AI-powered communication big data-driven automatic identification methods, systems, media, and equipment for travel history cards.

Abstract

This application relates to an AI-powered method for automatically recognizing travel itinerary cards based on big data. It utilizes image detection and text recognition models to identify images of travel itinerary cards based on big data, obtaining at least a text list. The text list includes text, and a route search text list is obtained based on this list. A smart text segmentation engine is then used to extract the route text from the route search text list. Specifically, the text in the search text list is segmented according to segmentation rules to obtain several sub-texts. Each sub-text is further segmented, and each segmented string is compared with strings in a preset database to obtain a difference score. The sum of the difference scores of several strings for each sub-text is calculated to obtain a difference score result value. The string with the smallest difference score result value in the database is extracted as the route text after the sub-text is recognized. This method can improve the success rate and accuracy of route text extraction.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to AI-based communication big data automatic identification methods, systems, media, and devices for travel cards. Background Technology

[0002] To help people conveniently prove their travel history and improve the efficiency of travel verification for businesses, communities, and transportation departments, the China Academy of Information and Communications Technology (CAICT) under the Ministry of Industry and Information Technology, in conjunction with the three major telecom operators (China Telecom, China Unicom, and China Mobile), has launched a communication big data travel card. This card provides travel query services to 1.6 billion mobile phone users nationwide, utilizing the communication big data of these three major telecom operators. It supports one-click querying of countries (regions) visited and domestic cities (staying for more than 4 hours) within the past 14 days.

[0003] The method of restricting the travel of people passing through medium- and high-risk areas by viewing and filtering individual dynamic travel card information has been used by many regions and organizations across the country.

[0004] Currently, the common method for ordinary users is to take a screenshot of their travel itinerary card using their mobile phone and send it to the service provider that collects the data. While mobile phone screenshots ensure image clarity and reduce the difficulty of information extraction, in situations with high network security requirements and where access to big data travel data services is not possible, this simple method is not feasible. For example, using an app would be a lengthy and inefficient process. Therefore, a simpler method is to use a camera to photograph the phone screen and then obtain the travel itinerary card information. However, this method has drawbacks: the camera's environment and the user's improper phone placement can lead to unclear images or missing information.

[0005] Currently, there are no existing products or technologies for extracting information from travel history code images. If OCR models are used directly to extract information, firstly, OCR models can only output the recognized text and text location, and cannot directly obtain structured information such as phone numbers, update dates, transit points, and colors from the travel history code; secondly, due to factors such as the shooting environment, OCR models suffer from omissions and misidentifications, resulting in low success rates and accuracy in extracting structured information from the travel history code. Therefore, the inventors believe that this problem urgently needs to be solved.

[0006] To improve the success rate and accuracy of travel code information extraction, this application provides an AI-powered method, apparatus, medium, and equipment for automatic identification of travel cards using big data communication technology.

[0007] Firstly, the AI-powered communication big data-based automatic identification method for travel cards provided in this application adopts the following technical solution:

[0008] The AI-powered method for automatically recognizing travel history cards based on big data communication includes the following steps:

[0009] Image detection and text recognition models are used to identify images from travel history cards based on big data, resulting in at least a text list, which includes at least text.

[0010] Search the text list based on the method described above.

[0011] Based on the search list of texts along the aforementioned path, the intelligent text segmentation engine is used to extract the text along the path, including the following steps:

[0012] The text in the searched text list is segmented according to segmentation rules to obtain several sub-texts.

[0013] Each subtext is segmented into several strings. Each segmented string is compared with strings in a pre-defined database to obtain a difference score. The sum of the difference scores of the several strings in each subtext is calculated to obtain a difference score result value. The database strings include region name strings.

[0014] Extract the strings corresponding to the smallest difference result values ​​from the database as the path text after the subtext is identified.

[0015] The above technical solution uses an OCR model or other image detection and recognition model to identify the text list from the image. Based on the position of the route location text on the travel card, the text range is narrowed down to obtain a route location search text list. However, since the route location text information in the search text list may be very long, a database storing region name strings is pre-set. The text in the route location search text list is segmented using segmentation rules to obtain several sub-texts. Each sub-text is further segmented to obtain multiple segmentation results. The difference degree between each segmented string and the region name string in the database is calculated. The difference degrees corresponding to all strings of a sub-text are summed to obtain the difference degree result value. If the difference degree result value is the smallest, it indicates a higher matching degree with the string in the database. Therefore, the string with the smallest difference degree result value in the database is extracted as the route location text after the sub-text is recognized. In this process, omissions or errors in the sub-text can be automatically corrected, which can solve the problems of text recognition errors and inability to extract structured information in general OCR models, thereby improving the success rate and accuracy of route location text extraction.

[0016] In conjunction with the first aspect, in some implementations of the first aspect, the segmentation rule includes: using a sliding window to select and segment the text to be extracted to obtain several sub-texts, wherein the length of the sliding window is equal to the longest string in the preset database plus N characters, and N is an integer not less than 1.

[0017] By using the above technical solutions and setting the length of the sliding window, we can ensure that the city name is completely selected, so that the segmented sub-text can reflect the complete city name.

[0018] In conjunction with the first aspect, in some implementations of the first aspect, the step of splitting each of the sub-texts into several strings includes: splitting each sub-text string once to obtain two strings.

[0019] The above technical solution obtains two strings through a single segmentation, resulting in the minimum difference value. This single segmentation scheme produces strings that represent the complete path text, achieving a locally optimal partitioning result with fewer segmentation attempts.

[0020] In conjunction with the first aspect, in some implementations of the first aspect, the difference between each segmented string and the strings in the preset database is compared to obtain the difference degree, including: calculating the edit distance between each string and the strings in the preset database to obtain the difference degree.

[0021] Using the above technical solutions, the difference is measured by calculating the edit distance between each string and the strings in the preset database. This allows for a quick determination of the difference value, which is the value of the edit distance. The smaller the edit distance, the greater the similarity between the two strings.

[0022] In conjunction with the first aspect, in some implementations of the first aspect, the database also includes the frequency of each string. If there are multiple city strings with the same edit distance as the strings in the database, the string with the highest frequency is selected as the string with the smallest edit distance.

[0023] Using the above technical solution, when there is a typo in the text in the route search text list, there may be multiple region name strings with the same edit distance as the strings in the database. In this case, selecting the string with the highest frequency as the string with the smallest edit distance can obtain a more accurate route text.

[0024] In conjunction with the first aspect, in some implementations of the first aspect, if the minimum value of the difference result is greater than a preset value, then the database corresponding to the difference result value will not be extracted as the path text after the subtext is identified.

[0025] Using the above technical solutions, since the text may contain non-regional names, such as "You arrived or passed through: Beijing in the previous 14 days" which is non-regional name text, setting a preset value can identify and remove such text, thereby reducing the amount of text to be analyzed further.

