Automatic auditing method and visualization system for attendance photos

Abstract

The invention discloses a method and a visualization system for automatically auditing attendance photos, and the method comprises the steps: obtaining attendance photos; judging whether the file format of the clock-in photo conforms to a preset file format or not; a PP-OCRv4 model is used for recognizing watermark text content in the intercepted punch-in photo, and a watermark text recognition result is output; carrying out noise suppression on the clock-in photo by adopting an edge preserving filtering algorithm, and dynamically adjusting the photo brightness based on a dynamic brightness compensation mechanism combined with a Retinex theory and adaptive histogram equalization so as to complete preprocessing of the clock-in photo; using a ViT model to identify and classify the background of the preprocessed clock-in photo, and outputting a background identification result; the validity of the clock-in photo is comprehensively judged according to the watermark text recognition result and the background recognition result; analyzing by adopting a cross table, and outputting an attendance result. The scientific and efficient offshore wind power operation and maintenance management can be effectively realized, the management efficiency is remarkably improved, and the cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular to an automated review method and visualization system for sea check-in photos. Background Art

[0002] With the rapid development of offshore wind power, the demand for operation and maintenance of offshore wind power facilities continues to grow. To ensure stable equipment operation and efficient maintenance, operation and maintenance personnel need to frequently go out to sea to ensure equipment reliability, promptly resolve faults, and optimize power generation performance, thereby improving the economic benefits and safety of wind farms.

[0003] Due to the special environment of offshore operations, offshore operation and maintenance personnel use the traditional attendance method of clocking in and out at sea to record attendance information. This method cannot accurately track the operation and maintenance personnel's offshore operations, and ground management personnel need to review the authenticity of the attendance records one by one later. Faced with a large amount of attendance data, a lot of manpower and time costs are required, and the accuracy of manual review is low. Management personnel cannot accurately grasp the attendance status of personnel in real time, and it is difficult to reasonably arrange personnel scheduling according to actual work needs, affecting the efficiency and quality of offshore wind power operation and maintenance work. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a method for automated review of sea clock-in photos. Through image text recognition and background recognition technology, the accuracy and compliance of clock-in information can be automatically verified, and attendance reports can be generated in combination with data analysis to achieve scientific and efficient offshore wind power operation and maintenance management, significantly improve management efficiency and reduce costs.

[0005] Another object of the present invention is to provide a system for automatically reviewing and visualizing photos of check-in at sea.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A method for automatically reviewing overseas check-in photos, comprising:

[0008] Obtain clock-in photos of offshore operation and maintenance personnel going out to sea;

[0009] Determine whether the file format of the check-in photo meets the preset file format. If so, proceed to the next step;

[0010] The watermark text area of ​​the clock-in photo is intercepted, and the PP-OCRv4 model is used to identify the watermark text content in the intercepted clock-in photo. The attendance information is extracted, and whether it meets the preset overseas clock-in specifications is determined. The watermark text recognition result is output;

[0011] An edge-preserving filtering algorithm is used to suppress noise in check-in photos. A dynamic brightness compensation mechanism combining Retinex theory and adaptive histogram equalization is used to dynamically adjust photo brightness, completing the preprocessing of check-in photos.

[0012] Use the ViT model to identify and classify the background of the pre-processed check-in photo, determine whether the background is a unit or a crane ship, and output the background recognition result;

[0013] The validity of the check-in photo is comprehensively judged based on the watermark text recognition results and background recognition results;

[0014] Use cross-tab analysis to output attendance results.

[0015] Furthermore, the preset file formats include jpg, jpeg, png, gif, and bmp.

[0016] Furthermore, the area of ​​the watermark text content of the check-in photo is intercepted, specifically,

[0017] The image.size attribute in the Python image processing library PIL is used to obtain the width and height of the check-in photo, and then the crop method is used to capture the preset area in the lower left corner of the check-in photo. This area is the fixed position area of ​​the watermark text content of the check-in photo and is used as the object of text recognition.

