OCR for Drawing Images Using Bidirectional LSTM with CTC
Hee-Ran Shin, Jang‐Sik Park, Jong-Kwan Song · 2019
Nowadays, the system is changing with the Optical Character Recognition (OCR) technology in various industrial. It is technology that allows computers detect and convert form handwritten or scanned images into searchable machine encoded data. Searchable data can save a huge amount of time and effort. Therefore, the machine industry has also improved portability and accessibility by converting existing dealer customers' mechanical parts drawing books into mobile. Dealers can easily order mechanical parts without looking for drawing, just search a serial number of mechanical parts on the database. In this paper, we propose the OCR on the drawing images to convert images into searchable data. Proposed OCR consist of three parts: pre-processing, deep learning and postprocessing. The pre-processing part removed the guide lines and shape for improving the precision. The guide lines and shape usually cause false recognition as `1' or `7'. The deep learning part extract features and classify the images. We adopt Shi, et al.'s CRNN architecture [1]. The post-processing, drop the low probability of object. Most of left lines and shapes are seem to the low probability. As result of proposed OCR, it shows 98.52% of average recall rate and 92.25% average precision.