[0026] In conjunction with the first aspect, in some implementations of the first aspect, the method for obtaining the text list by means of searching includes the following steps:

[0027] Image processing is used to obtain a "white" information area from the image of the travel data itinerary card. This information area contains text information such as mobile phone number, update date, and transit locations. The text information is then identified to form a sub-text list, which includes the mobile phone number, update date, transit locations, and the location coordinates of the text.

[0028] Based on the location coordinates of the text, the text below the update date in the sub-text list is obtained as the search text list by means of the method.

[0029] Using the above technical solution, since the truly useful information to be extracted from the itinerary card is all within the "white" information area, obtaining the text information within the "white" information area can solve the problem of not being able to determine the end of the text in the transit location text. It can also reduce the analysis of other useless text information, thereby improving the efficiency of extracting useful information. The text below the update date is the transit location text, which can be quickly located.

[0030] In conjunction with the first aspect, in some implementations of the first aspect, all text in the search text list is concatenated sequentially.

[0031] The above technical solution addresses the issue that the text detected and recognized by the image detection and recognition model is not a continuous string of text, but rather consists of many separate texts, each with its own location coordinates. Therefore, the texts need to be sequentially concatenated to form a continuous string of text before further text processing can proceed.

[0032] In conjunction with the first aspect, in some implementations of the first aspect, the text list also includes the confidence score of each text. If a text with a confidence score less than 0.8 appears in the text list, failure is returned directly.

[0033] If the confidence level is below 0.8 using the above technical solutions, it indicates that the text recognition result is unreliable and unusable.

[0034] In conjunction with the first aspect, in some implementations of the first aspect, if the number of characters in the first line of the search text list is less than 10, then it is removed.

[0035] By using the above technical solution, if the number of characters in the first line of the search text list is less than 10, the text will be removed, thus preventing the arrow in the middle from being recognized as text.

[0036] In conjunction with the first aspect, in some implementations of the first aspect, after recognizing the images of the travel data itinerary card using the image detection and recognition model to obtain at least a text list, the following steps are performed:

[0037] If the number of recognized text entries is less than 3, the text list will be rotated 180°.

[0038] Using the above technical solution, if the number of recognized text entries is less than 3, it means that the user's phone is upside down and the image of the trip card is reversed. Therefore, the text list needs to be rotated 180°.

[0039] In conjunction with the first aspect, in some implementations of the first aspect, after recognizing the images of the travel data itinerary card using the image detection and recognition model to obtain at least a text list, the following steps are performed:

[0040] Set a text box for each line of text in the text list, calculate the angle A between the long side of the text box and the horizontal direction of the image, and if the angle A is greater than 45°, it will not be included in the subsequent text extraction.

[0041] Using the above technical solutions, if the included angle is greater than 45°, it is considered misdirected text and will not be included in subsequent text extraction.

[0042] In conjunction with the first aspect, in some implementations of the first aspect, after recognizing the images of the travel data itinerary card using the image detection and recognition model to obtain at least a text list, the following steps are performed:

[0043] Set a text box for each line of text in the text list, and reorder the text in the text list from top to bottom and from left to right using the coordinates of the top left corner of each text box.

[0044] Using the above technical solution, since the text in the search text list may be incorrectly sorted, but the text's position coordinates are correct, the text can be moved to the correct position based on its position coordinates, thus obtaining the correct text list.

[0045] In conjunction with the first aspect, in some implementations of the first aspect, based on the text list and the image recognition of run code colors, the following steps are included:

[0046] First, determine whether the text "Please accept the green travel card" is contained in the text list. If the matching degree of the text is above the preset value and the position coordinates of the text are above the position coordinates of the mobile phone number, then determine that the color of the travel code is green; otherwise, continue to judge.

[0047] Based on the position coordinates of the text, if a single text "个" or "T" is recognized between the update date and the places passed through, then it is considered a green code; otherwise, continue to judge.

[0048] Identify the image. If there is an arrow "↑", then determine that the color of the travel code is green; otherwise, it is non - green.

[0049] Through the above technical solutions, only the green travel card has the text "Please accept the green travel card". In view of the problems such as reflection or missing characters that may occur in the captured image, if the matching degree of the text is above the preset value and the position coordinates of the text are above the position coordinates of the mobile phone number, it can be considered that the color of the travel card is green. Since only the green travel card has the arrow "↑", therefore, if a single text "个" or "T" can be recognized in the text and this text is between the update date and the places passed through, then it can be considered that the color of the travel card is green, or if an arrow "↑" can be recognized in the image, then it can also be considered that the color of the travel card is green. By providing multiple ways to identify the color of the travel card, it is possible to more accurately identify whether the color of the travel card is green or non - green, and the recognition result will not be unavailable due to certain defects in the image.

[0050] Combined with the first aspect, in some implementation manners of the first aspect, the identifying the image includes the following steps:

[0051] First, convert the picture information into the HSV color space and set a mask to convert non - green to black.

[0052] Then perform binarization and Gaussian blur processing.

[0053] Further perform edge detection.

[0054] Further perform erosion and dilation.

[0055] Finally, remove the circular contour, find the contour with a perimeter within the preset range (the number of pixels is from 300 to 2000) and obtain the circumscribed polygon of this contour, obtain the convex hull of the circumscribed polygon, find that the number of side lengths of the convex hull is between 3 and 6 and there are two extra sides in the circumscribed polygon inside the convex hull, then judge that this circumscribed polygon is the arrow "↑".

[0056] Through the above technical solution, the picture information is converted into the HSV color space and a mask is set to convert non-green colors into black, which is equivalent to only extracting the green part, reducing the interference of other colors on edge detection. Then, binarization and Gaussian blur processing are performed to reduce noise interference. Finally, erosion and dilation are carried out to reduce small cracks and increase edge connectivity.

[0057] Combined with the first aspect, in some implementation manners of the first aspect, the text list further includes the position coordinates of the four vertices of the text box where the text is located. Based on the text list, the mobile phone number is extracted. The extraction method of the mobile phone number includes the following steps:

[0058] Traverse the text list, take the Chinese character text in the text list, calculate the edit distance between the Chinese character text and the text "Dynamic Travel Card". If the edit distance is less than 2 and the position coordinates of the text are in the upper half of the image, then this line of text is designated as a candidate text. Take the numerical text in the text list. If the number of consecutive longest digits is greater than 4 and the total number of digits in the text is less than 12, then it is also listed as a candidate text.

[0059] Traverse the candidate texts, obtain all the digits in the candidate texts, and form a numerical text.

[0060] If the number of digits in the numerical text is greater than 6 and less than 12, and the first digit of the digits is 1, then take the first 3 and the last 4 digits of the numerical text, separated by "****" in the middle, as the final mobile phone number.

[0061] If the number of digits in the numerical text is greater than 3 and less than 7, then use the position coordinates of the text box for judgment. If there are 3 digits on the left side of the text box and the first digit is 1, then these 3 digits and the last 4 digits on the right side plus the "****" in the middle form the mobile phone number.