[0018] Furthermore, the PP-OCRv4 model is used to identify the watermark text content in the intercepted punch-in photo, extract the attendance information, determine whether it meets the preset overseas punch-in specifications, and output the watermark text recognition results, specifically,

[0019] The configuration parameters of the PP-OCRv4 model are as follows:

[0020] ocr=PaddleOCR(use_angle_cls=True, lang="ch", use_gpu=True, det_db_thresh=0.2, det_db_box_thresh=0.6, det_db_unclip_ratio=1.6,

[0021] use_dilation=False,drop_score=0.6)

[0022] Among them, use_angle_cls is set to True, which means starting the direction classifier to process tilted text; lang is set to "ch" to optimize the recognition of Chinese content; use_gpu is set to True, which means enabling GPU acceleration to improve processing speed; det_db_thresh is the threshold for optimizing watermark text detection; det_db_box_thresh is the threshold for optimizing watermark text box detection; use_dilatio is set to False, which means prohibiting the expansion operation to avoid watermark text adhesion to improve recognition accuracy; drop_score is used to filter low-confidence recognition results to improve overall recognition accuracy;

[0023] The PP-OCRv4 model is used to identify the watermark text content in the intercepted punch-in photo, and the attendance information including the check-in time, the name of the check-in person, the check-in person's employee number, the longitude and latitude, and the project name are extracted. If all the information in the attendance information meets the preset overseas punch-in specifications, it is determined that the watermark text content of the punch-in photo meets the requirements, and the watermark text recognition result is qualified; otherwise, it is unqualified.

[0024] Furthermore, an edge-preserving filtering algorithm is used to suppress noise in the check-in photos, and a dynamic brightness compensation mechanism based on the Retinex theory combined with adaptive histogram equalization is used to dynamically adjust the photo brightness to complete the preprocessing of the check-in photos. Specifically,

[0025] Adjust the image size of the check-in photo to 128 pixels in height and width;

[0026] A joint weight function is constructed based on the spatial proximity weight and pixel similarity weight to accurately suppress image noise and enhance the texture features of the unit and the crane. The calculation formula of the joint weight function is as follows:

[0027]

[0028] Where g(i,j) is the output pixel value of the bilateral filter; f(k,l) is the pixel value of the input image at coordinate (k,l); Ω is the filter window centered at the current pixel (i,j); c(i,j,k,l) ​​is the spatial proximity weight, which is calculated as follows:

[0029]

[0030] Among them, σ d is the standard deviation of spatial distance, which controls the decay rate of spatial proximity; s(i,j,k,l) ​​is the pixel similarity weight, which is calculated as follows:

[0031]

[0032] Among them, σr is the standard deviation of pixel values, which controls the decay rate of pixel similarity;

[0033] Then, the dynamic brightness compensation mechanism is used to dynamically adjust the brightness according to the lighting conditions of the image, and the image I(x, y) is decomposed into the reflection component R(x, y) and the illumination component L(x, y) using Retinex decomposition, that is:

[0034] I(x,y)=R(x,y)·L(x,y)

[0035] Combined with adaptive histogram equalization parameters, the illumination component L(x,y) is dynamically adjusted to achieve intelligent optimization of image brightness under different lighting conditions.

[0036] Furthermore, the ViT model is used to identify and classify the background of the pre-processed punch-in photo, determine whether the background is a unit or a crane ship, and output the background recognition result, specifically,

[0037] Divide the image into 8×8 blocks, each of which is 16×16 pixels in size;

[0038] Image feature extraction is performed based on each small block using depth-wise separable convolution;

[0039] The small block of data is combined with the positional encoding and input into the Transformer encoder for feature extraction and modeling. The number of Transformer encoders is set to 7 in a stacked arrangement, and the number of heads in the multi-head attention mechanism structure is set to 12.

[0040] The ViT model decodes the extracted features through the classification head to identify and classify the photo background. The classification head uses an MLP structure with 4 hidden layers and a dropout parameter of 0.5.

[0041] If the background is a unit or a lifting vessel, the background recognition result is output as qualified. If the background is unclear, the background recognition result is output as manual judgment. In other cases, the background recognition result is output as unqualified.

[0042] Furthermore, the validity of the check-in photo is comprehensively judged based on the watermark text recognition results and background recognition results. Specifically,

[0043] If the watermark text recognition result and the background recognition result include unqualified, the check-in photo is unqualified. If the watermark text recognition result and the background recognition result include manual judgment, the check-in photo needs to be manually judged. If the watermark text recognition result and the background recognition result are both qualified, the check-in photo is qualified.

[0044] Furthermore, we use cross-tabulation analysis to output attendance results, specifically,

[0045] Use the pivot method in Python to create a crosstab, with the name of the sign-in person, the employee number of the sign-in person, and the project name as the row index of the DataFrame, and the sign-in time as the column name. Fill all the judgment results into the corresponding positions, and finally save the DataFrame as a table file and output it to the folder path specified by the user.

