Pushing the Performance Limit of Scene Text Recognizer without Human Annotation