsing accuracy, high semantic parsing accuracy for complex images or images that focus on details, etc., and have been successfully applied in the fields of content-based image retrieval, medical image analysis, auxiliary guide, early childhood education, etc.
[0040] Step 1: Image feature information extraction and encoding
[0041] 1.1) Image feature extraction and encoding
[0042] The image feature extraction encoder in the model of the present invention adopts the convolutional neural network VGG-16 model to extract features from the input image, and the data of the last convolution layer of the model is processed by adaptive average pooling, the size of the image feature vector is uniformly set to 14×14×2048, and this feature vector is sent to the decoding end as the global information of the image for cross-modal interaction.
[0043] Step 2: Image feature information decoding
[0044] 2.1) Gated Recurrent Unit
[0045] In order to improve the accuracy and richness of language description, a bidirectional recurrent neural network model is adopted in the global interaction model of the present invention. In addition, in order to avoid the problem of a sharp increase in parameter scale caused by directly adopting a long short-term memory unit (LSTM), the model adopts a gated recurrent unit (GRU) as the logical unit of the recurrent neural network, thereby achieving the purposes of reducing the model parameter scale, improving the model convergence speed and reducing the degree of model overfitting.
[0046] The GRU unit merges the input gate and forget gate of the LSTM unit into an update gate. Instead of building linear self-updates on additional memory units, it directly builds on hidden states after linear accumulation and relies on gates to regulate. After simplifying the gating logic components inside the LSTM unit, the GRU unit reduces the overall parameter size by about 30%, accelerates the convergence speed of the model training process, and reduces the possibility of model overfitting. The working principle and update formula of the GRU unit gating logic component are as follows:
[0047] r t =δ(W r x t +U r h t-1 )
[0048] Reset Gate: r t Responsible for determining the hidden unit h at the previous moment t-1 New Memory The importance of r t Approximately equal to 0, h t-1 It will not be passed to New Memory Among them, W r and U r It is the Reset Gate update parameter.
[0049] z t =δ(W z x t +U z h t-1 )
[0050] Update Gate: z t Responsible for deciding how much of the previous hidden state h to pass t-1 Give the hidden state h at the current moment t .W z and U z Is the Update Gate update parameter. If z t If it is approximately equal to 1, h t-1 Almost directly copied to h t , on the contrary, if z t Approximately equal to 0, New Memory Passed directly to h t , as shown in formula (3.5):
[0051]
[0052] New Memory: New Memory is the current input x tand the Hidden Stateh of the previous moment t-1 W and U are the New Memory update parameters. The new vector is calculated and summarized Contains the above information and the new input x t .
[0053]
[0054] Hidden State:h t By h t-1 and Add them together, and the weight of the two is controlled by Update Gate z(t).
[0055] Where δ represents the sigmoid function and ⊙ represents the bitwise AND operation. Figure 2 shown.
[0056] 2.2) Gated Recurrent Units with Global Interaction
[0057] In the recurrent neural network at the decoding end of the NIC baseline model, the generation of image semantic information mainly depends on the input information at the current time point and the hidden state at the previous moment (which implies the image information input at the beginning). This process proceeds step by step until the end marker of the sentence is encountered. However, as this process continues, the role of the image information sent to the language model at the beginning becomes weaker and weaker, and in the entire semantic generation process, there will be blurring or loss of some image information, resulting in the semantic description being unable to express the image content richly and comprehensively. Therefore, when a longer sentence is needed to describe the image, at the end of the semantic generation stage, the model almost "blindly" executes to the end of the sentence. In order to solve this problem, a global interaction mechanism of image information is introduced in this model, that is, global image information will be introduced in each time series in the semantic generation process of the recurrent neural network. The schematic diagram of the global interactive GRU unit structure is shown in the figure. Figure 3 shown.