[0046] Another object of the present invention is achieved through the following technical solutions:

[0047] A system for automatically reviewing and visualizing sea check-in photos, used to implement the above-mentioned method for automatically reviewing sea check-in photos, comprising:

[0048] The acquisition module is used to obtain the check-in photos of offshore operation and maintenance personnel going out to sea;

[0049] A format determination module, used to determine whether the file format of the punch-in photo conforms to a preset file format;

[0050] The watermark text recognition module is used to intercept the watermark text content area of ​​the clock-in photo, use the PP-OCRv4 model to identify the watermark text content in the intercepted clock-in photo, extract the attendance information, determine whether it meets the preset overseas clock-in specifications, and output the watermark text recognition results;

[0051] The image preprocessing module is used to suppress noise in check-in photos using an edge-preserving filtering algorithm and dynamically adjust the photo brightness based on a dynamic brightness compensation mechanism that combines Retinex theory with adaptive histogram equalization;

[0052] The background recognition module is used to identify and classify the background of the pre-processed punch-in photo using the ViT model, determine whether the background is a unit or a lifting vessel, and output the background recognition result;

[0053] A comprehensive judgment module is used to comprehensively judge the validity of the check-in photo based on the watermark text recognition results and background recognition results;

[0054] The visual output module is used to output attendance results using cross-tabulation analysis.

[0055] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0056] 1. The present invention uses image text recognition and image background recognition technologies to automatically review the validity of sea check-in photos and output analysis results, realizing automated and intelligent attendance management, significantly reducing manual review operations, improving review efficiency and attendance management efficiency, improving review accuracy, and reducing the error rate of manual review. It overcomes the limitations of traditional attendance methods in special maritime environments and realizes accurate tracking of operation and maintenance personnel's sea operations.

[0057] 2. The present invention can effectively improve the standardization of attendance and the scientific nature of operation and maintenance management, providing a scientific and efficient solution for offshore wind power operation and maintenance. The method is highly compatible and suitable for wind power operation and maintenance image recognition in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Flow chart of the method of the present invention.

[0059] Figure 2 This is a schematic diagram of the output of watermark text recognition results.

[0060] Figure 3 This is a diagram of the interface of the web-based overseas check-in tool. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0062] Example 1:

[0063] like Figure 1 As shown, this embodiment provides a method for automatically reviewing photos of overseas check-in, including the steps of:

[0064] S1. Obtain check-in photos of offshore operation and maintenance personnel, keeping the background as clear as possible;

[0065] S2. Determine whether the file format of the check-in photo meets the preset file format. If so, proceed to the next step. If not, skip and do not process. The preset file formats include jpg, jpeg, png, gif, and bmp.

[0066] S3, intercept the area of ​​the watermark text content in the punch-in photo, use the PP-OCRv4 model to recognize the watermark text content in the intercepted punch-in photo, extract the attendance information, determine whether it meets the preset overseas punch-in specifications, and output the watermark text recognition result; specifically,

[0067] The image.size attribute in the Python image processing library PIL is used to obtain the width and height of the check-in photo. Then, the crop method is used to crop the preset area in the lower left corner of the check-in photo. This area is the fixed position area of ​​the watermark text content in the check-in photo and is used as the object of text recognition.

[0068] The configuration parameters of the PP-OCRv4 model are as follows:

[0069] ocr=PaddleOCR(use_angle_cls=True, lang="ch", use_gpu=True, det_db_thresh=0.2, det_db_box_thresh=0.6, det_db_unclip_ratio=1.6,

[0070] use_dilation=False,drop_score=0.6)

[0071] Among them, use_angle_cls is set to True, which means starting the direction classifier to process tilted text; lang is set to "ch" to optimize the recognition of Chinese content; use_gpu is set to True, which means enabling GPU acceleration to improve processing speed; det_db_thresh is the threshold for optimizing watermark text detection; det_db_box_thresh is the threshold for optimizing watermark text box detection; use_dilatio is set to False, which means prohibiting the expansion operation to avoid watermark text adhesion to improve recognition accuracy; drop_score is used to filter low-confidence recognition results to improve overall recognition accuracy;

[0072] The PP-OCRv4 model is used to identify the watermark text content in the intercepted punch-in photo, and the attendance information including the check-in time, the name of the check-in person, the employee number of the check-in person, the longitude and latitude, and the project name are extracted. If all the information in the attendance information meets the preset overseas punch-in specifications, such as the check-in time is between 7:00 and 23:00, the name of the check-in person and the employee number of the check-in person are filled in correctly, and the distance between the longitude and latitude and the longitude of the punch-in project is no more than 100KM, it is determined that the watermark text content of the punch-in photo meets the requirements, and the watermark text recognition result is qualified. Otherwise, it is unqualified. The recognition result is as shown in the attached figure. Figure 2 As shown;

