A ticket itinerary identification system and method using super-resolution technology
The ticket itinerary recognition system, which utilizes super-resolution technology, solves the problem of low recognition accuracy caused by image blur by employing term localization, super-resolution, and text recognition modules, thus achieving efficient text recognition.
Patent Information
- Application Number
- CN202311064354.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-23
- Publication Date
- 2026-02-03
- Estimated Expiration
- 2043-08-23
AI Technical Summary
In existing technologies, image blurring caused by camera shake and focus deviation affects the resolution and recognition accuracy of ticket itinerary images, making it difficult to solve effectively.
Super-resolution technology is employed, which detects text regions through the term localization module, uses the super-resolution module to sharpen the image, combines character feature extraction and generator modules to generate a sharpened version, and finally achieves accurate recognition through the text recognition module.
It improves the accuracy of ticket itinerary recognition, especially under low-quality image conditions, and significantly enhances the text recognition effect.
Smart Images
Figure CN117058703B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a ticket itinerary recognition system and method using super-resolution technology, belonging to the field of information technology, and particularly to the field of automatic ticket itinerary recognition technology. Background Technology
[0002] Passenger ticket itinerary recognition refers to the structured recognition of fields on flight itineraries, including name, departure station, destination station, flight number, date, time, fare, ID number, insurance premium, fuel surcharge, other taxes and fees, total amount, and e-ticket number. Automatically recognizing these fields on passenger ticket itineraries can be widely applied in internal expense reimbursement scenarios, significantly reducing manual data entry, lowering accounting costs, and automating financial and tax reimbursement processes.
[0003] In practical applications, issues such as camera shake and focus deviation leading to defocusing can produce low-quality images, resulting in low resolution and blurry images on passenger tickets and itineraries. This severely impacts text recognition and fails to guarantee a high recognition accuracy. Effectively addressing the recognition accuracy problem caused by image blur has become a pressing technical challenge in the field of automatic passenger ticket and itinerary recognition. Summary of the Invention
[0004] In view of this, the purpose of this invention is to propose a ticket itinerary recognition system and method using super-resolution technology. By using super-resolution technology, the text area of the ticket itinerary is clarified, thereby improving the recognition accuracy.
[0005] To achieve the above objectives, this invention proposes a ticket itinerary recognition system using super-resolution technology, the system comprising the following modules:
[0006] The term location module detects the coordinates of the detection box containing terms in the ticket itinerary image, crops the text region image block containing the term based on the detection box coordinates, and assigns the category of the text region image block. The category of the text region image block refers to the category of the term contained in the text region image block. The term refers to names, flight numbers, fares, totals, origin, destination, ID card numbers, e-ticket numbers, issuance dates, fuel surcharges, insurance premiums, civil aviation development funds, verification codes, printing serial numbers, etc.
[0007] Super-resolution module: This module receives the output information from the term location module and performs super-resolution processing on the text region image blocks output by the term location module to obtain a clearer version of the text region image blocks;
[0008] Text recognition module: The function of this module is to receive the sharpened version of the text region image block output by the super-resolution module, perform text recognition, and obtain the recognition result.
[0009] The term localization module consists of a backbone network CSPDarknet submodule, an enhanced feature extraction network FPN submodule, and a YOLO Head classifier and regressor submodule.
[0010] The term localization module works as follows: First, the CSPDarknet backbone network submodule extracts features from the ticket itinerary image to obtain three effective feature layers; then, the FPN enhanced feature extraction network submodule fuses the three effective feature layers obtained above to integrate feature information at different scales; finally, the YOLO Head classifier and regressor submodule obtains the coordinates of the detection box containing the term, the text region image block containing the term, and the category of the text region image block in the ticket itinerary image.
[0011] The super-resolution module consists of a character feature extraction and fusion submodule and a super-resolution image generator submodule;
[0012] The character feature extraction and fusion submodule is used to extract character structure features, character stroke features, and character semantic features from the text region image block, and fuse the above features to obtain the character fusion features in the text region image block; the character feature extraction and fusion submodule consists of a character structure feature extraction unit, a character stroke feature extraction unit, a character semantic feature extraction unit, and a character fusion feature processing unit;
[0013] The super-resolution image generator submodule utilizes the character fusion feature and, based on the diffusion generation model, performs super-resolution processing on the text region image block to obtain a clearer version of the text region image block.