[0058] Figure 3 The V in the green dashed box represents the feature vector sent to the language model to represent the global information of the image. The update formula of the global image information interaction unit is as follows:
[0059] r t =δ(W r x t +U r h t-1 +G r V t-1 )
[0060] z t =δ(W z xt +U z h t-1 +G z V t-1 )
[0061]
[0062]
[0063] The structure and working principle of the gate logic component of the GRU unit after adding the global image information are the same as those of the GRU unit before adding the global image information, except that the input of the global image information is added to the update formula. t Update the formula for ResetGate, W r , U r , G r is the corresponding update parameter of the formula; z t Update Gate update formula, W z , U z , G z is the corresponding update parameter of the formula; is the New Memory update formula, W and U are the corresponding update parameters of the formula; h t Update the formula for Hidden State; where x in the formula t 、h t-1 、V t-1 They represent the current external input, the hidden state at the previous moment, and the global image information at the previous moment respectively. δ represents the sigmoid function, and ⊙ represents the bitwise AND operation.
[0064] 2.3) Bidirectional Recurrent Neural Networks with Global Interaction
[0065] In the generation and training process of the semantic description of the target image, if a unidirectional recurrent neural network is used, each word will only pay attention to the text information in front of it in the word order during the generation process, but not the text information behind it. However, when decoding a word in the entire parsing process, it is usually necessary to know the information around the word, that is, the previous and next context information. The introduction of the bidirectional recurrent neural network model has largely solved this problem. The unidirectional recurrent neural network only outputs information in one direction at the same time t, while the main structure of the bidirectional recurrent neural network is composed of two unidirectional recurrent neural networks in opposite directions superimposed. At each time series t, the two recurrent neural networks in opposite directions will receive external input at the same time. Each layer of the network will update parameters independently, without affecting each other, and independently generate their own update states and output results at time t, and output information in two directions. Finally, the output results of the two unidirectional recurrent neural networks in opposite directions are directly linearly superimposed as the final output of the bidirectional recurrent network. Except for the different directions, the basic structures of the two recurrent neural networks are completely symmetrical, and they fully follow the update rules in 2.1 and 2.2. Take the New Memory update formula as an example:
[0066]
[0067]
[0068] in, is the New Memory calculation result of the gated recurrent unit at time t of the forward propagation, is the New Memory calculation result of the gated recurrent unit at time t of the backward propagation. The parameters W and U are the corresponding update parameters of the formula. ⊙ still represents the bitwise AND operation. In order to distinguish the forward and backward update formulas, only the superscripts F and B are used to represent the forward and backward update formulas respectively. The schematic diagram of the bidirectional GRU recurrent neural network in time sequence is as follows Figure 4 shown.
[0069] 2.4) Overall framework of image semantic understanding model based on global interaction
[0070] The overall architecture of this model follows the basic structure of encoding and decoding from input image to output text. Global interaction is mainly reflected in two aspects: first, the introduction of a bidirectional GRU unit network for image semantic generation, that is, in the process of generating semantic information, real-time attention is paid to the previous and next contextual information and connections, and no longer only to one-way semantic information; second, on the basis of the bidirectional GRU unit, the global information of the image is introduced into the GRU unit, and in the process of generating text, real-time attention is paid to the global information of the image to guide semantic generation, avoiding the problem of image information ambiguity at the end of semantic parsing caused by the image feature information being only sent to the recurrent neural network at time t0 when the recurrent neural network begins to parse the image information. The overall structure diagram of the image semantic understanding model based on global interaction is shown in the figure. Figure 5 shown.
[0071] 2.5) Attention Mechanism in Global Interaction Model
[0072] This paper adopts the soft attention mechanism mode, which strengthens the image information of the area that needs to be paid attention to when generating the next word of the target image, instead of randomly filtering out the image information that does not meet the conditions. Figure 6a , Figure 6b shown.