[0062] Otherwise, it indicates that the extraction of the mobile phone number fails and no further information extraction is performed.

[0063] Through the above technical solutions, (1) the edit distance between the Chinese text and the text "dynamic travel card" is less than 2, and the position coordinates of the Chinese text are in the upper half of the image, so the text can be directly used as a candidate text. (2) During the recognition process, due to the quality of the captured image, the model may recognize "****" as a number. Therefore, the case where the longest consecutive number is greater than 4 and the total number of numbers in the text is less than 12 is listed as a candidate text for the mobile phone number. (3) If the number of numbers in the digital text is greater than 6 and less than 12, it means that the text meets the number of numbers in the mobile phone number. If the first number is 1, it can be judged that the text is a mobile phone number. However, in fact, the first three and last four digits of the mobile phone number on the travel card are not numbers, but four stars "****". Therefore, the four digits in the middle are replaced with four stars "****" to obtain the correct travel card mobile phone number. (4) If the number of digits in the numeric text is greater than 3 and less than 7, the numeric text may only contain the last 4 digits of the mobile phone number, and the first 3 digits may be on the left side of the text. In this case, the position coordinates of the text box are used to make a judgment. If there are 3 digits on the left side of the text box and the first digit is 1, then the 3 digits and the last 4 digits on the right side plus the middle "****" form the mobile phone number.

[0064] In conjunction with the first aspect, in some implementations of the first aspect, the text list also includes the position coordinates of the four vertices of the text box containing the text, and the update date is extracted based on the text list. The method for extracting the update date includes the following steps:

[0065] Based on the text list, a search text list with the update date is obtained.

[0066] Iterate through the search text list for the update date. If the number of consecutive longest digits in the text is greater than 4, and the total number of digits is 13 or 14, and the edit distance between the non-numeric part of the text and the text "updated on:" is less than 2, then the text is directly identified as update date information; otherwise, the text is added as a candidate to the candidate list.

[0067] Iterate through the candidate list. If the total number of numbers is 13, add 0 to the end and convert it to time format. Compare the converted time with the current time. If the time difference is within 24 hours, the correct time is considered to have been extracted. Otherwise, the time update failed and no further information extraction is performed.

[0068] The above technical solution addresses the issue of missing the last digit in the update date. Instead, a zero is added to the last digit to complete the update date. Since itinerary card updates have a time limit (trips within 24 hours are considered valid), all converted times are compared to the current time. If the time difference is within 24 hours, the correct time is considered extracted. This method solves the problem of missing the last digit in the update date.

[0069] Secondly, this application provides an AI-powered communication big data travel card automatic recognition system, which includes the following modules:

[0070] Image recognition module: Used to recognize images of travel data cards using image detection and recognition models, and to obtain at least a text list, wherein the text list includes at least text.

[0071] The path-search text list acquisition module is used to obtain a path-search text list based on the text list.

[0072] The route location text extraction module is used to extract route location text based on the route location search text list using an intelligent text segmentation engine. The extraction of route location text using the intelligent text segmentation engine includes the following steps:

[0073] The text in the searched text list is segmented according to segmentation rules to obtain several sub-texts.

[0074] Each subtext is segmented into several strings. Each segmented string is compared with strings in a pre-defined database to obtain a difference score. The sum of the difference scores of the several strings in each subtext is calculated to obtain a difference score result value. The database strings include region name strings.

[0075] Extract the strings corresponding to the smallest difference result values ​​from the database as the path text after the subtext is identified.

[0076] Thirdly, this application provides a computer-readable storage medium for storing a computer program, which, when invoked, executes the AI ​​communication big data travel card automatic identification method in the first aspect or any possible implementation thereof.

[0077] Fourthly, this application provides a computer device including a memory and a processor, wherein the memory is used to store a computer program, and the computer program, when invoked by the processor, executes the AI ​​communication big data travel card automatic identification method in the first aspect or any possible implementation thereof.

[0078] In summary, this application includes at least one of the following beneficial technical effects:

[0079] (1) Use an OCR model or other image detection and recognition model to recognize the image to obtain a text list. Based on the position of the route text on the travel card, narrow the text range to obtain a route search text list. However, since the route text information in the route search text list may be very long, a database storing regional strings is preset first. The text in the route search text list is segmented using segmentation rules to obtain several sub-texts. Each sub-text obtained by segmentation is further divided to obtain multiple segmentation results. The difference degree between each segmented string and the string in the database is calculated. The difference degree of all strings of the same sub-text is summed to obtain the difference degree result value. If the difference degree result value is the smallest, it means that the matching degree with the string in the database is higher. Therefore, the string with the smallest difference degree result value in the database is extracted as the route text after the sub-text. In this process, if there are typos or omissions in the text list recognized by the model, they can be corrected. This can solve the problem of text recognition errors and inability to extract structured information in general OCR models, thereby improving the success rate and accuracy of route text extraction.

[0080] (2) Each subtext is split once to obtain two strings. By splitting once to obtain two strings, the splitting scheme with the smallest difference result value can be obtained. The string split by this scheme is the complete path text, thus achieving a local optimal partitioning result with fewer splitting times.

[0081] (3) The difference between each string after segmentation and the strings in the preset database is compared to obtain the difference. This includes: calculating the edit distance between each string and the strings in the preset database to obtain the difference. The difference is measured by calculating the edit distance between each string and the strings in the preset database, which can quickly obtain the value of the difference, that is, the value of the edit distance. The smaller the edit distance, the greater the similarity between the two strings, that is, the smaller the difference.

[0082] (4) Since the useful information that needs to be extracted from the itinerary card is in the "white" information area, obtaining the text information in the "white" information area can reduce the analysis of other useless text information and thus improve the extraction efficiency of useful information. The text below the update date is the transit text, which can be quickly located.

[0083] (5) Only the green travel card has the text "Please accept the green travel card". Given that the captured image may have problems such as reflection or missing characters, if the matching degree of the text is above the preset value and the position coordinates of the text are above the position coordinates of the mobile phone number, it can be considered that the color of the travel card is green. Since only the green travel card has the arrow "↑", therefore, if a single text "个" or "T" can be recognized in the text and this text is located between the update date and the places passed through, it can be considered that the color of the travel card is green, or if the arrow "↑" can be recognized in the image, it can also be considered that the color of the travel card is green. By providing multiple ways to identify the color of the travel card, the color of the travel card can be more accurately identified as green or non-green, and the recognition result will not be unavailable due to certain defects in the image.

[0084] (6) This application can also improve the success rate and accuracy of extracting the mobile phone number and the update date, and thus improve the success rate and accuracy of extracting the main information of the travel card for the big data of travel. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a recognition schematic diagram of the OCR model for extracting the main information of the travel card for the big data of travel in Embodiment 1 of this application.

[0086] Figure 2 It is an incomplete image captured in Embodiment 1 of this application.

