Intelligent detection method for ambient neutron dose monitor based on deep learning

Abstract

The invention discloses an intelligent verification method for a neutron ambient dose monitor based on deep learning, and relates to the technical field of machine vision, and the method comprises the steps: S1, collecting a picture of a camera video data stream, and carrying out the preprocessing of the camera video data stream based on an OPENCV library, so as to improve the processing speed and recognition success rate in a text recognition process; s2, realizing text detection and text recognition of the instrument picture based on a PaddleOCR library; s3, performing post-processing on the output result of the step S2 based on a PaddleOCR library so as to filter out wrongly recognized data in the recognition result array, wherein the post-processing comprises the steps of extracting a dose rate and filtering a recognition result sequence; and S4, after data identification of all the measurement points is completed, calculating statistical fluctuation, a calibration factor, a relative inherent error and relative expansion uncertainty of the instrument based on the data of each measurement point. According to the method provided by the invention, the numerical value of the ambient neutron dose monitor can be automatically read, the identification time is less than 0.5 s, and the verification efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine vision, and in particular relates to an intelligent calibration method for a neutron ambient dose monitor based on deep learning. Background Art

[0002] Neutron ambient dose equivalent monitoring is an important part of radiation protection monitoring around nuclear facilities such as nuclear power plants and accelerators. In order to ensure the accuracy of the measurement values ​​of the monitoring instrument, according to the requirements of the "Calibrating Procedures for Neutron Ambient Dose Equivalent (Rate) Instruments" (JJG852-2019), neutron ambient dose monitors need to be calibrated once a year. However, the current calibration of neutron ambient dose monitors relies heavily on manual readings, which not only brings human errors to the final results, but also has the problem of low calibration efficiency. Therefore, how to automatically identify the value of the instrument during the calibration process, improve work efficiency, and avoid clerical errors and omissions caused by manual records has become the current focus of research. Summary of the invention

[0003] In view of the defects existing in the prior art, the purpose of the present invention is to provide a deep learning-based intelligent calibration method for a neutron ambient dose monitor. This method can automatically read the numerical value of the neutron ambient dose monitor and filter out erroneous recognition results caused by missing digits and decimal points.

[0004] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] A deep learning-based intelligent verification method for a neutron ambient dose monitor comprises the following steps:

[0006] S1. Collect images of the camera video data stream of the neutron ambient dose monitor, and pre-process the camera video data stream based on the OPENCV library to improve the processing speed and recognition success rate in the text recognition process;

[0007] S2. Implement text detection and text recognition for instrument images based on PaddleOCR library;

[0008] S3, post-processing the output result of step S2 based on the PaddleOCR library to filter out the wrongly recognized data in the recognition result array, wherein the post-processing includes extracting the dose rate and filtering the recognition result sequence;

[0009] S4. After completing the data identification of all measurement points, the statistical fluctuation, calibration factor, relative inherent error, and relative expanded uncertainty of the instrument are calculated based on the data of each measurement point.

[0010] Furthermore, in the above-mentioned intelligent calibration method for neutron ambient dose monitor based on deep learning, the preprocessing in step S1 specifically includes: graying, trimming and correcting, denoising and cropping the image.

[0011] Further, in the above-mentioned deep learning-based intelligent verification method for neutron ambient dose monitor, in step S1:

[0012] The grayscale of color images is processed based on OPENCV; the correction algorithm is used to trim and correct the image; the image is denoised based on the image restoration technology of multi-resolution convolutional neural network; based on user controls, the image is cropped by manually selecting the cropping area to extract the region of interest in the image, namely the ROI area.

[0013] Furthermore, in the above-mentioned intelligent verification method for neutron ambient dose monitor based on deep learning, the specific steps of trimming and correcting the image in step S1 are:

[0014] S11. Identify points in the digital image where the brightness changes significantly and present them in the form of contours;

[0015] S12, finding the edge with the largest area among all the detected edges, which is the desired contour, and calculating its minimum circumscribed rectangle, and obtaining the four endpoints, center point coordinates and rotation angle of the minimum circumscribed rectangle;

[0016] S13. Based on the camera projection model, any quadrilateral area on the original image is mapped to a new position and shape by processing the four endpoints of the minimum circumscribed rectangle in the image.

