A slab number recognition method based on deep learning

Abstract

The application discloses a kind of based on deep learning slab number identification method, including the following steps: obtaining the video recording of the field slab running process, a large number of images with slab number are obtained by frame interception to video recording;The slab number on each slab image is labeled, and the data set required for the slab number identification model training is made;Train and debug identification model;Field deployment identification model, each time the video of the entire track car passes through is identified, finally the matching result is processed, and the slab number and corresponding slab number are obtained;The result of identification is automatically queried in the warehousing plan of database, secondary verification is carried out, and identification is successful after success, and alarm prompt after failure.The slab number identification method of the application not only can reduce the amount of human work, but also can improve the accuracy of the slab number in the warehouse, facilitate the management of the warehouse area, and can realize automatic query and check, effectively speed up the slab warehousing rhythm, and improve the efficiency of the entire production workshop.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of metallurgical machinery and automation, artificial intelligence, in particular to a method for automatically identifying slab numbers based on deep learning. BACKGROUND

[0002] Hot rolling production is an important production line in a steel enterprise, and the continuous casting workshop and the hot rolling workshop belong to the upstream and downstream relationship. The hot continuous casting slab is sent to the hot rolling workshop by the roller after being cooled, flame treated, and having the slab number sprayed and printed on the surface and the internal quality of the continuous casting slab being qualified. Then, the slab number of the slab is manually identified by an operator, and the data of the slab in the slab yard is checked. After the slab number is confirmed to be correct, the subsequent operation instruction is issued to guide the slab to enter the designated position of the slab yard for stacking and waiting for the completion of the established rolling plan. With the continuous development of intelligent manufacturing technology, the steel enterprise also welcomes the wave of reform. At present, the automatic degree of the slab storage process in the hot rolling production line is low, especially the identification process of the slab number. Most steel enterprises rely on workers to observe the slab number in the running process for 24 hours without interruption, and then check the data in the slab yard every time. The labor intensity is not only large but also the work efficiency is low. At the same time, it is greatly affected by human factors. Long-time work is easy to cause misidentification or missed identification, which may cause the product to be judged as waste or cause serious economic losses to the entire production line. SUMMARY

[0003] The present application provides a method for automatically identifying slab numbers based on deep learning to realize automatic checking of the storage plan and improve production efficiency.

[0004] In order to achieve the above purpose, the technical scheme provided by the present application is as follows:

[0005] A slab number identification method based on deep learning, comprising the following steps:

[0006] Obtaining a video recording of the running process of the slab on site, and obtaining a large number of images with slab numbers by frame interception on the video recording;

[0007] Using machine vision recognition technology to label the slab number on each slab image, and making a data set required for training the slab number identification model;

[0008] Using a deep learning neural network to train and debug the identification model to achieve an accuracy of more than 99% on the collected data set;

[0009] Deploying the identification model on site, identifying a section of video of the entire running process each time, finally processing the matching result, obtaining the number of slabs and the corresponding slab numbers, and outputting the identification result;

[0010] The result of the identification is automatically queried in the database storage plan, secondary verification is performed, and the identification is successful after the success, and the alarm is prompted after the failure.

[0011] More specific steps are as follows:

[0012] A large number of images with slab numbers are obtained;

[0013] An RGB camera is set up at a position during the travel of the rail car, which is used to record the complete process of each rail car passing in front of the camera, and capture the slab number when the slab moves;

[0014] The video is downloaded from the field, and the video without rail car passing or the rail car without slab is deleted, and the remaining video is processed into an image through the command in ffmpeg;

[0015] The obtained slab image is processed, and the image without slab number is filtered out.

[0016] The data set required for training the slab number recognition model is made:

[0017] The image labeling tool labelme is used to label the obtained image with slab number;

[0018] The head of the slab is labeled, which is used to obtain the number of slabs;

[0019] During the labeling process, the images that are too blurred, misprinted, without slab number and slab head are deleted;

[0020] The json format obtained by labeling the image is converted into a txt file.

[0021] The recognition model is trained and debugged:

[0022] The training set and the verification set of the present application are divided;

[0023] The recognition model selects yolov5-L, and the deep learning recognition model is trained;

[0024] The trained model reaches an accuracy of more than 99% on the verification set.

