A study on object recognition using deep learning for optimizing categorization of radioactive waste. (December 2020)

• Record Type: Journal Article
• Authors: Kim, Jeong-Guk; Jang, Sung-Chan; Kang, Il-Sik; Lee, Dong-Ju; Lee, Jin-Woo; Park, Hee-Seoung
• Is Part Of: Progress in nuclear energy. Volume 130(2020)
• Publication Date: 2020-12
• Subjects: Deep learning -- Object recognition -- Computer vision -- Radioactive waste -- Categorization; Nuclear energy -- Periodicals; Nuclear engineering -- Periodicals; 333.7924
• DOI: 10.1016/j.pnucene.2020.103528
• Languages: English
• ISSNs: 0149-1970

Abstract

Abstract: The smart management of radioactive waste by deep learning technology is becoming of great concern, as it could decrease the workload and errors of workers in categorizing radioactive waste, thereby reducing the waste volume. In this paper, we try to maximize the efficiency of categorization for new or temporary workers instead of skilled workers by training the categorization using deep learning technology. The waste recognition system based on the deep learning technology was trained with a total of 86, 084 images for 50 epochs with a subdivision of 8 and a batch of 128, which were extracted from video data that were taken in a waste sorting site. The image recognition was applied for four typical radioactive wastes (vinyl, rubber, cotton, and paper) with no object with hands (no object) and without hands (empty). The waste recognition was tested with a total of 21, 521 images to evaluate the accuracy. It was determined that the accuracy of the image recognition with a deep neural network was 99.67%. Graphical abstract: Image 1 Highlights: Dataset for deep learning was extracted from video taken during manual categorization work. The accuracy of the ResNet model increased as the layer increased. Regularization by the random removal was applied to the last layer to solve the overfitting problem in the model training. Trained model performed parallel processing using GPU to accelerate the detection.