[0014] The character structure feature extraction unit consists of a patch embedding subunit and three downsampling network subunits with the same structure connected sequentially. The downsampling network subunit includes a mixingblocks component and a merging component. The mixingblocks component is responsible for extracting character structure features, and the merging component is responsible for performing convolution and normalization operations on the character structure features obtained by the mixing blocks component, downsampling in the height direction, and outputting the final character structure features.
[0015] The working process of the character structure feature extraction unit is as follows: First, the patch embedding subunit uses a 4×4 convolutional kernel with a stride of 4 to downsample the text region image block, dividing the text region image block into small patches, and converting each small patch into a feature map; then, through the three downsampling network subunits, the height of the aforementioned feature map is downsampled to obtain the character structure features in the text region image block.
[0016] The mixingblocks subunit consists of two parts: a global feature mixing component and a local feature mixing component. The global feature mixing component adopts a transformer encoder structure and uses a multi-head self-attention mechanism to establish long-term dependencies between different characters and extract the overall features of the character structure. The local feature mixing component only considers the neighborhood of each character and uses a local feature self-attention mechanism. Each patch responds only to the pixels within a local window with a width and height of 4 ohms centered on it, and the remaining pixels are masked to extract detailed features of the character structure.
[0017] The specific process of the character stroke feature extraction unit extracting character stroke features from the text region image block is as follows: the text region image block is smoothed using median filtering to remove noise; then the noise-removed text region image block is binarized; finally, the skeleton of the obtained binary image is extracted to obtain the final character stroke features.
[0018] The specific process of the character semantic feature extraction unit extracting character semantic features from text region image blocks is as follows: the text region image block is input into a trained CRNN recognition network, which performs convolution and pooling operations on the text region image block to extract features from it; then the obtained features are input into a bidirectional LSTM network to generate a feature sequence of fixed length, where each element of the feature sequence is a feature vector representing the semantic prediction probability of a character, and finally the semantic features of all characters in the text region image block are extracted.
[0019] The specific working process of the character fusion feature processing unit is as follows: First, the character semantic features and character structural features are concatenated and input into the encoder sub-network of the transformer network. This encoder sub-network includes a multi-head self-attention layer and a feedforward network layer, which globally correlates the elements of the input features. The input of the decoder sub-network of the transformer network includes two parts: one is the output of the encoder sub-network, and the other is the new image features formed by convolutional fusion of the character stroke features and the text region image blocks. These features are fed into the multi-head self-attention layer and the feedforward network layer to learn the correlation between different features. In this way, the decoder sub-network can effectively fuse the correlation information between features to finally obtain the character fusion features.
[0020] The super-resolution image generator submodule consists of a U-Net network unit; the U-Net network unit uses the character fusion feature to perform iterative denoising processing on the text region image block, and finally obtains a clearer version of the text region image block.
[0021] The text recognition module is built on a convolutional recurrent neural network (CRNN), which consists of three parts: convolutional layer units, recurrent layer units, and transcription layer units. The convolutional layer units are responsible for extracting the feature sequences of the sharpened version of the text region image blocks. The recurrent layer units predict the label distribution of each feature vector in the feature sequence. The error of the recurrent layer units is backpropagated and finally transformed into a new feature sequence, which is then fed back to the convolutional layer units. The transcription layer units integrate the label distributions of the feature vectors predicted by the convolutional layer units and output the text recognition result. The recurrent layer units are composed of a bidirectional LSTM recurrent neural network.
[0022] This invention also proposes a method for recognizing passenger ticket itineraries using super-resolution technology, the method comprising the following steps:
[0023] (1) Input the ticket itinerary image into the term location module, detect the coordinates of the detection box containing the term in the ticket itinerary image, crop out the text area image block containing the term based on the detection box coordinates, and give the category of the text area image block; the term refers to name, flight number, ticket price, total, origin, destination, ID number, electronic ticket number, issuance date, fuel surcharge, insurance premium, civil aviation development fund, verification code, printing serial number, etc.