[0025] The recognition model is deployed on site, and some images in a video in front of the camera passing through the entire rail car are recognized each time, and the matching result is finally processed to obtain the number of slabs and the corresponding slab number, and the specific steps are as follows:

[0026] According to the slab number obtained from each image, the images are first sorted from bottom to top, and then sorted from left to right. The character with the maximum y coordinate of the upper left corner pixel value in the image recognition result is selected as the last row, and then the center point coordinate y is obtained according to the upper left corner coordinate y1 and the lower right corner coordinate y2 of the character.center , reiterate all identified characters in the image of two y-axis coordinates, and determine whether the just obtained center point coordinate y center is within the current character two y-axis coordinates, set a different constant value deviation according to different field conditions, and then remove the line character from the image recognition result, loop the above iteration process to find the characters in each line of the image, iterate the x coordinate of each line of characters, and sort from left to right according to the x from small to large, complete the character recognition of an image. For the same slab head identification, the y coordinate of each line is sorted and selected to complete the complete identification algorithm of an image.

[0027] For a vehicle carrying a slab, many images will be processed one by one, and finally a summary will be made for a line of characters in each image. According to the above y center judgment algorithm, after matching the coordinates of all lines of characters, the line character with the most occurrences in the same line is considered as the correct recognition result, and the same applies to the slab head line coordinate matching. Finally, the line matching of the slab head and the slab string is performed to complete the complete recognition process of a vehicle.

[0028] Secondary verification of the recognition result:

[0029] Establish a database link to realize communication with the WMS database and the continuous casting workshop database;

[0030] According to the recognized slab number, search the slab original PDI data in the WMS database and the continuous casting database;

[0031] If the query is successful, it means that the secondary verification is successful, and this identification is correct;

[0032] If the query fails, an alarm will prompt manual intervention: workers will intervene to check the slab number image provided by the recognition client and confirm. If there are situations such as slab number separation, image blur, unclear handwritten slab, and non-standard slab number, contact the continuous casting workshop personnel for inquiry and correction to complete manual input.

[0033] In the above scheme, the client is connected to the camera above the track car running site, and stores the running situation photographed, and a database is established in the client for recognition result verification.

[0034] In the technical solution, the camera collects slab image videos of the slab in the running process at different positions, and the own data set is made. After the labeling processing, the data set is input into the deep learning neural network for training to obtain the recognition model. After the recognition program is deployed on site, the track car in front of the camera is recognized for 30 seconds through the trigger recognition signal. The single image recognition and single image post-processing are performed. Then, the multi-image recognition processing is performed on all image recognition processing results in 30 seconds to obtain the final slab number result. The result is written into the database. Finally, according to the recognized slab number, the WMS database and the continuous casting database are connected. Whether the recognition result exists in the database is inquired to judge the success and failure of the automatic recognition, and the secondary verification is completed. If the secondary verification fails, the client performs alarm prompting. The human eye confirms the slab image on the client interface, and the slab number is modified. For some abnormal and non-standard slab numbers, the workers in the continuous casting workshop are contacted for inquiry and modification.

[0035] Compared with the prior art, the present application has the following advantages:

[0036] The present application automatically recognizes the slab number in the running process through the neural network. Then, the final slab number and the corresponding slab number are obtained after a series of sorting algorithm processing, so as to replace the worker recognition process. The present application not only saves the warehouse slab number verification process, but also reduces the history of slab management confusion, the phenomenon of low productivity, high labor intensity and mixed steel, and the problems of seriously affecting the production rhythm and the determination of product quality in the later period. The present application realizes the full-automatic tracking and control of materials from continuous casting to furnace charging, and gets rid of the "bottleneck" affecting the production rhythm for many years.

[0037] The slab number recognition method of the present application not only can reduce the amount of human work, but also can improve the accuracy of the warehouse slab number, facilitate the management of the warehouse area, and realize automatic query and checking, effectively speed up the slab storage rhythm, and improve the efficiency of the entire production workshop. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a field device diagram suitable for the slab number recognition method based on deep learning of the present application;

[0039] Figure 2 is a flowchart of the slab number recognition method of the present application;

[0040] Figure 3 is a specific application flowchart of the slab number recognition method of the present application in the case;

[0041] Figure 4 is a multi-image algorithm processing schematic diagram of the slab recognition result in the embodiment of the present application;

[0042] Figure 5It is a schematic diagram of the slab number recognition client in the embodiment of the application. DETAILED DESCRIPTION

[0043] In order to more conveniently and in detail explain the problems and technical advantages to be solved by the application, the drawings and specific examples thereof will be described in detail below.