[0087] Figure 3 It is an included angle schematic diagram of misaligned text in the image captured in Embodiment 1 of this application.

[0088] Figure 4 Flowchart for extracting the mobile phone number in Embodiment 1 of this application.

[0089] Figure 5 It is a flowchart for extracting the update date in Embodiment 1 of this application.

[0090] Figure 6 It is a schematic diagram for obtaining the information area of "white" in Embodiment 1 of this application.

[0091] Figure 7 It is a schematic diagram of the initialized database in Embodiment 1 of this application.

[0092] Figure 8 It is an image of the travel card for the big data of travel with an arrow to be extracted in Embodiment 1 of this application.

[0093] Figure 9 It is for Figure 8 The result diagram after the first step of processing.

[0094] Figure 10This is the counterpart of Embodiment 1 of this application. Figure 9 The image after binarization and Gaussian blurring.

[0095] Figure 11 This is the counterpart of Embodiment 1 of this application. Figure 10 The image obtained by edge detection.

[0096] Figure 12 This is the counterpart of Embodiment 1 of this application. Figure 11 The image obtained by erosion and expansion. Detailed Implementation

[0097] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1 to 12 The present application will be further described in detail with reference to the embodiments.

[0098] Example 1.

[0099] The AI-powered travel data travel card automatic recognition method in this embodiment uses an OCR model to recognize the image of the travel data travel card. First, a camera is used to take a picture of the travel data travel card on the user's mobile phone to obtain the image of the travel data travel card. The image is then transmitted to the OCR model for recognition to obtain a text list.

[0100] The OCR module (OCR model) uses a deep learning-based text detection and recognition model to identify text in images and the location of the text.

[0101] Specifically, it is mainly based on deep learning. It includes two parts: text detection and text recognition. The text detection part uses the SAST model with a ResNet50 backbone. The recognition model uses the SRN model with a ResNet50 backbone. In practical applications, the OCR model can be replaced with other models.

[0102] To improve model accuracy, this embodiment annotates images from actual travel data itinerary cards before performing transfer learning training. The annotation method involves labeling all text appearing in the images, including both the text content and its location, such as... Figure 1 As shown.

[0103] The number of images selected should be no less than 200,000. During training, the last 5 layers of the detection model and the last 8 layers of the recognition model are retrained, while the other layers are kept fixed, thereby accelerating the training speed and completing transfer learning.

[0104] The OCR model takes an RGB image as input and outputs the text content, location, and confidence score. For example... Figure 2 The following is a list of text output by the OCR model for an incomplete image:

[0105] [{"val": "Dynamic itinerary card of 188****8958", "prop": 0.99426204, "bbox": [[327.0, 38.0], [570.0, 40.0], [570.0, 65.0], [327.0, 64.0]]}]

[0106] {"val": "Updated on: 2022.04.22 15:55:30", "prop": 0.9443428, "bbox": [[292.0, 85.0], [608.0, 81.0], [608.0, 108.0], [292.0, 112.0]]}

[0107] {"val": "You arrived or passed through: Wuhan, Hubei within the past 14 days", "prop": 0.99921364, "bbox": [[272.0, 402.0], [651.0, 380.0], [652.0, 408.0], [274.0, 429.0]]}

[0108] {"val": "City", "prop": 0.99968636, "bbox": [[277.0, 432.0], [300.0, 432.0], [300.0, 453.0], [277.0, 453.0]]}

[0109] {"val": "The result includes the countries (regions) you visited within the past 14 days and stays over 4 hours", "prop": 0.96395385, "bbox": [[258.0, 481.0], [624.0, 458.0], [626.0, 482.0], [259.0, 506.0]]}

[0110] {"val": "The color card only reminds of the visited places", "prop": 0.95303756, "bbox": [[352.0, 506.0], [487.0, 498.0], [488.0, 515.0], [353.0, 523.0]]}

[0111] {"val": "journey and is not related to health status", "prop": 0.9349943, "bbox": [[477.0, 498.0], [586.0, 488.0], [588.0, 505.0], [478.0, 515.0]]}

[0112] {"val": "This service is jointly provided", "prop": 0.976952 , "bbox":[[417.0, 534.0],[526.0, 525.0],[528.0, 545.0],[419.0, 554.0]]}

[0113] {"val": "CAICI", "prop": 0.9229716 , "bbox":[[290.0, 577.0],[331.0,577.0],[331.0, 590.0],[290.0, 590.0]]}]

[0114] The format is JSON. In each text entry, the `val` field represents the content, the `prop` field represents the confidence level, and `bbox` represents the pixel coordinates of the four vertices of the text box. In this scheme, a text box refers to the smallest quadrilateral that encloses the text.

[0115] Because the above text list contains a lot of useless information, this solution pre-extracts the "white" information area from the travel card image to remove it. The "white" information area contains all the information that needs to be extracted. The process of obtaining the "white" information area is as follows: read the image, perform binarization, perform median filtering, perform edge detection, obtain the smallest contour among the contours containing the phone number and update date in the bounding polygon, and obtain the smallest bounding rectangle of the contour. This smallest bounding rectangle is the corresponding "white" information area. Figure 6 As shown, the final white outline is the obtained rectangle, and the text in the "white" information area is used to form a sub-text list.

[0116] Next, the sub-text list is processed.

[0117] The preliminary processing module mainly includes several parts: reverse text detection, misdirected text removal, and text position sorting.

[0118] The reverse text detection primarily addresses the issue of images rotating 180° due to the phone being upside down. This solution uses a simple method of counting the number of text entries. If the number of recognized text entries is less than 3, the image is considered potentially reversed, and the OCR model is requested to rotate the image 180°. If the recognition result remains unchanged, the image is not further processed. The case of less than 3 text entries includes situations where the number of recognized text entries is 0 (a result caused by the image being upside down). After requesting the OCR model to rotate the image 180° (i.e., turning the image to the correct orientation), the recognition result will change, and further processing will continue.

[0119] Misaligned text removal primarily eliminates interference caused by image distortion and the presence of non-run-code text in some images. This method calculates the angle between the long side of the text box and the horizontal direction of the image. If the angle is greater than 45°, it is considered misaligned text. Figure 3 As shown, misdirected text is not included in subsequent processing.

[0120] The sub-text list may contain text in incorrect positions (even if the coordinates are correct), so text sorting is necessary to facilitate subsequent processing. Using the coordinates of the top-left corner of each text box, the text list is sorted from top to bottom and left to right.

[0121] Next, extract the mobile phone number, update date, transit locations, and trip card color in sequence.

[0122] Mobile phone number extraction module:

[0123] Text feature description:

[0124] The text is a single line, starting with the number 1, consisting of 7 digits. The first 3 digits and the last 4 digits are separated by 4 asterisks "****". The text following this is the fixed text "Dynamic Itinerary Card".