[0017] Furthermore, in the intelligent calibration method of neutron ambient dose monitor based on deep learning as described above, a deep learning algorithm based on PaddleOCR is used in step S2 to implement OCR processing of the instrument image, and the OCR processing includes text detection and text recognition. The text detection is used to select the text box contained in the image, and its output is used as the input of the text recognition algorithm; the text recognition is used to extract and recognize the text in the text box.

[0018] Furthermore, in the above-mentioned intelligent verification method for neutron ambient dose monitor based on deep learning, the DBNet model is used in step S2 to implement text detection, specifically:

[0019] Extract features from different stages of the DBNet model backbone network and fuse them to obtain feature maps;

[0020] A probability map and a threshold map are obtained according to the feature map prediction, wherein the probability map is a text area, and the threshold map is a contraction and expansion of the text area;

[0021] The probability map and the threshold map are converted into binary maps using a binarization formula, and added to the entire training;

[0022] During training, the losses are calculated for the probability map, threshold map and binary map respectively, and finally the weighted total loss is calculated.

[0023] Further, in the above-mentioned intelligent verification method of neutron ambient dose monitor based on deep learning, the CRNN-CTC model is used in step S2 to realize text recognition. The CRNN-CTC model structure includes three parts, from bottom to top: convolutional layer CNN, recurrent layer RNN and transcription layer CTC loss. The specific process of text recognition is:

[0024] Use a standard CNN network to extract features of text images;

[0025] The feature vectors are fused using the RNN network to extract the contextual features of the character sequence, and then the probability distribution of each column feature is obtained;

[0026] Finally, the text sequence is predicted through the transcription layer CTC.

[0027] Further, in the above-mentioned intelligent verification method for neutron ambient dose monitor based on deep learning, the step of extracting dose rate in step S3 is specifically as follows:

[0028] S311, preliminarily screening all text boxes extracted in the ROI area, and if the character string contained in the text box can be converted into a floating point number, then determining that the text box is a digital information text box;

[0029] S312, filtering the text box containing "Sv / h" or "Gy / h", and extracting the numbers in the text box. If the text box does not contain numbers, it means that the text box contains dose rate units; if the text box contains numbers, it means that the text box contains dose rate values ​​and dose rate units, then the remaining characters except the numbers in the text box are deleted and the text box is also listed as a digital information text box;

[0030] S313. After completing step S312, multiple digital information text boxes are obtained. The area occupied by each text box is calculated using the text box coordinates, and the text boxes are sorted according to the area size. The digital information text box with the largest area is the dose rate value.

[0031] Further, in the above-mentioned intelligent verification method for neutron ambient dose monitor based on deep learning, the filtering and identification result sequence in step S3 is specifically:

[0032] S321, saving the recognition results into an array according to the chronological order, where the length of the array is related to the number of digits to be imported;

[0033] S322, selecting the last 10 numbers of the array as the recognition result sequence, rearranging the sequence according to the numerical value, and calculating the median thereof;

[0034] S323, traverse all numbers in the array, and determine whether the absolute value of the difference between the selected number and the median is less than half of the median. If so, retain the number; if not, remove the number from the sequence;

[0035] S324. After filtering in step S323, if there are eliminated numbers in the result sequence, insert the numbers in the spare array into the result sequence, and perform the number filtering operation in step S323 again until there are no erroneous recognition results in the result sequence or the spare array is used up.

[0036] Furthermore, the intelligent verification method for neutron ambient dose monitor based on deep learning as described above further includes after step S4:

[0037] Based on the NPOI library, the original record template and certificate template are read and data is imported to complete the generation of original records and certificates.