[0073] S4 is used to suppress noise in the check-in photo using an edge-preserving filtering algorithm, and dynamically adjust the photo brightness based on a dynamic brightness compensation mechanism combining Retinex theory and adaptive histogram equalization to complete the preprocessing of the check-in photo. Specifically,

[0074] Adjust the image size of the check-in photo to 128 pixels in height and width;

[0075] A joint weight function is constructed based on the spatial proximity weight and pixel similarity weight to accurately suppress image noise, enhance the texture features of the unit and the crane, provide high-quality image input for the ViT model, and improve the accuracy of background recognition. The calculation formula of the joint weight function is as follows:

[0076]

[0077] Where g(i,j) is the output pixel value of the bilateral filter; f(k,l) is the pixel value of the input image at coordinate (k,l); Ω is the filter window centered at the current pixel (i,j); c(i,j,k,l) ​​is the spatial proximity weight, which is calculated as follows:

[0078]

[0079] Among them, σ d is the standard deviation of spatial distance, which controls the decay rate of spatial proximity; s(i,j,k,l) ​​is the pixel similarity weight, which is calculated as follows:

[0080]

[0081] Among them, σ r is the standard deviation of pixel values, which controls the decay rate of pixel similarity;

[0082] Then, the dynamic brightness compensation mechanism is used to dynamically adjust the brightness according to the lighting conditions of the image, and the image I(x, y) is decomposed into the reflection component R(x, y) and the illumination component L(x, y) using Retinex decomposition, that is:

[0083] I(x,y)=R(x,y)·L(x,y)

[0084] Adaptive histogram equalization parameters are combined to dynamically adjust the illumination component L(x,y), completing intelligent optimization of image brightness under different lighting conditions. This effectively solves the problem of blurred and difficult-to-recognize image features caused by complex and changing lighting conditions at sea, and enhances the robustness of the model in complex lighting environments.

[0085] S5. Use the ViT model to identify and classify the background of the pre-processed punch-in photo, determine whether the background is a unit or a lifting ship, and output the background recognition result; specifically,

[0086] Divide the image into 8×8 blocks, each of which is 16×16 pixels in size;

[0087] Image feature extraction is performed based on each small block using depth-wise separable convolution;

[0088] The small block of data is combined with the positional encoding and input into the Transformer encoder for feature extraction and modeling. The number of Transformer encoders is set to 7 in a stacked arrangement, and the number of heads in the multi-head attention mechanism structure is set to 12.

[0089] The ViT model decodes the extracted features through the classification head to identify and classify the photo background. The classification head uses an MLP structure with 4 hidden layers and a dropout parameter of 0.5.

[0090] If the background is a unit or a lifting vessel, the background recognition result is output as qualified. If the background is unclear, the background recognition result is output as manual judgment. In other cases, the background recognition result is output as unqualified.

[0091] S6. Comprehensively judge the validity of the check-in photo according to the watermark text recognition results and background recognition results according to logical rules; specifically,

[0092] If the watermark text recognition result and background recognition result both contain "unqualified", the check-in photo is unqualified. If the watermark text recognition result and background recognition result both contain "manual judgment", the check-in photo needs to be manually judged. If the watermark text recognition result and background recognition result both contain "qualified", the check-in photo is qualified.

[0093] S7, using cross-tab analysis, output attendance results; specifically,

[0094] Use the pivot method in Python to create a crosstab, with the name of the sign-in person, the employee number of the sign-in person, and the project name as the row index of the DataFrame, and the sign-in time as the column name. Fill all the judgment results into the corresponding positions, and finally save the DataFrame as a table file and output it to the folder path specified by the user.

[0095] Example 2:

[0096] This embodiment provides a system for automatically reviewing and visualizing sea check-in photos, which is used to implement the method for automatically reviewing sea check-in photos described in Example 1. The system includes:

[0097] The acquisition module is used to obtain the check-in photos of offshore operation and maintenance personnel going out to sea;

[0098] A format determination module, used to determine whether the file format of the punch-in photo conforms to a preset file format;

[0099] The watermark text recognition module is used to intercept the watermark text content area of ​​the clock-in photo, use the PP-OCRv4 model to identify the watermark text content in the intercepted clock-in photo, extract the attendance information, determine whether it meets the preset overseas clock-in specifications, and output the watermark text recognition results;

[0100] The image preprocessing module uses an edge-preserving filtering algorithm to suppress noise in check-in photos and dynamically adjusts the photo brightness using a dynamic brightness compensation mechanism that combines Retinex theory with adaptive histogram equalization.