[0125] Identify the existing problems:

[0126] • When taking the picture, the phone was positioned too high, or the picture was not a travel code image and did not contain a phone number;

[0127] • The asterisk may not be recognized, or it may be identified as other symbols, or even numbers;

[0128] • Blurry images can lead to incorrect or incomplete digit recognition;

[0129] • Blurry images cause the text "Dynamic Travel Card" to be misrecognized or incomplete;

[0130] • The recognition will break at the asterisk, meaning the phone number will be recognized as two separate text entries;

[0131] Implementation method: such as Figure 4 As shown, since the text below the phone number includes at least one update date text and one transit location text, the phone number is not retrieved from the last two texts. Therefore, the search text list for the phone number is taken from the entire sub-text list excluding the last two texts.

[0132] Traverse the search text list of mobile phone numbers, extract the Chinese characters in the text, calculate the edit distance between the Chinese characters and the text "Dynamic Travel Card", and if the edit distance is less than 2 and the position is in the upper half of the image, then determine the line text as the candidate text. If the number of consecutive longest digits is greater than 4 and the total number of digits in the text is less than 12, it is also listed as a candidate text.

[0133] Traverse the candidate text, obtain all the digits in the candidate text, and form a digit text.

[0134] If the middle "****" is recognized as a digit, then the number of digits in the digit text is greater than 6 and less than 12. If the first digit of the digits is 1, then take the first 3 and the last 4 digits of the digit text, separated by "****" in the middle, as the final mobile phone number.

[0135] If there are the following situations:

[0136] 1. Some of the middle asterisks may be recognized as digits;

[0137] 2. The asterisks may not be recognized, resulting in the mobile phone number being separated due to too large a distance in the middle.

[0138] For example, the middle asterisks in 123****4567 may not be recognized, and in this case, the text is separated into 123 and 4567. It is also possible that some asterisks are recognized as digits, such as being recognized as 123 and 04567.

[0139] So there are three possibilities of recognizing 4 / 5 / 6 digits.

[0140] Therefore, if the number of digits in the digit text is greater than 3 and less than 7, then the digit text may only contain the last 4 digits of the mobile phone number, and the first 3 digits may be on the left side of this text. Use the position of the text box to judge. If there are 3 digits on the left side of the text box and the first digit is 1, then these 3 digits and the last 4 digits on the right side plus the "****" in the middle form the mobile phone number.

[0141] Other situations indicate that the extraction of the mobile phone number fails, and no subsequent information extraction is performed.

[0142] Update date extraction module:

[0143] Text feature description:

[0144] Single-line text. The fixed text is "Updated on:", and the date is after the text, in the format of "2022.05.01 00:00:00", and the position is below the mobile phone number and in the upper half of the image.

[0145] Recognition problems:

[0146] • It cannot recognize periods, spaces, and colons.

[0147] • Due to the space after the full-width colon in the middle, the Chinese characters and the following date may be split into two separate texts.

[0148] • Fixed text may be misrecognized.

[0149] • The last digit may be lost.

[0150] Implementation method:

[0151] like Figure 5 As shown, since there is at least one transit point text below the update date, and it is below the phone number, the phone number is not retrieved from the last text in the subtext list. Assuming the phone number is the m-th text, the part between the (m+1)-th row and the last text in the entire subtext list is taken as the search text list for the update date.

[0152] Iterate through the list of search texts with the update date. If the number of consecutive longest digits in the text is greater than 4 and the total number of digits is 13 or 14, and the edit distance between the non-numeric part of the text and the text "updated on:" is less than 2, then the text is directly identified as the only candidate. Otherwise, the text is simply added to the candidate list.

[0153] Iterate through the candidate list; if the total number of numbers is 13, add a 0 to the end. Then convert it to a time format. Since the update time of the itinerary card has a time limit, compare the converted time with the current time. If the time difference is within 24 hours, the correct time is considered to have been extracted. Otherwise, it indicates that the update time extraction failed, and no further information extraction is performed.

[0154] Route extraction module:

[0155] Text feature description:

[0156] The single-line text, with the fixed phrase "You arrived or transited through: within the previous 14 days," is followed by one or more lines of region names. Region names are listed down to the second-level administrative division. Risk regions are marked with an asterisk (*), and all region names have a fixed note at the end. See below:

[0157] You arrived in or transited through: Beijing*, Tianjin (Note: *) within the previous 14 days.

[0158] This indicates that there are currently medium- or high-risk areas in the city, and does not...

[0159] This indicates that the user has actually visited these medium- and high-risk areas.

[0160] Problems were identified:

[0161] • The fixed text “You arrived or transited through within the previous 14 days:” is in gray text, which may lead to misidentification or failure to detect it at all.

[0162] • Because the fixed text “You arrived or passed through in the previous 14 days:” ends with a full-width colon, creating a large gap, it may be recognized as two separate texts with the following area name, making it difficult to locate the starting point of the transit point.

[0163] • The separator between regions is a comma, which is very likely to be unrecognizable.

[0164] • Region names with many strokes are very likely to be misidentified.

[0165] • It may be impossible to determine when the route ends just by reading the text.

[0166] The recognition method is as follows:

[0167] The problem of not being able to determine the end of the text can be solved by pre-extracting the "white" information area as described above.

[0168] If the bounding rectangle cannot be obtained, a failure is returned and no further detection is performed (Note: For images with long routes that cannot be fully displayed in one image, special stitching processing is required).

[0169] The search algorithm retrieves a list of text from the area below the update date, within the "white" information zone. To prevent the arrow in the middle from being recognized as text, text with fewer than 10 characters in the first line is removed. Additionally, if any text in the list has a confidence level less than 0.8, the algorithm returns a failure.

[0170] The text is concatenated sequentially from the sub-text list and sent to the intelligent text segmentation engine to obtain the segmented and matched text. Strings that cannot be matched at the beginning or end of the string (the database stores location strings; strings like "You arrived or passed through within the last 14 days" are unmatchable) are removed. Then, each segmented string is checked for multiple matches. If multiple matches are found, the process returns failure; otherwise, the segmentation and matching are successful.

[0171] Intelligent text segmentation engine:

[0172] This embodiment uses a database-based and locally optimal method to achieve intelligent text segmentation, which can address issues such as misspellings and omissions in the text.

[0173] First, add the region names and their corresponding frequencies of occurrence to the database. The region names are the same as those used in the actual itinerary code, such as Beijing City and Turpan City, Xinjiang Uygur Autonomous Region, without splitting. The word frequency is initialized to 1, using the number of times the word appears. The database will update the frequency of the region names in real time according to the actual processed data. The initialized database is as Figure 7 shown ( Figure 7 only showing a part of the database content).

[0174] If a text contains N characters, there are 2N - 1 possible ways of splitting. For example, "Beijing City" has 3 characters and can be split into "Bei / Jing City", "Beijing / City", "Beijing City", "Bei / Jing / City" (4 = 2^3 - 1) ways.