[0024] (2) The super-resolution module receives the output information of the term location module and performs super-resolution processing on the text region image block output by the term location module to obtain a clearer version of the text region image block.
[0025] (3) The text recognition module receives the cleared version of the text region image block output by the super-resolution module, performs text recognition, and obtains the recognition result.
[0026] The beneficial effects of this invention are as follows: The system and method proposed in this invention address the image blurring problem that often occurs in real-world scenarios. By using a super-resolution module, the text region is clarified. The super-resolution module uses an attention mechanism to combine image features with character fusion features, ignoring background information that accounts for a large proportion of the image, thereby achieving targeted clarification of the text region and greatly improving the recognition accuracy of passenger ticket itineraries. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of a ticket itinerary recognition system using super-resolution technology proposed in this invention;
[0028] Figure 2 This is a schematic diagram of the term location module proposed in this invention;
[0029] Figure 3 This is an image of a passenger ticket itinerary in an embodiment of the present invention;
[0030] Figure 4 In the embodiments Figure 3 The image shown is a visualization of the ticket itinerary image after processing by the term location module.
[0031] Figure 5 This is a schematic diagram of the structure of the super-resolution module proposed in this invention;
[0032] Figure 6 This is a schematic diagram of the character feature extraction and fusion submodule proposed in this invention;
[0033] Figure 7 This is an example diagram of a text region image block in an embodiment of the present invention;
[0034] Figure 8 The local feature blending component in the embodiment is... Figure 7 A schematic diagram illustrating the processing of the text region image block shown;
[0035] Figure 9 This is an example of... Figure 7 The stroke feature map obtained by extracting the skeleton from the text region image block shown;
[0036] Figure 10 This is an example of... Figure 7 The image block showing the text region is a sharpened version obtained after super-resolution processing.
[0037] Figure 11 This is a flowchart illustrating a method for recognizing passenger ticket itineraries using super-resolution technology proposed in this invention. Detailed Implementation
[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.
[0039] See Figure 1 This invention introduces a ticket itinerary recognition system using super-resolution technology, the system comprising the following modules:
[0040] The term location module detects the coordinates of the detection box containing terms in the ticket itinerary image, crops the text region image block containing the term based on the detection box coordinates, and assigns the category of the text region image block. The category of the text region image block refers to the category of the term contained in the text region image block. The term refers to names, flight numbers, fares, totals, origin, destination, ID card numbers, e-ticket numbers, issuance dates, fuel surcharges, insurance premiums, civil aviation development funds, verification codes, printing serial numbers, etc.
[0041] Super-resolution module: This module receives the output information from the term location module and performs super-resolution processing on the text region image blocks output by the term location module to obtain a clearer version of the text region image blocks;
[0042] Text recognition module: The function of this module is to receive the sharpened version of the text region image block output by the super-resolution module, perform text recognition, and obtain the recognition result.
[0043] See Figure 2 In this embodiment, the YOLO V5 object detection model is used to construct the term localization module. The term localization module consists of a backbone network CSPDarknet submodule, an enhanced feature extraction network FPN submodule, and a YOLO Head classifier and regressor submodule. For detailed information on YOLO V5, please refer to the following link: https: / / github.com / ultralytics / yolov5.
[0044] The term localization module works as follows: First, the CSPDarknet backbone network submodule extracts features from the ticket itinerary image to obtain three effective feature layers; then, the FPN enhanced feature extraction network submodule fuses the three effective feature layers obtained above to integrate feature information at different scales; finally, the YOLO Head classifier and regressor submodule obtains the coordinates of the detection box containing the term, the text region image block containing the term, and the category of the text region image block in the ticket itinerary image.
[0045] Figure 3 The image shown is an image of a passenger ticket itinerary. Figure 4 The term location module is for... Figure 3 The visualization results after image processing are shown: all detection boxes are drawn in the image according to the coordinates of the detection boxes, and the category and confidence level are marked on each detection box.