[0044] The embodiment of the application provides a deep learning-based automatic slab number recognition method, as shown in the figure. Figure 1 The figure shows the on-site equipment situation to which the slab number recognition method of the application is applicable, and the specific recognition method is as follows:

[0045] The camera is located on one side of the track, the three-dimensional scanning continuously tracks and positions the trolley, and the camera position and coordinates are fixed. When the trolley on the track is loaded with slabs and drives forward, the recognition signal is triggered when the trolley coordinates approach the camera coordinates, and the recognition mark in the database is changed from 0 to 1. The recognition program continuously accesses the recognition information in the database, and when the recognition signal becomes 1, the recognition is started. The camera image is captured and recognized in a loop, and the process is a single thread lasting for 30 seconds. The installation position of the camera on site can be modified, and the time is approximately equal to the time when a complete trolley passes the position of the camera. Finally, all single recognition results are processed to obtain the final result, and the database information is modified, the recognition time, the recognition signal, the number of slabs, the slab number and the slab image are set to 0, and the recognition program task is completed. Finally, the WMS database and the continuous casting database are verified again, and if the verification is successful, the recognition result is correct, and if the verification fails, the client is manually modified and manually input.

[0046] In the device configuration, the application adopts a client for manual setting and recognition verification. The client connects a camera above the track trolley running site, stores the running situation captured by the camera, and establishes a database in the client for verification of the recognition result.

[0047] As shown in the figure, Figure 2 A flowchart for automatically recognizing the slab number based on deep learning is shown.

[0048] The camera captures the image in the process of the track trolley running in the slab yard to obtain multiple effective slab number images, and can completely record the process of each trolley passing. The process is not only used in subsequent recognition, but also needs to collect a large number of images to make a data set during model training.

[0049] The deep learning neural network algorithm is used to recognize the slab image, and the single character symbol on the slab and the recognition of the slab head are obtained. At present, with the development of machine vision image algorithm, the neural network can effectively classify, detect, segment and the like by using image processing. The slab number and slab head are recognized by means of the yolov5 algorithm, and the more detailed process is as follows:

[0050] In order to train a neural network, a large amount of data iteration training is required for the neural network, so it is necessary to make a special data set, and then if manpower is used to collect images on site, the images shot are not restored in terms of light, angle, clarity and the like, and a lot of time and effort is consumed. Therefore, it is an optimal scheme to design and write an automatic image collection program. The application is arranged with a Dahua camera on site to capture a large number of slab images in the moving process of the rail car, and with the help of the recognition program idea, before the formal recognition on site, the host computer software (dispatching software WMS) gives a recognition signal to the database through the laser ranging value of the rail car reaching the recognition point. After the recognition program obtains the signal, the IP, username and password information of the camera are obtained, a frame of image is captured and saved to the local, and after 30 seconds, the database recognition signal is modified, and the image collection program enters the waiting state, waiting for the next recognition signal from the host computer software. About 15598 images are collected. The collected images are labeled, and the images with blurred, misprinted and no number are deleted, and finally 13844 training set images and 1754 test set images are obtained, and the data set is completed.

[0051] The model is trained by using the prepared data set, and it is found by observation that there are insufficient light and motion blur in the images on site, so some commonly used data augmentation methods are added when training the model. The model yolov5-L is selected, and the hyperparameters are set as follows: learning rate is 0.01, SGD is 0.937, warmup_epochs is 3.0, hsv_h is 0.015, hsv_s is 0.7, hsv_v is 0.4, translate is 0.1, scale is 0.9, left-right flip probability is 0.5, mosaic is 1.0. After many debugging experiments, it is found that mixup data enhancement has a negative effect on the result. It is analyzed that if mixup is used, more chaotic images will be generated, which are quite different from the real characters, thereby causing the result accuracy to decrease.

[0052] The model was trained using transfer learning, employing a pre-trained YOLOv5-L model on the COCO dataset. The training consisted of 300 epochs, a batch size of 2, an image size of 1280x1280, and a label-smoothing value of 0.1. A 3080 GPU was used for training. The final results on the validation set, with an NMS IoU threshold of 0.484 and a confidence threshold of 0.1, showed an accuracy of approximately 99%.