[0175] For a string with a large number of characters, it is basically very difficult to traverse all possibilities. However, for characters with a certain order, it is often not necessary to traverse all possible splits, and only the locally optimal splitting situation needs to be considered. For example, for the text "Wuwei City, Gansu Province, Aksu Region, Xinjiang Uygur Autonomous Region, Ningxia Hui Autonomous Region, Dachuan City, Wuzhong City, Ningxia Hui Autonomous Region, Turpan City, Xinjiang Uygur Autonomous Region, Shihezi City, Xinjiang Uygur Autonomous Region, Jiuquan City, Gansu Province, Hami City, Xinjiang Uygur Autonomous Region, Changji Hui Autonomous Prefecture, Xinjiang Uygur Autonomous Region", in the sub-text "Wuwei City, Gansu Province, Aksu Region, Xinjiang Uygur Autonomous Region,", the locally optimal division results are "Wuwei City, Gansu Province" and "Aksu Region, Xinjiang Uygur Autonomous Region", and this local string is obtained by the sliding of the sliding window. In this scheme, the size of the sliding window is the length of the longest string in the database plus 1 (i.e., the preset value is set to 1), because the text is only split once within the window, so the window size should be greater than the longest string in the database.

[0176] In this scheme, the length of the longest string in the database is 19, so the window size is 20 characters long.

[0177] For the above text, after the window is initialized, the first sub-text obtained is "Wuwei City, Gansu Province, Aksu Region, Xinjiang Uygur Autonomous Region, Ning". In the window, the text is only divided into 2 strings, so the division results include "Gan / Susu Province, Wuwei City, Aksu Region, Xinjiang Uygur Autonomous Region, Ning", "Gansu / Susu Province, Wuwei City, Aksu Region, Xinjiang Uygur Autonomous Region, Ning", "Gansu Province / Wuwei City, Aksu Region, Xinjiang Uygur Autonomous Region, Ning", etc. Calculate the minimum distance between the two divided strings and the strings in the database. This distance uses the Damerau - Levenshtein edit distance.

[0178] The Damerau-Levenshtein edit distance is a string metric used to measure the edit distance between two character sequences. The Damerau–Levenshtein Distance between two words is the minimum number of operations required to transform one word into the other. Different from the Levenshtein Distance, in addition to the insertion, deletion, and substitution of single characters, it also includes the transposition of two adjacent characters.

[0179] For example, in the segmentation method of "Wuwei City, Gansu Province / Aksu Region, Xinjiang Uygur Autonomous Region, Ning", the word with the minimum edit distance in the database that can match "Wuwei City, Gansu Province" is "Wuwei City, Gansu Province". Since there is an identical string in the database, the edit distance is 0. Therefore, the string "Wuwei City, Gansu Province" in the database is used as the recognition result of the待识别字符串 "Wuwei City, Gansu Province".

[0180] The word with the minimum edit distance in the database that can match "Aksu Region, Xinjiang Uygur Autonomous Region, Ning" is "Aksu Region, Xinjiang Uygur Autonomous Region", and the edit distance is 3, that is, adding a character "维", deleting the character ",", and deleting the character "宁", a total of 3 operations. Therefore, the string "Aksu Region, Xinjiang Uygur Autonomous Region" in the database is used as the recognition result of the待识别字符串 "Aksu Region, Xinjiang Uygur Autonomous Region, Ning", and at this time, the missing character "维" can be automatically added.

[0181] The edit distance of this segmentation method is the sum of the edit distances of these two substrings, which is 3 and the smallest among all segmentation methods, so it is the best segmentation method.

[0182] Therefore, the recognized text of the first sub-text "Wuwei City, Gansu Province Aksu Region, Xinjiang Uygur Autonomous Region, Ning" is "Wuwei City, Gansu Province Aksu Region, Xinjiang Uygur Autonomous Region".

[0183] Next, the window moves to after "Wuwei City, Gansu Province", that is, the window content becomes "Aksu Region, Xinjiang Uygur Autonomous Region, Ningxia Hui Autonomous".

[0184] It should be noted that the original text has some unclear expressions like "待识别字符串", which are directly translated as they are in the above translation. You may need to check and clarify the context for a more accurate translation. Also, the tags etc. are left unchanged as required.According to the previous calculation method, the best segmentation is "Aksu Prefecture, Xinjiang Uygur Autonomous Region / , Ningxia Hui Autonomous". The edit distance from this to "Aksu Prefecture, Xinjiang Uygur Autonomous Region / , Yinchuan City, Ningxia Hui Autonomous Region" in the database is 7 (1 + 6). That is, the edit distance between the string to be recognized "Aksu Prefecture, Xinjiang Uygur Autonomous Region" and the string in the database "Aksu Prefecture, Xinjiang Uygur Autonomous Region" is 1 (the character 'Wei' needs to be added). The edit distance between the substring to be recognized ", Ningxia Hui Autonomous" and the string in the database "Yinchuan City, Ningxia Hui Autonomous Region" is 6, that is, deleting the character ',', changing the character 'Bai' to 'Zi', and adding the characters 'Qu Yinchuan City' in total requires 6 operations.

[0185] Based on the above analysis, "Aksu Prefecture, Xinjiang Uygur Autonomous Region" can be matched to "Aksu Prefecture, Xinjiang Uygur Autonomous Region", and the missing character 'Wei' can also be corrected through matching. By the same method, the subsequent "Ningxia Hui Autonomous Big City" will also be matched to "Yinchuan City, Ningxia Hui Autonomous Region" and directly corrected to "Yinchuan City, Ningxia Hui Autonomous Region", and the misspelled characters can also be corrected.

[0186] Therefore, by calculating in the above way successively, the segmentation result can finally be obtained, and the misspelled and missing characters in the middle will also be corrected.

[0187] In addition, if the text contains words other than region names, such as "You arrived at or passed through: Beijing within the previous 14 days", when segmenting, the maximum distance can be set (the preset value, set to 4 in this scheme). If it exceeds the maximum distance, the matching fails. The best segmentation of the above text is "You arrived at or passed through: / Beijing", but the edit distance between "You arrived at or passed through: " and the strings in the database is greater than 4, so it is marked as "matching failed", and "Beijing" is marked as "matching successful".

[0188] Color extraction module:

[0189] Color feature description:

[0190] In the green travel card, there is a green arrow in the middle, the surrounding of the white information area is green, and there is fixed text "Please accept the green travel card" above the mobile phone number. In the yellow travel card, there is a yellow shield graphic in the middle, the surrounding of the white information area is yellow, and there is no fixed text reflecting the color above the mobile phone number. In the red travel card, there is a red heart graphic in the middle, the surrounding of the white information area is red, and there is no fixed text reflecting the color above the mobile phone number.

[0191] Due to interference such as light and eye protection mode, and because yellow and red are treated the same in actual business, only green and non - green are distinguished.

[0192] Problems in extraction:

[0193] • Due to shooting reasons, the fixed text "Please accept the green travel card" may not be captured.

