The Comparison of Deep Learning Driven Optical Character Recognition for Hard Disk Head Slider Serial Number

This paper presents the development of an optical character recognition system by using deep learning which is trained by the Faster_Rcnn_Inception_V2 and SSD_Mobilenet_V1 model through the Tensorflow object detection API library. To build the character set recognition system, 466 images of the character set of hard disk head slider were used for training the model to build the character set recognition system. For the Faster_Rcnn_Inception_V2 model, the model was trained by 12570 epochs with the 0.0456 of loss value. In case of the SSD_Mobilenet_V1 model, the model was trained by 99051 epochs with the 0.9329 of loss value. From the results, the character set recognition system using Faster_Rcnn_Inception_V2 model, SSD_Mobilenet_V1 model and the template matching method, the accuracy in character set recognition of these methods are 83.5%, 1.5% and 4.5% respectively. The software creation of this paper focuses on the Python language and the TensorFlow library.