[0053] The GPU deployed on-site for the recognition program is a 1060. It obtains recognition signals from the database, then the camera captures images, and the recognition program performs the recognition. The results are then sent to the database. This process constitutes a complete recognition workflow. Under normal circumstances, the entire single-image recognition processing takes 30 seconds. According to on-site recognition records, it can process approximately 20 to 30 frames of images each time, capturing the slab number at various angles as the railcar passes in front of the camera.

[0054] For each image, the recognition results are sorted by line and the characters are selected. The final recognition results are summarized, and the results of the most frequent occurrences in each line are obtained from bottom to top using a multi-image processing algorithm. The same principle applies to the recognition results of the slab head. Finally, a line matching algorithm is used with the string to obtain the final result.

[0055] The slabs transported by the track vehicles on-site vary, generally ranging from one to four slabs. However, due to issues with the marking machine in the continuous casting workshop, some slabs experience marking failures. Therefore, it's necessary to identify the head of the slab during the identification process. This helps address both the issue of slab quantity and row matching. This is because the warehousing clamps can only hold a maximum of two slabs at a time, making the number of slabs on the track vehicle and the order of their numbers two key problems to be solved. To address this, post-processing of single and multiple image recognition results is proposed. For single image recognition results, an unordered array is obtained, a portion of which contains the character category of the entire image and its (x,y) coordinates at the top left and bottom right corners of the image.

[0056] For each image, row matching is performed. The top-left y-axis coordinates of the results are iterated through, and the largest value is selected as the bottom row of the slabs carried on the railcar. The bottom-right y-axis coordinates of the identified results are then used to determine the y-coordinate. center Iterate through the y-axis coordinates of the other characters, and if their y-coordinates are... center Similarly, the coordinates fall between the upper left corner (y0) and lower right corner (y1) of the two y-axis. Here, a constant deviation 'a' is set based on different on-site conditions, resulting in a range [y0+a, y1-a]. If y centerIf the result falls within the range, it is combined into a new array, and after all the results are looped, the last row of the slab character result is obtained, and all the characters in the row are deleted in the single image. Continue to loop through the recognition results of this image, select a new y as the second row, traverse the other two coordinate values of the character y axis, and similarly obtain a range [y1+a, y2-a]. If it falls within the range, it is combined into a new value, and the second row content in the recognition result is deleted. In this way, all the results are looped and processed until the original recognition result array is empty, and the bottom-up slab string line matching process is completed. Then sort each layer of string respectively, compare the x coordinate array of each line of string, and sort from left to right from small to large, and complete the x sorting to obtain the final correct format of the recognition result of the single image on the rail car. Similarly, the head of the slab is sorted from bottom to top, and because each row of the slab head has only one, if multiple are present in a single image, the recognition is wrong, so only the y coordinate line of the slab head is judged, and no x sorting is performed.

[0057] Generally, the slab number is composed of 12 digit characters, and in particular, more than 12 characters have been found to be printed on a row of slabs on site. For this, after sorting each row of slab number from left to right, an x coordinate distance judgment is added, and if the number of recognized characters in the row is more than 12, the distance between each character in the row is calculated. The largest distance is selected by traversing, and the characters in the row are divided into left and right strings according to the distance, the lengths of the two strings are judged, and the longer one is selected as the recognition result of the row. If the lengths of the two are equal, the left (or right) string is selected as the recognition result of the row, and the length of the selected string is looped and judged. If it is greater than 12, continue to traverse to find the maximum distance, and loop the above process until the length of the selected string is equal to or less than 12. If the continuous casting plant has a situation of missing spray of slab number, and the number of slab number in a row is less than 12, the normal recognition returns the result of the row, and waits for subsequent manual processing.

[0058] Because of the limitation of the installation position of the camera on site, a single image cannot capture all the slabs, and some slab numbers are printed on the head of the slab, some are located in the middle, and some are located at the tail of the slab. For this irregular position printing, a multi-image processing method is proposed. With the movement of the rail car, the images in the process of the rail car passing through are recognized from the head to the tail, and the correct arrangement rule of the slab number and the slab head is obtained by using the deep learning neural network recognition and single image processing algorithm. Using the recognition processing results of multiple single images in different positions and different angles, the multi-image processing algorithm can effectively solve the problem of irregular position of slab number printing, and finally using the slab number and the slab head for line matching can also solve the problem of no number on the slab.