[0046] See Figure 5 The super-resolution module consists of a character feature extraction and fusion submodule and a super-resolution image generator submodule.
[0047] The character feature extraction and fusion submodule is used to extract character structure features, character stroke features, and character semantic features from the text region image block, and fuse the above features to obtain the character fusion features in the text region image block;
[0048] See Figure 6 The character feature extraction and fusion submodule consists of a character structure feature extraction unit, a character stroke feature extraction unit, a character semantic feature extraction unit, and a character fusion feature processing unit.
[0049] The super-resolution image generator submodule utilizes the character fusion feature and, based on the diffusion generation model, performs super-resolution processing on the text region image block to obtain a clearer version of the text region image block.
[0050] The character structure feature extraction unit consists of a patch embedding subunit and three downsampling network subunits with the same structure connected sequentially. The downsampling network subunit includes a mixingblocks component and a merging component. The mixingblocks component is responsible for extracting character structure features, and the merging component is responsible for performing convolution and normalization operations on the character structure features obtained by the mixing blocks component, downsampling in the height direction, and outputting the final character structure features.
[0051] The working process of the character structure feature extraction unit is as follows: First, the patch embedding subunit uses a 4×4 convolutional kernel with a stride of 4 to downsample the text region image block, dividing the text region image block into small patches, and converting each small patch into a feature map; then, through the three downsampling network subunits, the height of the aforementioned feature map is downsampled to obtain the character structure features in the text region image block.
[0052] Figure 7 The image shown is an example of a text region image patch in an embodiment of the present invention. In this embodiment, all text region image patches are uniformly sized to 128×32 pixels. After a 4×4 convolution with a stride of 4, the width and height become 32×8, meaning each patch is 4×4 in size, and each patch undergoes convolution calculation independently. The downsampling network subunit is a convolutional layer with a kernel width and height of 1×2. Each downsampling layer is responsible for compressing the height to half of its original value. After three downsampling layers, the width remains unchanged, while the height changes from 8 to 1. The final feature vector is a vector with a height of 1 and a width of 32, containing multiple channels, used to represent character structure features.
[0053] The mixingblocks subunit consists of two parts: a global feature mixing component and a local feature mixing component. The global feature mixing component adopts a transformer encoder structure and uses a multi-head self-attention mechanism to establish long-term dependencies between different characters and extract the overall features of the character structure. The local feature mixing component only considers the neighborhood of each character and uses a local feature self-attention mechanism. Each patch responds only to the pixels within a local window with a width and height of 4 ohms centered on it, and the remaining pixels are masked to extract detailed features of the character structure.
[0054] See Figure 8 In this embodiment, the Global Mixing component... Figure 8 Multi-head attention computation is performed on the entire text region image block shown; the Local Mixing component only applies to... Figure 8 The self-attention calculation is performed on the area within the black box of the entire text region image block shown. After the calculation is completed, the black box moves from left to right and from top to bottom, and the self-attention calculation is performed only within the box each time.
[0055] The specific process of the character stroke feature extraction unit extracting character stroke features from the text region image block is as follows: The text region image block is smoothed using median filtering to remove noise; then, the noise-removed text region image block is binarized; finally, the resulting binary image is subjected to skeleton extraction to obtain the final character stroke features. In this embodiment, the skeleton extraction uses the Zhang-Suen algorithm. Detailed information about this algorithm can be found in the literature: Zhang TY, Suen CYA fast parallel algorithm for thinning digital patterns[J]. Communications of the ACM, 1984, 27(3):236-239. Figure 9 In this embodiment, the character stroke feature extraction unit extracts the character stroke features. Figure 7 The stroke feature map is obtained by extracting the skeleton from the text region image block shown.
[0056] The specific process of the character semantic feature extraction unit extracting character semantic features from text region image blocks is as follows: the text region image block is input into a trained CRNN (Convolutional Recurrent Neural Network) recognition network, which performs convolution and pooling operations on the text region image block to extract features from it; then the obtained features are input into a bidirectional LSTM network to generate a feature sequence of fixed length, where each element of the feature sequence is a feature vector representing the semantic prediction probability of a character, and finally the semantic features of all characters in the text region image block are extracted.
