Research Of Radiation Degradation Image Information Recognition Technology Based On Deep Learning

Modern economy cannot develop without energy.Nuclear energy is widely used in food and Chinese medicine irradiation,sterilization and secondary processing as an unique source of renewable energy.In the nuclear radiation environment,the accurate classification of goods category information in degraded images and the detection of abnormal changes in location information can provide a reliable theoretical basis and technical support for automated transportation and early warning protection against nuclear accidents.The traditional degraded image information recognition technology has many shortcomings and can no longer meet the requirements of intelligent production.How to use deep learning network models to overcome its shortcomings and improve the efficiency and precision of category and location information recognition of goods in nuclear radiation scenarios will become the focus of research,so this paper investigates the denoising and information recognition technology of radiation degraded images from the algorithm level.In terms of image denoising,an improved denoising algorithm is designed based on the obtained feature laws by deriving the numerical characteristics of noise through real irradiation experiments and using Image J image processing tools.The algorithm first determines the noise location based on the combination of the neighboring pixel value comparison method and the four-direction method,and then uses a weighted nuclear norm minimization algorithm to filter the image with the extracted noise location.The improved denoising algorithm is compared with many familiar denoising algorithms,and the highest peak signal-to-noise ratios at low and high cumulative doses can reach 30.85 and 23.49,respectively,which proves the effectiveness of the improved denoising algorithm and the ability of detail information protection.For image information recognition,to ensure that the algorithm can achieve a balance of accuracy and speed,the YOLOv4 network of the first stage is used as the basis of the framework,and is improved in various aspects for practical application scenarios.To reduce the parameters and computation of the network,CSPDarknet was used to replace and retrofit its original backbone network;In order to ensure the accuracy of information recognition,the classification regression layer is decomposed into two parts and integrated at the end of prediction;to suppress the positive and negative sample imbalance problem,the Focal Loss function is utilized in place of the original cross-entropy loss function;the Mosaic data enhancement,Softer-NMS replacing NMS,etc.to reduce the probability of occurrence of problems such as wrong detection,missed detection and rechecking.According to the simulation results on the homemade nuclear radiation image dataset,the average accuracy of the improved convolutional network is improved from 84.34% to 85.57%,and the accuracy-recall curves for three categories of goods recognition are better than the network before the improvement,in addition to the decrease of some network parameters and computation,so that the network capacity is reduced from 253 M to 212 M.Finally,this dissertation applies the object position information obtained by the improved convolutional neural network to tilt detection,designs three tilt detection methods according to the morphological features such as shape,volume and size of three different categories of goods,and provides early warning for goods with abnormal changes in position information according to the preset threshold.Experiments show that the algorithm in this dissertation can effectually and precisely identify the category and position information of nuclear radiation degradation images and make corresponding early warning for tilted goods in the field of view,which has high practical application value.