Research on Optical Character Recognition for Nuclear Power Components in Underwater Environment Based on Deep Learning

Nuclear underwater fuel assembly OCR can recognize and interpret characters, signs and symbols in the underwater environment of high radiation and low visibility, which is of great significance to ensure the safe operation of nuclear power plants. Four batches of fuel assembly images in different nuclear power environments are collected. According to the character recognition requirements, a one-step character recognition method is proposed. The character detection model of the one-step character recognition method has a mAP50 of 0.963 on the test set, and there are many character misrecognition problems in the prediction results. The reason for character misidentification is that the ratio of the character area to the size of the captured image is less than 1:5000, so the character area target is too small. Aiming at the problem of one-step character recognition, this paper proposes a two-step character recognition method. The mAP50 of the character detection model of the two-step character recognition method can reach 0.994 on the test set. The two-step character recognition method can effectively solve the problem of one-step character recognition method, but there is a problem of misidentification of fuel assembly characters in other nuclear power plants in the prediction results. According to the problems existing in the two-step character recognition method, the paper puts forward the three-step character recognition method at last. The character classification model of the three-step character recognition method can reach top1_accuracy of 0.996 on the test set. Using the fifth batch of fuel assembly images in nuclear power environment to test the three-step character recognition method, the test results show that the three-step character recognition method can effectively suppress the problems existing in the one-step character recognition method and the two-step character recognition method, improve the generalization ability of fuel assembly character recognition, and have better field application value.