[0038] Compared with the prior art, the intelligent calibration method for neutron ambient dose monitor based on deep learning provided by the present invention has the following beneficial effects:

[0039] The present invention realizes the optical character recognition (OCR) of the neutron ambient dose monitor by calling OPENCV and PaddleOCR, and realizes the generation of original records and certificates by calling NPOI; the post-processing algorithm of recognition result sequence filtering is adopted to filter out the erroneous recognition results caused by missing numbers and decimal points. After the filtering algorithm, the accuracy of text recognition can reach 100%.

[0040] In summary, the method provided by the invention can automatically read the dose rate value of the neutron ambient dose monitor, with an identification time of less than 0.5s, which can significantly improve the verification efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flow chart of an intelligent calibration method for a neutron ambient dose monitor based on deep learning provided in an embodiment of the present invention;

[0042] Figure 2 Another flow chart of an intelligent verification method for a neutron ambient dose monitor based on deep learning provided in an embodiment of the present invention;

[0043] Figure 3 It is the software interface of the intelligent verification system for neutron ambient dose monitor;

[0044] Figure 4 This is a schematic diagram of the image grayscale processing result;

[0045] Figure 5 This is a schematic diagram of image trimming and correction processing results;

[0046] Figure 6 This is a schematic diagram of the background noise reduction processing results;

[0047] Figure 7 This is a schematic diagram of the image cropping processing result;

[0048] Figure 8 This is a schematic diagram of OCR processing results;

[0049] Fig. 9 Schematic diagram of error recognition sequence;

[0050] Fig.10 Flowchart of the OCR result sequence filtering algorithm. DETAILED DESCRIPTION

[0051] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0052] The present invention provides a deep learning-based intelligent calibration method for a neutron ambient dose monitor, and based on the method provides an intelligent calibration system for a neutron ambient dose monitor. Figure 1 The flowchart of the method provided in the embodiment of the present invention is shown. Figure 2 The working process of the system is shown, and the method specifically comprises the following steps:

[0053] S1. Collect images of the camera video data stream of the neutron ambient dose monitor, and pre-process the camera video data stream based on the OPENCV library to improve the processing speed and recognition success rate in the text recognition process.

[0054] First, based on the OPENCV library, the camera video stream is converted into an image data format that is easy for the system software to process, such as the Mat format; after completing the acquisition of the video data stream, the acquired image is preprocessed based on the OPENCV library. The preprocessing content includes grayscale, edge trimming and correction, noise reduction and cropping. The following is a detailed description of each preprocessing operation method.

[0055] Image grayscale: grayscale processing of color images based on OPENCV, the effect is as follows Figure 4 shown.

[0056] Image trimming and correction: The images taken by the camera may show some distortion, which will affect the OCR recognition rate. Therefore, a correction algorithm is needed to correct the image. The specific steps are as follows:

[0057] S11, edge detection: Identify points in digital images where brightness changes significantly and present them in the form of contours;

[0058] S12, contour discovery: find the edge with the largest area among all detected edges, which is the desired contour, and calculate its minimum circumscribed rectangle, and obtain the four endpoints, center point coordinates and rotation angle of the minimum circumscribed rectangle;

[0059] S13, perspective transformation: The basic principle of perspective transformation is based on the camera projection model. By processing the four control points in the image (i.e., the four endpoints of the rectangle obtained in step S12), any quadrilateral area on the original image is mapped to a new position and shape. With the help of perspective transformation, image distortion can be corrected and perspective transformation can be achieved. The image trimming and correction processing results are shown in Figure 1. Figure 5 shown.

[0060] Noise reduction: When a camera is used to shoot a meter screen, moiré and screen reflection may appear on the image screen. This is the noise in the image background. When there are too many noise points in the image, it will greatly interfere with the program's detection and recognition of text in the image, so noise reduction processing is required. The present invention uses an image restoration technology based on a multi-resolution convolutional neural network to perform noise reduction on the image. The processing results are as follows: Figure 6 shown.