[0073] Figure 6a is a GRU unit without attention mechanism. Figure 6b Schematic diagram of GRU unit with attention mechanism. In the model, V = {v1,v2,...,v L} means that after feature extraction by convolutional neural network, the features of L different regions of a given image are output, h = {h1,h2,...h t ...,h n-1} represents the hidden layer state of the GRU unit of the recurrent neural network. The weight calculation of the attention mechanism can be done in many ways. Here we only use cosine similarity to introduce the weight calculation method of the attention mechanism. At this time, the attention mechanism calculates the cosine similarity of the semantic vector of each time series of the recurrent neural network and the image features of L different regions of the image. When the cosine similarity of the two is large, it means that the language parsing model is parsing this part of the image content at this moment. At this time, a higher weight is assigned to this part of the image area, so that the semantic parsing model pays more attention to this part of the image content. The attention weight calculation method is as follows:
[0074] e ti =f att (v i ,h t )
[0075] Here the nonlinear function f attis the cosine distance metric function between vectors. The above formula can be specifically expressed as:
[0076]
[0077]
[0078] a ti Indicates the image region v assigned to the i-th image region at time t i The attention weight is a value in the interval [0,1], which is used to indicate the importance of different image regions for the current semantic parsing model to predict the next word. Therefore, a visual context vector can be calculated by weighted sum to indicate how the semantic parsing model pays attention to information at different positions in the image at time t in the time series:
[0079]
[0080] At this point, the semantic parsing model in the decoding stage can calculate the probability distribution of the next word predicted at the current moment through the hidden layer state at the current moment and the information of the image area that needs to be focused on at the current moment.
[0081] 2.6) Bidirectional recurrent neural network model with global interaction based on attention mechanism
[0082] This section adds the attention mechanism mentioned in Section 2.5) on the basis of Sections 2.3) and 2.4). In the semantic parsing process, a bidirectional GRU unit recurrent neural network model is used, and the global image information is still fed into the GRU unit in each time series. The model diagram of the semantic parsing stage after adding the attention mechanism is shown in the figure below: Figure 7 shown.
[0083] Figure 7 The Attention layer in is the attention mechanism layer, which is used to decide which part of the image information to focus on when decoding the next word. Both the forward and reverse decoders use standard GRU units. The update method of the GRU unit completely follows the rules in 2.2, except that the model in this chapter also adds the image area data that is focused on. The update formula is briefly expressed as follows:
[0084] h t =GRU(x t ,c t ,h t-1 )
[0085] Among them, x t represents the external input of the GRU unit at the current time t, h t-1 represents the hidden state of the GRU unit at the previous moment, c tRepresents the image data after being processed by the attention mechanism, that is, the information intensity of the image data that is focused on when predicting the next word is higher than the information intensity of the image area that is not focused on. The external input at each moment is the mean of the global feature vector of the image And the word vector S at the current moment t Cascaded, external input x t The formula is as follows:
[0086]
[0087] In the formula, S t Represents the word vector of the input word at time t. Among them, the global feature vector of the image is the average value of the feature vectors of each region of the image:
[0088]
[0089] After adding the attention mechanism, when the model predicts the next word, it will use the attention mechanism to strengthen the image information that needs to be focused on based on the original global image information, and send it to the semantic parsing model, which plays the role of strengthening the information of the image area that needs to be focused on and weakening the information of the image area that is not focused on.
[0090] 2.7) Training the model
[0091] In the encoding stage, i.e., the image feature extraction stage, the weights of the VGG-16 model pre-trained on ImageNet are directly used as the initial values to import the model in this paper to accelerate the training process. The main parameters of the encoder model are not updated, and only the output layer parameters of the model are updated.
[0092] The decoding stage adopts a global interaction mechanism. The special feature of the global interaction mechanism is that the image information is no longer just before the semantic parsing begins, that is, t -1 The language model is not input at the time (t = 0 is the first time series of the parsing model), but the image information is sent to the GRU unit in each time series, and is not affected by the time series. It interacts with the semantic information in real time during the entire process of image semantic generation, thereby guiding the semantic generation. In the process of generating image semantics, it not only pays attention to the semantic information in a single direction, but also pays attention to the contextual information before and after the current word, so that the generated semantic description is more consistent with the image content described.
[0093] After adding the attention mechanism, in the decoding stage of the image semantic understanding model, when the recurrent neural network predicts the next word in different time series, it uses the attention weights calculated through training to learn to focus on different image feature information when generating semantic words, weaken non-focused image information, and no longer focus only on global image information. In this way, computers can think like humans to effectively recognize and understand images semantically.