[0057] In the embodiment, Figure 7 The text region image block is input into a trained CRNN (Convolutional Recurrent Neural Network) recognition network to obtain a feature vector representing the semantic prediction probability of the character. Specifically, the CRNN consists of three parts: convolutional layers, recurrent layers, and transcription layers. The recurrent layer unit is composed of a bidirectional LSTM recurrent neural network. First, the text region image block is input into a trained CRNN (Convolutional Recurrent Neural Network) to obtain a feature vector representing the semantic prediction probability of the character. Figure 7 The text region image block is input into the CRNN convolutional layer to extract features and obtain a feature sequence; then the feature sequence is input into a bidirectional LSTM network to capture semantic information in the sequence through information transfer over time; the transcription layer maps the features output by the LSTM recurrent layer to the probability distribution of character categories. Figure 7After the above processing, the character with the highest probability of character category is extracted, which can be obtained as T---V---9---8------5----4---, representing the character with the highest probability value of each predicted character at each width position. "-" indicates that no character was predicted at that position. After removing "-", the final recognition result is TV9854. However, in order to facilitate subsequent feature integration, only the character category probability distribution feature vector output by the LSTM recurrent layer is extracted, instead of the final recognition result TV9854. More detailed information about CRNN can be found in: Shi B, Bai X, Yao C. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition[J].IEEE transactions on pattern analysis and machine intelligence,2016,39(11):2298-2304.
[0058] The specific working process of the character fusion feature processing unit is as follows: First, the character semantic features and character structural features are concatenated and input into the encoder sub-network of the transformer network. This encoder sub-network includes a multi-head self-attention layer and a feedforward network layer, which globally correlates the elements of the input features. The input of the decoder sub-network of the transformer network includes two parts: one is the output of the encoder sub-network, and the other is the new image features formed by convolutional fusion of the character stroke features and the text region image blocks. These features are fed into the multi-head self-attention layer and the feedforward network layer to learn the correlation between different features. In this way, the decoder sub-network can effectively fuse the correlation information between features to finally obtain the character fusion features.
[0059] The super-resolution image generator submodule consists of a U-Net network unit. This U-Net network unit uses the character fusion features to iteratively denoise the text region image blocks, finally obtaining a clearer version of the text region image blocks. Detailed information about the U-Net network can be found in the following two references: Reference 1: Dhariwal P, Nichol A. Diffusion models beat gans on image synthesis[J]. Advances in neural information processing systems, 2021, 34: 8780-8794 and Reference 2: Ronneberger O, Fischer P, Brox TU-net: Convolutional networks for biomedical imagesegmentation[C] / / Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015: 234-241.
[0060] The text recognition module is built on a convolutional recurrent neural network (CRNN), which consists of three parts: convolutional layer units, recurrent layer units, and transcription layer units. The convolutional layer units are responsible for extracting the feature sequences of the sharpened version of the text region image blocks. The recurrent layer units predict the label distribution of each feature vector in the feature sequence. The error of the recurrent layer units is backpropagated and finally transformed into a new feature sequence, which is then fed back to the convolutional layer units. The transcription layer units integrate the label distributions of the feature vectors predicted by the convolutional layer units and output the text recognition result. The recurrent layer units are composed of a bidirectional LSTM recurrent neural network.
[0061] In the embodiment, Figure 10 The pair shown Figure 7 The image block of the text region shown is processed by super-resolution to obtain a clearer version. After being input into the text recognition module, the final text recognition result is TV9854.