[0061] Cropping: The interface of the neutron dose rate meter is usually mixed with other digital information, and different meters have different layout rules. It is impossible to use a fixed model to uniformly process multiple types of meters. Therefore, the present invention provides a user control to crop the image. The image is cropped by manually selecting the cropping area, and the subsequent model operation is based on the cropped image. This ensures that the information is extracted accurately while reducing the image resolution, thereby reducing the system's resource usage and processing time during image processing. The image cropping result is as follows: Figure 7 shown.

[0062] S2. Implement text detection and text recognition for instrument images based on the PaddleOCR library.

[0063] After completing the preprocessing of the image, the present invention uses a deep learning algorithm based on PaddleOCR to implement OCR processing of instrument images. The algorithm can realize offline OCR recognition while occupying less memory, ensuring the efficient operation of the software. OCR processing is mainly divided into two parts, text detection and text recognition. Text detection is used to select the text box contained in the image, and its output is used as the input of the text recognition algorithm; text recognition is used to extract and recognize the text in the text box. In a specific embodiment of the present invention, text detection is implemented based on the DBNet model, and text recognition is implemented based on the CRNN-CTC model, specifically:

[0064] Text detection: The present invention uses the DBNet model to implement text detection. The DBNet structure is as follows: First, it is a backbone network that extracts features from different stages and then fuses them to obtain a feature map. After obtaining the feature map, two results will be predicted, namely the probability map and the threshold map. The probability map is the text area, and the threshold map is a contraction and expansion of the text area. Afterwards, the binarization formula is used to calculate an approximate binary map, and the binarization is added to the entire training. During training, the losses of these three maps are calculated separately, and finally a weighted total loss is calculated.

[0065] Text recognition: The present invention uses the CRNN-CTC model to realize text recognition. The entire CRNN network structure consists of three parts, from bottom to top: CNN (convolutional layer), RNN (recurrent layer) and CTC loss (transcription layer). During the recognition process, the standard CNN network is first used to extract the features of the text image, and then the RNN network is used to fuse the feature vectors to extract the contextual features of the character sequence, and then the probability distribution of each column feature is obtained, and finally the transcription layer (CTC) is used to predict the text sequence. The OCR processing results are as follows: Figure 8 shown.

[0066] Since the numerical refresh rates of different instruments are different, the experimenter needs to set the OCR refresh time in the software in advance, and perform OCR calculations according to the refresh time cycle until a complete set of data for the measurement point is completed.

[0067] S3, post-processing the output result of step S2 to filter out the wrongly recognized data in the recognition result array, wherein the post-processing includes extracting the dose rate and filtering the recognition result sequence.

[0068] This step is also based on the deep learning algorithm of PaddleOCR to post-process the recognition results. The dose rate extraction algorithm adopted by the present invention filters the text recognition results based on the area of ​​the detection text box. The output of text recognition depends on the text contained in the ROI area. These texts not only contain neutron dose rate values, but also other information, such as alarm information, battery power, etc., and effective information needs to be extracted from multiple recognition texts. In addition, the dose rate value may be surrounded by the same text box as the dose rate unit, and the two need to be separated. The specific steps for extracting the dose rate are:

[0069] S311, preliminarily screening all text boxes extracted in the ROI area, if the string contained in the text box can be converted into a floating point number, then the text box is determined to be a digital information text box. In a specific embodiment, the float.TryParse method in the C# language is used to screen text boxes containing only numbers, and these text boxes are defined as digital information text boxes.

[0070] S312, filtering the text box containing "Sv / h" or "Gy / h", and trying to extract the numbers in the text box. If the text box does not contain numbers, it means that the text box only represents the dose rate unit; if the text box contains numbers, it means that the text box contains the dose rate value and the dose rate unit, then delete the remaining characters except the numbers in the text box and list the text box as a digital information text box;

[0071] S313. After the above steps are completed, multiple digital text boxes are obtained. The areas occupied by the text boxes are calculated using the text box coordinates, and the text boxes are sorted according to their areas. The digital text box with the largest area is the dose rate value.