[0094] Step 2: Model Testing
[0095] The experimental results of the baseline model and the global interactive image semantic understanding model of the present invention are compared in Table 1:
[0096] Table 1
[0097]
[0098] Among them, Baseline represents the NIC baseline model, MMD-1 represents the global interaction image semantic understanding model, and MMD-2 represents the global interaction image semantic understanding model based on the attention mechanism. From the comparison of the result data in the table, it can be seen that the effect of the global interaction image semantic understanding model based on the attention mechanism is better than that of the NIC baseline model and the image semantic understanding model based on the global interaction under the four BLEU indicators.
[0099] Under the assumption that all image data are independent and identically distributed, new third-party data is obtained to verify the model effect. Multiple rounds of model training are carried out during the experiment, and the model results of each round of training are saved. The semantic understanding example of the global interactive image semantic understanding model based on the attention mechanism under third-party image data is as follows, for example Figure 8 shown.
[0100] (1) a brown dog is chasing a white dog on the grass.
[0101] (2) a motorcyclist in black outerwear is riding a motorcycle on the road.
[0102] (3) a little girl in a white shirt is sitting on the grass holding a few yellow flowers.
[0103] (4)Two men in white shirt and shorts are playing soccer.
Claims
1. A global interactive image semantic parsing method based on attention mechanism, characterized in that: The following steps are involved: 1) Image feature information extraction: The image information is extracted from the input image through a convolutional neural network to generate a high-dimensional feature vector of the image information, and this feature vector is sent to the semantic parsing model as the global information of the image for decoding; 2) Decode image feature information through semantic parsing model: The global information of the image is fed into the forward GRU network and the backward GRU network. The forward GRU network and the backward GRU network each independently generate an update state at time t, which is used to input into the next GRU unit in each network. The outputs of the forward GRU network and the backward GRU network are directly linearly superimposed. The global information of the image is processed by the attention model to obtain the image area c that the semantic parsing model should focus on when predicting the next text information at the current moment t , respectively sent to the forward GRU network and the backward GRU network, and superimposed with the outputs of the GRU units in the forward GRU network and the backward GRU network, and finally the image text corresponding to the input image at the current moment is obtained.
2. According to claim 1, the image semantic parsing method based on global interaction of attention mechanism is characterized in that: The following steps are involved: When the CNN model and the semantic parsing model are pre-trained through steps 1) to 2), in the step of decoding the image feature information through the semantic parsing model, the global information of the image is processed by the attention model to obtain image data c for highlighting the image area of interest t , c t The text information of the input image is superimposed to obtain image data containing the image text information, which is input into the forward GRU and backward GRU respectively.
3. According to claim 2, the image semantic parsing method based on global interaction of attention mechanism is characterized in that: When training the semantic parsing model, the input information of the model is the mean of the information of different regions of the input image. And the current word vector S t Cascaded: In the formula, S t The word vector representing the text information corresponding to the input image at time t; L represents the number of features in different regions of the image, v i Represents the information of the i-th image region.
4. According to claim 1, the image semantic parsing method based on global interaction of attention mechanism is characterized in that: The attention model performs similarity calculation based on the global information of the image and the update state at the previous moment; the update state represents the update state of the GRU unit, and the global information of the image represents the regional image features in the image; The larger the cosine similarity between the two, the more the image information in this area is currently being parsed, and the attention mechanism will make the semantic parsing model pay more attention to the image information in this area.
5. According to claim 1, the image semantic parsing method based on global interaction of attention mechanism is characterized in that: The weight e of the attention model ti It is obtained by the following formula: Nonlinear function f att Represents the cosine distance metric function between vectors, h t represents the hidden state of the GRU unit at time t, a ti Indicates the image region v assigned to the i-th image region at time t i The attention weight is used to indicate the importance of different image regions for the current semantic parsing model to predict the next text information; The information at different positions of the image is paid attention to differently, and the following formula is used to obtain: Among them, c t After being processed by the attention model, the image information of different regions is accumulated according to the weights to represent the image information that needs to be paid attention to at the current time t. L represents the number of features in different regions of the image, and v i Represents the information of the i-th image region.
Citation Information
Patent Citations
Image semantic description generation method
CN110188779A
Machine reading understanding method for guiding attention based on knowledge
CN111241807A
Scene character recognition method and system based on semantic enhancement encoder decoder framework
CN111753827A
Method, apparatus and storage medium for recognizing character
US20200143191A1