[0101] The background recognition module is used to identify and classify the background of the pre-processed punch-in photo using the ViT model, determine whether the background is a unit or a lifting vessel, and output the background recognition result;

[0102] A comprehensive judgment module is used to comprehensively judge the validity of the check-in photo based on the watermark text recognition results and background recognition results;

[0103] The visual output module is used to output attendance results using cross-tabulation analysis.

[0104] The system of the present invention can realize the development and deployment of Web-based overseas check-in tools based on the Streamlit architecture, such as Figure 3 As shown, it includes a visual interface, automatic review, one-click analysis, and information feedback. When using this tool, managers can launch the analysis model program with a single click and quickly obtain the final attendance results in their designated folder.

[0105] The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solution and invention concept of the present invention within the scope disclosed by the present invention, which falls within the scope of protection of the present invention.

Claims

1. A method for automatically reviewing photos of people checking in at sea, characterized in that: include, Obtain clock-in photos of offshore operation and maintenance personnel going out to sea; Determine whether the file format of the check-in photo meets the preset file format. If so, proceed to the next step; The watermark text area of ​​the clock-in photo is intercepted, and the PP-OCRv4 model is used to identify the watermark text content in the intercepted clock-in photo. The attendance information is extracted, and whether it meets the preset overseas clock-in specifications is determined. The watermark text recognition result is output; An edge-preserving filtering algorithm is used to suppress noise in check-in photos. A dynamic brightness compensation mechanism combining Retinex theory and adaptive histogram equalization is used to dynamically adjust photo brightness, completing the preprocessing of check-in photos. Use the ViT model to identify and classify the background of the pre-processed check-in photo, determine whether the background is a unit or a crane ship, and output the background recognition result; The validity of the check-in photo is comprehensively judged based on the watermark text recognition results and background recognition results; Use cross-tab analysis to output attendance results.

2. The method for automatically reviewing photos of passengers checking in at sea according to claim 1, characterized in that: The preset file formats include jpg, jpeg, png, gif, and bmp.

3. The method for automatically reviewing photos of passengers checking in at sea according to claim 1, characterized in that: The area where the watermark text content of the check-in photo is intercepted is: The image.size attribute in the Python image processing library PIL is used to obtain the width and height of the check-in photo, and then the crop method is used to capture the preset area in the lower left corner of the check-in photo. This area is the fixed position area of ​​the watermark text content of the check-in photo and is used as the object of text recognition.

4. The method for automatically reviewing photos of passengers checking in at sea according to claim 1, characterized in that: Use the PP-OCRv4 model to identify the watermark text content in the intercepted punch-in photo, extract the attendance information, determine whether it meets the preset overseas punch-in specifications, and output the watermark text recognition results, specifically, The configuration parameters of the PP-OCRv4 model are as follows: ocr=PaddleOCR(use_angle_cls=True,lang="ch",use_gpu=True, det_db_thresh=0.2, det_db_box_thresh=0.6, det_db_unclip_ratio=1.6, use_dilation=False,drop_score=0.6) Among them, use_angle_cls is set to True, which means starting the direction classifier to process tilted text; lang is set to "ch" to optimize the recognition of Chinese content; use_gpu is set to True, which means enabling GPU acceleration to improve processing speed; det_db_thresh is the threshold for optimizing watermark text detection; det_db_box_thresh is the threshold for optimizing watermark text box detection; use_dilatio is set to False, which means disabling the expansion operation to avoid watermark text adhesion and improve recognition accuracy; drop_score is used to filter low-confidence recognition results to improve overall recognition accuracy; The PP-OCRv4 model is used to identify the watermark text content in the intercepted punch-in photo, and the attendance information including the check-in time, the name of the check-in person, the check-in person's employee number, the longitude and latitude, and the project name are extracted. If all the information in the attendance information meets the preset overseas punch-in specifications, it is determined that the watermark text content of the punch-in photo meets the requirements, and the watermark text recognition result is qualified; otherwise, it is unqualified.