[0062] See Figure 11The present invention also proposes a method for recognizing passenger ticket itineraries using super-resolution technology, the method comprising the following steps:
[0063] (1) Image of passenger ticket itinerary ( Figure 3 The data is input into the term localization module, which detects the coordinates of the detection box containing the term in the ticket itinerary image. Based on these coordinates, the text region image block containing the term is cropped, and the category of the text region image block is given. The term refers to name, flight number, fare, total, origin, destination, ID number, e-ticket number, issuance date, fuel surcharge, insurance premium, civil aviation development fund, verification code, printing serial number, etc. (see...) Figure 4 );
[0064] (2) The super-resolution module receives the output information from the term localization module and performs super-resolution processing on the text region image blocks output by the term localization module to obtain a clearer version of the text region image blocks (see...). Figure 10 );
[0065] (3) The text recognition module receives the sharpened version of the text region image block output by the super-resolution module, performs text recognition, and obtains the recognition result. See Table 1, which shows the results. Figure 3 The recognition results of the Chinese terms.
[0066] Table 1
[0067]
[0068] The inventors conducted extensive experiments on the system and method proposed in this invention. In these experiments, they collected images of passenger ticket itineraries taken in real-world scenarios. The inventors labeled the locations of various information entries in the images, including name, flight number, fare, total, origin, destination, ID number, e-ticket number, issuance date, fuel surcharge, insurance premium, civil aviation development fund, verification code, and printing serial number. They also recorded the center coordinates, width, height, and category of each key information entry. Furthermore, to improve the system's detection capabilities, data augmentation techniques such as flipping and scaling were applied to the image data. A separate Chinese dataset was collected, and images were processed using Gaussian blur and other techniques to create original image-low-resolution (HR-LR) image pairs. The data was then divided into training, validation, and test sets in an 8:1:1 ratio. The experiments yielded excellent results, demonstrating the effectiveness of the proposed system and method.
Claims
1. A ticket itinerary recognition system using super-resolution technology, characterized in that: The system includes the following modules: The term location module detects the coordinates of the detection box containing terms in the ticket itinerary image, crops the text region image block containing the term based on the detection box coordinates, and assigns the category of the text region image block. The category of the text region image block refers to the category of the term contained in the text region image block. The term refers to name, flight number, fare, total, origin, destination, ID number, e-ticket number, issuance date, fuel surcharge, insurance premium, civil aviation development fund, verification code, and printing serial number. Super-resolution module: This module receives the output information from the term location module and performs super-resolution processing on the text region image blocks output by the term location module to obtain a clearer version of the text region image blocks; The super-resolution module consists of a character feature extraction and fusion submodule and a super-resolution image generator submodule; The character feature extraction and fusion submodule is used to extract character structure features, character stroke features, and character semantic features from the text region image block, and fuse the above features to obtain the character fusion features in the text region image block; the character feature extraction and fusion submodule consists of a character structure feature extraction unit, a character stroke feature extraction unit, a character semantic feature extraction unit, and a character fusion feature processing unit; The specific working process of the character fusion feature processing unit is as follows: First, the character semantic features and character structural features are concatenated and input into the encoder sub-network of the transformer network. The encoder sub-network includes a multi-head self-attention layer and a feedforward network layer, which globally associates the elements of the input features. The input of the decoder sub-network of the transformer network includes two parts: one is the output of the encoder sub-network, and the other is the new image features formed by the convolutional fusion of the character stroke features and the text region image blocks. These features are fed into the multi-head self-attention layer and the feedforward network layer to learn the correlation between different features. In this way, the decoder subnetwork effectively fuses the correlation information between features, ultimately obtaining the character fusion feature; The super-resolution image generator submodule utilizes the character fusion feature and, based on the diffusion generation model, performs super-resolution processing on the text region image block to obtain a clearer version of the text region image block. Text recognition module: The function of this module is to receive the sharpened version of the text region image block output by the super-resolution module, perform text recognition, and obtain the recognition result.
2. The ticket itinerary recognition system using super-resolution technology according to claim 1, characterized in that: The term localization module consists of a backbone network CSPDarknet submodule, an enhanced feature extraction network FPN submodule, and a YOLOHead classifier and regressor submodule. The term localization module works as follows: First, the CSPDarknet backbone network submodule extracts features from the ticket itinerary image to obtain three effective feature layers; then, the FPN enhanced feature extraction network submodule fuses the three effective feature layers obtained above to integrate feature information at different scales; finally, the YOLO Head classifier and regressor submodule obtains the coordinates of the detection box containing the term, the text region image block containing the term, and the category of the text region image block in the ticket itinerary image.