[0194] • The fixed text "Please accept the green travel card" may be misrecognized.

[0195] • Due to the influence of the eye protection mode and shooting light, the green travel card will be yellowish, making color judgment relatively difficult.

[0196] Implementation method:

[0197] If the matching degree of the text "Please accept the green travel card" in the sub - text list is above 85% (i.e., the preset value is 85%) and the position of the text is appropriate (above the mobile phone number), it is determined to be green. If not, continue to judge.

[0198] If a single character "个" or "T" can be recognized between the update date and the places passed by, it is considered a green code. Otherwise, continue to judge.

[0199] Use image - processing methods to extract the arrow, as follows:

[0200] Read the picture as Figure 8 shown.

[0201] To reduce the interference of other colors on edge detection, convert the picture to the hsv color space and set a mask to convert non - green colors to black, as Figure 9 shown.

[0202] To reduce noise interference, binarize the picture and perform Gaussian blur at the same time, as Figure 10 shown.

[0203] Then perform edge detection, as Figure 11 shown.

[0204] To reduce small cracks and increase edge connectivity, perform erosion and dilation, as Figure 12 shown.

[0205] Finally, remove the circular contour, find the contour with a perimeter within a certain range (the number of pixels is from 300 to 2000), call the Opencv function to obtain the approximate polygon of the contour and the convex hull of the approximate polygon, and if the number of side lengths of the convex hull is between 3 and 6 and there are two extra sides in the approximate polygon inside the convex hull, then judge that the circumscribed polygon is the arrow "↑".

[0206] The method of this embodiment can solve the problems of misrecognition of text by the general OCR model and inability to extract structured information, and improve the success rate and accuracy of travel code information extraction.

[0207] Example 2.

[0208] This embodiment provides an AI-powered big data travel card automatic recognition system, which includes the following modules:

[0209] Image recognition module: Used to recognize images of travel data cards using image detection and recognition models, and to obtain at least a text list, wherein the text list includes at least text.

[0210] The path-search text list acquisition module is used to obtain a path-search text list based on the text list.

[0211] The route location text extraction module is used to extract route location text based on the route location search text list using an intelligent text segmentation engine. The extraction of route location text using the intelligent text segmentation engine includes the following steps:

[0212] The text in the searched text list is segmented according to segmentation rules to obtain several sub-texts.

[0213] Each subtext is segmented into several strings. Each segmented string is compared with strings in a pre-defined database to obtain a difference score. The sum of the difference scores of the several strings in each subtext is calculated to obtain a difference score result value. The database strings include region name strings.

[0214] Extract the strings corresponding to the smallest difference result values ​​from the database as the path text after the subtext is identified.

[0215] This embodiment consists of functional modules that correspond one-to-one with some of the method steps in Embodiment 1.

[0216] Example 3.

[0217] This embodiment provides a computer-readable storage medium for storing a computer program, which, when invoked, executes the AI ​​communication big data travel card automatic identification method in the first aspect or any possible implementation of the first aspect.

[0218] By storing the computer program that implements the above method in a computer-readable storage medium, the computer program can be easily obtained, thereby implementing the method of Embodiment 1 to improve the success rate and accuracy of itinerary code information extraction.

[0219] Example 4.

[0220] This embodiment provides a computer device, including a memory and a processor. The memory is used to store a computer program, which, when called by the processor, executes the AI ​​communication big data travel card automatic identification method in the first aspect or any possible implementation of the first aspect.

[0221] By storing the computer program that implements the above method in the memory of a computer device, the computer program can be easily executed by the processor, thereby implementing the method of Embodiment 1 to improve the success rate and accuracy of itinerary code information extraction.

[0222] It is understood that the memory in the embodiments of this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0223] Non-volatile memory can be ROM, programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0224] Volatile memory can be RAM, which is used as an external cache. There are many different types of RAM, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory.

[0225] The processor mentioned above can be a CPU, microprocessor, ASIC, or one or more integrated circuits that execute a program to control the method of transmitting feedback information described above. The processing unit and the storage unit can be decoupled and disposed on different physical devices, connected via wired or wireless means to implement their respective functions, thereby supporting the system chip in implementing the various functions described in the above embodiments. Alternatively, the processing unit and the memory can also be coupled to the same device.

[0226] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a computer-readable storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned computer-readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0227] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An AI-powered method for automatically recognizing travel itinerary cards based on big data communication, characterized by: It includes the following steps: Use an image detection and text recognition model to recognize the image of the travel data itinerary card, and at least obtain a text list. The text list at least includes text. Obtain a search text list of places passed through based on the text list. Use an intelligent text segmentation engine to extract the text of places passed through based on the search text list of places passed through. The step of using the intelligent text segmentation engine to extract the text of places passed through includes the following steps: Segment the text in the search text list of places passed through according to segmentation rules to obtain several sub-texts. Split each sub-text to obtain several strings of characters. Compare the difference degree between each string of characters after splitting and the strings in a preset database to obtain the difference degree. Calculate the sum of the difference degrees of several strings of characters for each sub-text to obtain a difference degree result value. Among them, the strings in the database include region name strings. Extract the several strings of characters corresponding to the smallest difference degree result value in the database as the text of the places passed through after recognition of this sub-text. The text list also includes the position coordinates of the four vertices of the text box where the text is located. Extract the update date based on the text list. The extraction method of the update date includes the following steps: Obtain a search text list of the update date based on the text list. Traverse the search text list of the update date. When the number of consecutive longest digits in the text is greater than 4 and the total number of digits is 13 or 14, if the edit distance between the non-digit part of the text and the text "Updated on:" is less than 2, then directly determine that this text is the update date information. Otherwise, add the text to the candidate list as a candidate. Traverse the candidate list. If the total number of digits is 13, add 0 at the end and then convert it to a time format. Compare the converted time with the current time. If the time difference is within 24 hours, it is considered that the correct time has been extracted. Otherwise, it means that the update time acquisition fails and no subsequent information extraction is performed. Recognize the color of the itinerary code based on the text list and the image, including the following steps: First, determine whether the text list contains the text "Please accept the green itinerary card". If the matching degree of the text is above the preset value and the position coordinates of the text are above the position coordinates of the mobile phone number, then determine that the color of the itinerary code is green. Otherwise, continue to judge. Based on the position coordinates of the text, if a single text "one" or "T" is recognized between the update date and the places passed through, then it is considered a green code. Otherwise, continue to judge. Recognize the image. If there is an arrow "↑", then determine that the color of the itinerary code is green. Otherwise, it is non-green. The image recognition process includes the following steps: First, the image information is converted to the HSV color space and a mask is set to convert non-green colors to black. Then, binarization and Gaussian blur processing are performed. Next, edge detection is performed, followed by erosion and dilation. Finally, the circular outline is removed, and an outline with a perimeter within a preset range is found. The OpenCV function is called to obtain the approximate polygon of the outline and the convex hull of the approximate polygon. If the number of side lengths in the convex hull is between 3 and 6 and the approximate polygon inside the convex hull has two extra sides, then the circumscribed polygon is determined to be the arrow "↑".