5. The method for automatically reviewing photos of people checking in at sea according to claim 1, characterized in that: The edge-preserving filtering algorithm is used to suppress noise in the check-in photos, and the dynamic brightness compensation mechanism based on the Retinex theory and adaptive histogram equalization is used to dynamically adjust the photo brightness to complete the preprocessing of the check-in photos. Specifically, Adjust the image size of the check-in photo to 128 pixels in height and width; A joint weight function is constructed based on the spatial proximity weight and pixel similarity weight to accurately suppress image noise and enhance the texture features of the unit and the crane. The calculation formula of the joint weight function is as follows: Where g(i,j) is the output pixel value of the bilateral filter; f(k,l) is the pixel value of the input image at coordinate (k,l); Ω is the filter window centered at the current pixel (i,j); c(i,j,k,l) ​​is the spatial proximity weight, which is calculated as follows: Among them, σ d is the standard deviation of spatial distance, which controls the decay rate of spatial proximity; s(i,j,k,l) ​​is the pixel similarity weight, which is calculated as follows: Among them, σ r is the standard deviation of pixel values, which controls the decay rate of pixel similarity; Then, the dynamic brightness compensation mechanism is used to dynamically adjust the brightness according to the lighting conditions of the image, and the image I(x, y) is decomposed into the reflection component R(x, y) and the illumination component L(x, y) using Retinex decomposition, that is: I(x,y)=R(x,y)·L(x,y) Combined with adaptive histogram equalization parameters, the illumination component L(x,y) is dynamically adjusted to achieve intelligent optimization of image brightness under different lighting conditions.

6. The method for automatically reviewing photos of passengers checking in at sea according to claim 1, characterized in that: The ViT model is used to identify and classify the background of the pre-processed punch-in photo, determine whether the background is a unit or a crane ship, and output the background recognition results, specifically, Divide the image into 8×8 blocks, each of which is 16×16 pixels in size; Image feature extraction is performed based on each small block using depth-wise separable convolution; The small block of data is combined with the positional encoding and input into the Transformer encoder for feature extraction and modeling. The number of Transformer encoders is set to 7 in a stacked arrangement, and the number of heads in the multi-head attention mechanism structure is set to 12. The ViT model decodes the extracted features through the classification head to identify and classify the photo background. The classification head uses an MLP structure with 4 hidden layers and a dropout parameter of 0.

5. If the background is a unit or a lifting vessel, the background recognition result is output as qualified. If the background is unclear, the background recognition result is output as manual judgment. In other cases, the background recognition result is output as unqualified.

7. The method for automatically reviewing photos of passengers checking in at sea according to claim 1, characterized in that: The validity of the check-in photo is judged comprehensively based on the watermark text recognition results and background recognition results. Specifically, If the watermark text recognition result and the background recognition result include unqualified, the check-in photo is unqualified. If the watermark text recognition result and the background recognition result include manual judgment, the check-in photo needs to be manually judged. If the watermark text recognition result and the background recognition result are both qualified, the check-in photo is qualified.

8. The method for automatically reviewing photos of people checking in at sea according to claim 1, characterized in that: Use cross-tabulation analysis to output attendance results, specifically, Use the pivot method in Python to create a crosstab, with the name of the sign-in person, the employee number of the sign-in person, and the project name as the row index of the DataFrame, and the sign-in time as the column name. Fill all the judgment results into the corresponding positions, and finally save the DataFrame as a table file and output it to the folder path specified by the user.

9. A system for automatically reviewing and visualizing photos of people checking in at sea, characterized by: The system is used to implement the method for automatically reviewing sea check-in photos as described in any one of claims 1 to 8, comprising: The acquisition module is used to obtain the check-in photos of offshore operation and maintenance personnel going out to sea; A format determination module, used to determine whether the file format of the punch-in photo conforms to a preset file format; The watermark text recognition module is used to intercept the watermark text content area of ​​the clock-in photo, use the PP-OCRv4 model to identify the watermark text content in the intercepted clock-in photo, extract the attendance information, determine whether it meets the preset overseas clock-in specifications, and output the watermark text recognition results; The image preprocessing module uses an edge-preserving filtering algorithm to suppress noise in check-in photos and dynamically adjusts the photo brightness using a dynamic brightness compensation mechanism that combines Retinex theory with adaptive histogram equalization. The background recognition module is used to identify and classify the background of the pre-processed punch-in photo using the ViT model, determine whether the background is a unit or a lifting vessel, and output the background recognition result; A comprehensive judgment module is used to comprehensively judge the validity of the check-in photo based on the watermark text recognition results and background recognition results; The visual output module is used to output attendance results using cross-tabulation analysis.

Images