[0072] During the verification process, there may be some misidentified numbers. The recognition errors come from missing numbers and missing decimal points. In the recognition result sequence, these misidentified results are reflected as sudden rise or fall "burrs", such as Fig. 9 As shown in . This feature can be used to filter the recognition result sequence and remove the wrong recognition results. Fig.10 As shown, the specific steps of filtering the recognition result sequence are:

[0073] S321. Save the recognition results into an array according to the chronological order. The length of the array is related to the number of digits to be imported.

[0074] Taking JJG 852 verification procedure as an example, 10 measurements are required to measure repeatability, so the array length is 20, with 10 extra numbers as backup to be replaced when an erroneous recognition result occurs.

[0075] S322, selecting the last 10 numbers of the above array as the recognition result sequence, rearranging the sequence according to the numerical value, and calculating the median.

[0076] S323, traverse all the numbers in the array, and determine whether the absolute value of the difference between the selected number and the median is less than half of the median. If so, retain the number; if not, remove the number from the sequence.

[0077] S324. After filtering in step S323, if there are eliminated numbers in the result sequence, insert the numbers in the spare array into the result sequence, and repeat the number filtering process of step S323 until there are no erroneous recognition results in the result sequence or the spare array is used up.

[0078] S325. If there are still incorrectly recognized numbers in the result sequence after all the spare arrays are replaced, a pop-up window will prompt the user that the current sequence parameters are insufficient and the user needs to wait for the software to continue reading more instrument numbers before entering data.

[0079] S4. After completing the data identification of all measurement points, the statistical fluctuation, calibration factor, relative inherent error, and relative expanded uncertainty of the instrument are calculated based on the data of each measurement point.

[0080] S5. Finally, the original record template and certificate template are read and data is imported based on the NPOI library to complete the generation of original records and certificates.

[0081] Based on the above method, the present invention provides an intelligent calibration system for neutron ambient dose monitor based on deep learning. The system software interface is as follows: Figure 3 As shown, the software includes a result saving module, an original image rotation module and a database.

[0082] The result saving module is used to save the recognition images and processing results. During the recognition process, the camera images and processing results are automatically saved to a folder to facilitate the judgment of abnormal recognition results and software optimization in the future.

[0083] The original image rotation module is used to realize the rotation of the original image to avoid image distortion caused by excessive image inclination, which in turn leads to OCR recognition errors.

[0084] The database is used to save recognition results, automatically record OCR recognition results and processing time, and display them in the interface. Users can copy recent recognition results through database query instructions and directly import them into the certificate processing software.

[0085] The present invention provides an intelligent calibration method for neutron ambient dose monitor based on deep learning. The optical character recognition (OCR) of the neutron ambient dose monitor is realized by calling OPENCV and PaddleOCR, and the generation of original records and certificates is realized by calling NPOI. The post-processing algorithm of the recognition result sequence filtering can filter out the erroneous recognition results caused by missing numbers and missing decimal points. After the filtering algorithm, the accuracy of text recognition can reach 100%. In short, the method provided by the invention can complete the automatic reading of the dose rate value of the neutron ambient dose monitor, and the recognition time is less than 0.5s, which can significantly improve the calibration efficiency.

[0086] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A deep learning-based intelligent calibration method for a neutron ambient dose monitor, comprising the following steps: S1. Collect images of the camera video data stream of the neutron ambient dose monitor, and pre-process the camera video data stream based on the OPENCV library to improve the processing speed and recognition success rate in the text recognition process; S2. Implement text detection and text recognition for instrument images based on PaddleOCR library; S3, post-processing the output result of step S2 based on the PaddleOCR library to filter out the wrongly recognized data in the recognition result array, wherein the post-processing includes extracting the dose rate and filtering the recognition result sequence; S4. After completing the data identification of all measurement points, the statistical fluctuation, calibration factor, relative inherent error, and relative expanded uncertainty of the instrument are calculated based on the data of each measurement point.

2. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 1 is characterized in that: The preprocessing in step S1 specifically includes: graying, trimming and correcting, denoising and cropping the image.

3. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 2 is characterized in that: In step S1: The grayscale of color images is processed based on OPENCV; the correction algorithm is used to trim and correct the image; the image is denoised based on the image restoration technology of multi-resolution convolutional neural network; based on user controls, the image is cropped by manually selecting the cropping area to extract the region of interest in the image, namely the ROI area.

4. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 3 is characterized in that: The specific steps of trimming and correcting the image in step S1 are: S11. Identify points in the digital image where the brightness changes significantly and present them in the form of contours; S12, finding the edge with the largest area among all the detected edges, which is the desired contour, and calculating its minimum circumscribed rectangle, and obtaining the four endpoints, center point coordinates and rotation angle of the minimum circumscribed rectangle; S13. Based on the camera projection model, any quadrilateral area on the original image is mapped to a new position and shape by processing the four endpoints of the minimum circumscribed rectangle in the image.

5. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 3 or 4, characterized in that: In step S2, a deep learning algorithm based on PaddleOCR is used to implement OCR processing of the instrument image. The OCR processing includes text detection and text recognition. The text detection is used to select the text box contained in the image, and its output is used as the input of the text recognition algorithm; the text recognition is used to extract and recognize the text in the text box.

6. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 5 is characterized in that: In step S2, the DBNet model is used to implement text detection, specifically: Extract features from different stages of the DBNet model backbone network and fuse them to obtain feature maps; A probability map and a threshold map are obtained according to the feature map prediction, wherein the probability map is a text area, and the threshold map is a contraction and expansion of the text area; The probability map and the threshold map are converted into binary maps using a binarization formula, and added to the entire training; During training, the losses are calculated for the probability map, threshold map and binary map respectively, and finally the weighted total loss is calculated.

7. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 6 is characterized in that: In step S2, the CRNN-CTC model is used to realize text recognition. The CRNN-CTC model structure includes three parts, from bottom to top: convolutional layer CNN, recurrent layer RNN and transcription layer CTCloss. The specific process of text recognition is as follows: Use a standard CNN network to extract features of text images; The feature vectors are fused using the RNN network to extract the contextual features of the character sequence, and then the probability distribution of each column feature is obtained; Finally, the text sequence is predicted through the transcription layer CTC.

8. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 7 is characterized in that: The steps of extracting the dose rate in step S3 are specifically as follows: S311, preliminarily screening all text boxes extracted in the ROI area, and if the character string contained in the text box can be converted into a floating point number, then determining that the text box is a digital information text box; S312, filtering the text box containing "Sv / h" or "Gy / h", and extracting the numbers in the text box. If the text box does not contain numbers, it means that the text box contains dose rate units; if the text box contains numbers, it means that the text box contains dose rate values ​​and dose rate units, then the remaining characters except the numbers in the text box are deleted and the text box is also listed as a digital information text box; S313. After completing step S312, multiple digital information text boxes are obtained. The area occupied by each text box is calculated using the text box coordinates, and the text boxes are sorted according to the area size. The digital information text box with the largest area is the dose rate value.

9. The intelligent verification method for neutron ambient dose monitor based on deep learning according to claim 8, characterized in that: The filtering and recognition result sequence in step S3 is specifically: S321, saving the recognition results into an array according to the chronological order, where the length of the array is related to the number of digits to be imported; S322, selecting the last 10 numbers of the array as the recognition result sequence, rearranging the sequence according to the numerical value, and calculating the median thereof; S323, traverse all numbers in the array, and determine whether the absolute value of the difference between the selected number and the median is less than half of the median. If so, retain the number; if not, remove the number from the sequence; S324. After filtering in step S323, if there are eliminated numbers in the result sequence, insert the numbers in the spare array into the result sequence, and perform the number filtering operation in step S323 again until there are no erroneous recognition results in the result sequence or the spare array is used up.

10. The intelligent verification method for neutron ambient dose monitor based on deep learning according to any one of claims 7 to 9, characterized in that: After step S4, the following steps are also included: Based on the NPOI library, the original record template and certificate template are read and data is imported to complete the generation of original records and certificates.

Images