[0059] The final result is then verified by a database query. If successful, automatic identification is completed. If the query fails, the system client will issue an alarm, and manual intervention will be required to enter the slab number.

[0060] To increase the reliability of the final identification, the identification results need to be verified a second time. The identified plate number is queried and matched with the WMS database and the continuous casting database. If the query is successful, the original PDI data of the slab is obtained, including the slab's length, width, and thickness specifications.

[0061] Before loading slabs into the warehouse, the length, width, and thickness of the slabs typically need to be measured. Each stack in the warehouse can hold a maximum of 12 slabs. Therefore, the overhead crane must know the number of slabs and their information before loading. The warehouse's loading plan contains various slab information, but it needs to identify which slab it is before loading to correctly determine which stack in the warehouse. Therefore, manual data entry is not only slow but also inaccurate, significantly impacting the warehouse loading schedule. This invention automatically identifies the slab number before loading, queries it against the WMS plan and continuous casting database to confirm its information, and achieves automatic verification and loading before the furnace, improving the overall production efficiency of the workshop line.

[0062] Figure 3 This is an example diagram of the automatic slab number identification based on deep learning provided by the present invention. Figure 2 Further, more detailed explanation:

[0063] The recognition program captures raw images of the site using a camera, and then uses a pre-trained neural network model to recognize these images, obtaining three slab head bs and 36 slab characters. Figure 2 The single-image row sorting algorithm obtains the three slab headers (b) from bottom to top and the corresponding string for each row. Then, the results are combined using a multi-image algorithm to output the final recognition result. A secondary verification is performed by querying and matching this result against information in the database to obtain valid slab size information. If the slab number query fails, manual intervention is required. The slab is modified via the client, completing the slab entry. Finally, the correct slab number is used to query the database for its various information to complete the warehousing and stockpiling plan.

[0064] Figure 4 An example was given. Figure 3 This article will provide a detailed introduction to the multi-frame algorithm in the text, but due to space limitations, only three locations will be described:

[0065] The first position recognition result after line matching and sorting is from bottom to top and from left to right: empty, 22, 228122, 228121; because the position of the slab number is irregular after triggering the recognition signal, there may be no shooting or no full shooting. The second position recognition result is: 228111112140, 228121112140, 228121112201, 228121112202; the slab number in this position is all shot and recognized. The third position recognition result is: 228111112140, 228121112140, 228121112201, 228121112202; then the multi-frame sorting algorithm is used to process the three recognition results, the string between each image is matched and summarized, and the summarized result is selected and processed, and the string with the most occurrences in a line is retained as the recognition result of the line. The first line has the following results: empty, 228111112140, 228121112140; the result with the most occurrences is selected as the recognition result of the line: 228111112140; the second line has the same result: 228121112140; the third line has the same result: 228121112201; the fourth line has the same result: 228121112202; the slab head in the detection result needs to be matched and selected, and four groups of slab head coordinates are obtained, then the four groups of strings and the four groups of slab heads are matched, which can effectively deal with the situation that there is no character in a line. Finally, the recognition result of the track trolley this time is: four slabs, 228111112140, 228121112140, 228121112201, 228121112202; the above is the complete multi-frame processing recognition algorithm, and the final recognition result given by the recognition program after processing is obtained.

[0066] Figure 5 The schematic diagram of the slab number recognition client is as follows:

