OCR with no shape training | IEEE Conference Publication | IEEE Xplore