Machine Learning-based Radiation Dose Meter Readings Identification Method Study

With the continuous development of China’s atomic energy industry,in order to ensure the accuracy of the radiation dose instruments used in the operation process of nuclear power plants,the measurement laboratory needs to detect the radiation dose instruments from different units through verification and testing.At present,the measurement verification of dose meters still requires the manual reading of the information of the dashboard,recording the reading data and other operations,which is not only inefficient,but also endangers the health and safety of the staff.The intelligent identification of dose meters can make the verification of radiometric meters safer and more convenient.This paper uses the text detection model and text recognition model to identify the reading of radiation dose meter.After the literature review,this paper conducts the feasibility analysis of the detection network and the identification network for the instrument reading identification,compares the traditional reading identification algorithm,and determines the technical scheme and model framework of this research work.Based on the radiation dose instrument information provided by a research institute,the radiometric meter image data set was collected,and the sample labels were obtained by manual annotation.Some instrument image mold lake will lead to low recognition accuracy,using image enhancement technology to remove the fuzzy phenomenon of instrument image.In the comparative test of digital instrument reading recognition accuracy,this experiment uses image processing technology,traditional machine learning BP network and Le Net network,and DBNet + CRNN model.The experimental results show that using the traditional text character recognition method is more accurate,but only one type of instrument can be identified at each time,which does not meet the practical application requirements of this subject.First,DBNet model is used for instrument reading text detection,and then CRNN network is used to identify the text of the detected area.This method can better deal with the reading identification of many kinds of radiometric meters,but the accuracy of instrument reading recognition needs to be improved.In the process of comparison experiment,the DBNet + CRNN model has low recognition accuracy and cannot replace the manual readings.First of all,the detection performance of the text candidate box of the predicted dose meter image is optimized,and then the text recognition algorithm of the detected reading box is optimized,so as to improve the accuracy of the reading.The main contents are as follows:(1)Combining Reg Net model set with standard residual structure and DBNet network,a text detection Reg＿DB model set based on residual convolutional neural network,which has the advantages of fast,efficient and applicable to most models of digital instruments,and helps to improve the accuracy of identification results.(2)The CRNN＿MV3 model improved combined with Mobilev3 network proposed in this paper improves the accuracy of identifying the instrument image reading by 4.84%.The CTC loss function used can identify the digital instrument detection box information of indefinite length,and the recognition task can also be completed in a short time when the data amount is very large.Experiments show that the proposed algorithm has a significant optimization effect when testing a large number of radiometric instrument images.The machine learning-based digital identification method of radiation dose meter proposed in this research work can accurately detect and extract the effective readings of the radiation measurement dashboard,and also quickly and accurately identify the text candidate box readings,which has broad practical application prospects.