3. A ticket itinerary recognition system using super-resolution technology according to claim 1, characterized in that: The character structure feature extraction unit consists of a patch embedding subunit and three downsampling network subunits with the same structure connected sequentially. The downsampling network subunit includes a mixingblocks component and a merging component. The mixingblocks component is responsible for extracting character structure features, and the merging component is responsible for performing convolution and normalization operations on the character structure features obtained by the mixingblocks component, downsampling in the height direction, and outputting the final character structure features. The working process of the character structure feature extraction unit is as follows: First, the patch embedding subunit uses a 4×4 convolutional kernel with a stride of 4 to downsample the text region image block, dividing the text region image block into small patches, and converting each small patch into a feature map; then, through the three downsampling network subunits, the height of the aforementioned feature map is downsampled to obtain the character structure features in the text region image block.
4. A ticket itinerary recognition system using super-resolution technology according to claim 3, characterized in that: The mixingblocks component consists of two parts: a global feature mixing component and a local feature mixing component. The global feature mixing component adopts a transformer encoder structure and uses a multi-head self-attention mechanism to establish long-term dependencies between different characters and extract the overall features of the character structure. The local feature mixing component only considers the neighborhood of each character and uses a local feature self-attention mechanism. Each patch responds only to the pixels within a local window with a width and height of 4 ohms centered on it, and the remaining pixels are masked to extract detailed features of the character structure.
5. A ticket itinerary recognition system using super-resolution technology according to claim 1, characterized in that: The specific process of the character stroke feature extraction unit extracting character stroke features from the text region image block is as follows: the text region image block is smoothed using median filtering to remove noise; then the noise-removed text region image block is binarized; finally, the skeleton of the obtained binary image is extracted to obtain the final character stroke features. The specific process of the character semantic feature extraction unit extracting character semantic features from text region image blocks is as follows: the text region image block is input into a trained CRNN recognition network, which performs convolution and pooling operations on the text region image block to extract features from it; then the obtained features are input into a bidirectional LSTM network to generate a feature sequence of fixed length, where each element of the feature sequence is a feature vector representing the semantic prediction probability of a character, and finally the semantic features of all characters in the text region image block are extracted.
6. A ticket itinerary recognition system using super-resolution technology according to claim 1, characterized in that: The super-resolution image generator submodule consists of a U-Net network unit; the U-Net network unit uses the character fusion feature to perform iterative denoising processing on the text region image block, and finally obtains a clearer version of the text region image block.
7. A ticket itinerary recognition system using super-resolution technology according to claim 1, characterized in that: The text recognition module is built on a convolutional recurrent neural network (CRNN), which consists of three parts: convolutional layer units, recurrent layer units, and transcription layer units. The convolutional layer units are responsible for extracting the feature sequences of the sharpened version of the text region image blocks. The recurrent layer units predict the label distribution of each feature vector in the feature sequence. The error of the recurrent layer units is backpropagated and finally transformed into a new feature sequence, which is then fed back to the convolutional layer units. The transcription layer units integrate the label distributions of the feature vectors predicted by the convolutional layer units and output the text recognition result. The recurrent layer units are composed of a bidirectional LSTM recurrent neural network.
8. A method for recognizing passenger ticket itineraries using super-resolution technology based on the system of claim 1, characterized in that: The method includes the following steps: (1) Input the ticket itinerary image into the term location module, detect the coordinates of the detection box containing the term in the ticket itinerary image, crop out the text area image block containing the term based on the detection box coordinates, and give the category of the text area image block; the term refers to name, flight number, ticket price, total, origin, destination, ID number, electronic ticket number, issuance date, fuel surcharge, insurance premium, civil aviation development fund, verification code, and printing serial number; (2) The super-resolution module receives the output information of the term location module and performs super-resolution processing on the text region image block output by the term location module to obtain a clearer version of the text region image block. (3) The text recognition module receives the cleared version of the text region image block output by the super-resolution module, performs text recognition, and obtains the recognition result.