2. The AI ​​communication big data travel card automatic identification method according to claim 1, characterized in that: The segmentation rules include: using a sliding window to select and segment the text to be extracted to obtain several sub-texts, wherein the length of the sliding window is equal to the longest string in the preset database plus N characters, and N is an integer not less than 1.

3. The AI ​​communication big data travel card automatic identification method according to claim 1, characterized in that: The step of splitting each subtext into several strings includes: splitting each subtext once to obtain two strings.

4. The AI ​​communication big data travel card automatic identification method according to claim 1, characterized in that: The difference between each segmented string and the strings in the preset database is compared to obtain the difference score. This includes calculating the edit distance between each string and the strings in the preset database.

5. The AI ​​communication big data travel card automatic identification method according to claim 1, characterized in that: If the minimum value of the difference result is greater than the preset value, then the database corresponding to the difference result value will not be extracted as the path text after the subtext is identified.

6. The AI ​​communication big data travel card automatic identification method according to claim 1, characterized in that: The method for obtaining the route search text list includes the following steps: Image processing is performed on the image of the travel data itinerary card to obtain a "white" information area. This information area contains text information such as mobile phone number, update date, and route locations. The text information is then identified to form a sub-text list. This sub-text list includes the text information of mobile phone number, update date, and route locations, as well as the position coordinates of the text. Based on the position coordinates of the text, the text below the update date in the sub-text list is obtained as the route search text list.

7. The AI ​​communication big data travel card automatic identification method according to claim 1, characterized in that: After identifying the images of the travel data itinerary card using the image detection and recognition model and obtaining at least a text list, the following steps are performed: If the number of identified text entries is less than 3, the text list is rotated 180°, and / or, a text box is set for each line of text in the text list, and the angle A between the long side of the text box and the horizontal direction of the image is calculated. If the angle A is greater than 45°, it is not included in subsequent text extraction, and / or, a text box is set for each line of text in the text list, and the text in the text list is reordered in order from top to bottom and from left to right using the coordinates of the upper left corner of the text box of each text entry.

8. The AI ​​communication big data travel card automatic identification method according to claim 1, characterized in that: The text list also includes the position coordinates of the four vertices of the text box where the text is located. Based on the text list, a mobile phone number is extracted. The method for extracting the mobile phone number includes the following steps: Traverse the text list, take the Chinese text in the text list, calculate the edit distance between the Chinese text and the text "Dynamic Travel Card", if the edit distance is less than 2, and the position coordinates of the text are in the upper half of the image, then determine the line of text as a candidate text. Take the numeric text in the text list. If the number of consecutive longest digits is greater than 4, and the total number of digits in the text is less than 12, then it is also listed as a candidate text. Traverse the candidate texts, obtain all the digits in the candidate texts, and form a numeric text. If the number of digits in the numeric text is greater than 6 and less than 12, and the first digit of the digits is 1, then take the first 3 and the last 4 digits of the numeric text, separated by "****", as the final mobile phone number; if the number of digits in the numeric text is greater than 3 and less than 7, then use the position coordinates of the text box for judgment. If there are 3 digits on the left side of the text box and the first digit is 1, then the 3 digits and the last 4 digits on the right side plus the "****" in the middle form the mobile phone number, otherwise it means that the extraction of the mobile phone number fails and no subsequent information extraction is performed.

9. An AI-powered communication big data-based automatic identification system for travel cards, characterized in that: It includes the following modules: Image recognition module: used to recognize the image of the travel big data itinerary card by using an image detection and recognition model, and at least obtain a text list. The text list at least includes text. If the number of recognized text lines is less than 3, then rotate the text list 180°. The text list also includes the position coordinates of the four vertices of the text box where the text is located. Based on the text list, the update date is extracted. The method for extracting the update date includes the following steps: Obtain the search text list of the update date based on the text list. Traverse the search text list of the update date. When the number of consecutive longest digits in the text is greater than 4, and the total number of digits is 13 or 14, if the edit distance between the non-numeric part of the text and the text "Updated on:" is less than 2, then directly determine that the text is the update date information, otherwise add the text as a candidate to the candidate list. Traverse the candidate list. If the total number of digits is 13, then add 0 at the end, and then convert it to a time format. Compare the converted time with the current time. If the time difference is within 24 hours, then it is considered that the correct time is extracted, otherwise it means that the acquisition of the update time fails and no subsequent information extraction is performed. Based on the text list and the color of the itinerary code recognized from the image, it includes the following steps: First, determine whether the text "Please accept the green itinerary card" is contained in the text list. If the matching degree of the text is above the preset value and the position coordinates of the text are above the position coordinates of the mobile phone number, then determine that the color of the itinerary code is green; otherwise, continue to judge. Based on the position coordinates of the text, if a single text "个" or "T" is recognized between the update date and the places passed by, then it is considered a green code; otherwise, continue to judge. Identify the image. If there is an arrow "↑", then determine that the color of the itinerary code is green; otherwise, it is non-green. The step of identifying the image includes the following steps: First, convert the picture information into the HSV color space and set a mask to convert non-green into black. Then, perform binaryzation and Gaussian blur processing. Further, perform edge detection. Further, perform erosion and dilation. Finally, remove the circular contour, find the contour with a perimeter within the preset range, call the Opencv function to obtain the approximate polygon of the contour and the convex hull of the approximate polygon. If the number of side lengths of the convex hull is between 3 and 6 and there are two extra sides in the approximate polygon inside the convex hull, then judge that the circumscribed polygon is the arrow "↑". Places Passed By Search Text List Acquisition Module: Used to obtain the places passed by search text list based on the text list. Places Passed By Text Extraction Module: Used to extract the places passed by text based on the places passed by search text list by using an intelligent text segmentation engine. The step of extracting the places passed by text by using the intelligent text segmentation engine includes the following steps: Segment the text in the places passed by search text list according to the segmentation rules to obtain several sub-texts. Divide each sub-text to obtain several strings of characters. Compare the difference degree of each string of characters after segmentation with the strings in the preset database to obtain the difference degree. Calculate the sum of the difference degrees of several strings of characters of each sub-text to obtain the difference degree result value. Among them, the strings in the database include region name strings. Extract the several strings of characters corresponding to the smallest difference degree result value in the database as the recognized places passed by text of this sub-text.

10. A computer-readable storage medium for storing a computer program, characterized in that: When the computer program is called, it executes the AI communication big data itinerary card automatic recognition method according to any one of claims 1 to 8.

11. A computer device comprising a memory and a processor, the memory being used to store a computer program, characterized in that: When the computer program is called by the processor, it executes the AI communication big data itinerary card automatic recognition method according to any one of claims 1 to 8.

Images