[0067] The figure is the visualization of the slab number recognition result on the slab yard track trolley, a total of two overpass trolleys are deployed with slab number recognition program, the client is connected with two angle cameras above the field to shoot the running situation of the overpass trolley, the middle part shows the recognition camera shooting the slab number passing situation, because the environment of the field is sometimes very bad, the light problem is an important factor affecting the recognition result, especially at night, so the supplemental light is added beside the recognition camera, and because the slab number itself has some unavoidable defects, for example: slab number printing failure, no number on the slab, slab temperature is too high to cause the slab number to separate, the trolley moves with a certain shaking to cause the image blurred by the camera, the different trolley moving causes the different angle of the slab number, even the printing error slab number, so adding multiple recognition can effectively solve the slab number printing position and angle problem. For some serious conditions such as no slab number, printing error, slab number falling off, manual input button is added below the client, if the recognition is wrong or the slab number is not standard, the alarm is initiated and the slab number is supplemented by the field staff. In the process of program testing and use, it is found that the error is less, and most of them are caused by the non-standard slab number, for the non-standard slab number, the field staff may not be able to correctly record the slab number for input, the staff responsible for printing in the continuous casting workshop needs to be consulted, considering that the communication between the trolleys is difficult, and it may not be able to solve the problem in time, so the log function is added, all the input failure conditions are recorded one by one, and finally modified by the relevant personnel, which not only can free a staff, but also can overall check the failure result.

[0068] The above is the preferred embodiment of the present application, for ordinary skilled in the art, can understand and realize the present application, and without departing from the principle range described in the present application, some adjustments and improvements can be made, these adjustments and improvements should also be considered within the protection scope of the present application.

Claims

1. A deep learning-based method for identifying slab numbers, characterized in that, Includes the following steps: The video recording of the slab movement process was acquired, and the video recording was extracted frame by frame to obtain a large number of images with slab numbers; Machine vision recognition technology was used to label the slab number on each slab image, and the dataset required for training the slab number recognition model was created. A deep learning neural network is used to train and debug the recognition model; The recognition model is deployed on-site, recognizing a segment of video of the entire vehicle's passage each time. The matching results are then processed to obtain the number of slabs and their corresponding slab numbers, and the recognition result is output. The specific steps are as follows: Using a single thread, after obtaining the recognition signal, the camera captures a single image, the model recognizes the single image, and after post-processing, the recognition result of the single image is saved. Then, the camera captures a single image again, the model recognizes it, and after post-processing, the recognition result of the single image is saved. The single image recognition program is set to recognize images in a loop for 30 seconds. Finally, all single image recognition results within 30 seconds are summarized, processed, and fed back to the database. The algorithm processes the recognition results of a single image by selecting and sorting the recognition results of individual characters of the slab number from bottom to top and from left to right, and combining them into a string. Similarly, the recognition results of the slab head detection results are processed by sorting them from bottom to top, thus completing the single image post-processing algorithm. The process involves processing the recognition results of multiple images, summarizing the string recognition results of each image by performing row matching, and then selecting the result that appears most frequently in that row as the final result for that row. Finally, the string recognition result of each line is matched with the row position of the slab head to complete the post-processing algorithm for multiple images; The final result of the post-processing algorithm for multiple images is written into the database, and several original images from the recognition process are written into the database as the basis for subsequent manual modification or inspection of the board number. The identification results are automatically queried in the database's entry plan for secondary verification. If successful, the identification is considered successful; otherwise, an alarm is triggered.

2. The slab number identification method according to claim 1, characterized in that, The method for acquiring images with slab numbers is as follows: RGB cameras are installed along the slab's travel path to record the complete process of each trolley passing in front of the camera and capture the slab number when the slab moves. From the acquired video recording, extract images frame by frame; Process the obtained slab images and filter out images without slab numbers.

3. The slab number identification method according to claim 1, characterized in that, The method for creating the dataset required for training the slab number recognition model is as follows: Use the image annotation tool labelme to annotate the obtained images with slab numbers and slab heads; If, during the labeling process, issues are found such as slab number distortion, printing errors, missing slab number, or images of the slab head; Convert the JSON format obtained from the labeled images into a TXT file.

4. The slab number identification method according to claim 1, characterized in that, The training and debugging process of the recognition model is as follows: The collected dataset was divided into training and validation sets. Set hyperparameters, use data augmentation, and train a deep learning recognition model; Evaluate the results of the trained neural network model.

5. The slab number identification method according to claim 4, characterized in that, The model used is YOLOv5-L.

6. The slab number identification method according to claim 1, characterized in that, The process of performing secondary verification on the recognition results is as follows: Establish a database link to enable communication with the WMS database and the continuous casting workshop database; Based on the identified slab number, search for the original PDI data of the slab in the WMS database and the continuous casting database; If the query is successful, it means that the secondary verification was successful and the identification was correct. If the query fails, an alarm will be triggered to prompt manual intervention.

Images