OCRopus

In addition to the recognition scripts themselves, there are a number of scripts for ground truth editing and correction, measuring error rates, determining confusion matrices, etc. OCRopus commands will generally print a stack trace along with an error message; this is not generally indicative of a problem (in a future release, we'll suppress the stack trace by default since it seems to confuse too many users). To recognize pages of text, you need to run separate commands: binarization, page layout analysis, and text line recognition. The default parameters and settings of OCRopus assume 300dpi binary black-on-white images. If your images are scanned at a different resolution, the simplest thing to do is to downscale/upscale them to 300dpi. The text line recognizer is fairly robust to different resolutions, but the layout analysis is quite resolution dependent. There are some things the currently trained models for ocropus-rpred will not handle well, largely because they are nearly absent in the current training data. That includes all-caps text, some special symbols (including "?"), typewriter fonts, and subscripts/superscripts. This will be addressed in a future release, and, of course, you are welcome to contribute new, trained models. This will create a directory "linegen/..." containing training data suitable for training OCRopus with synthetic data. Project Announcements : The text line recognizer has been ported to C++ and is now a separate project, the CLSTM project, available here: New GPU-capable text line recognizers and deep-learning based layout analysis methods are in the works and will be published as separate projects some time in 2017. Please welcome @zuphilip and @kba as additional project maintainers. @tmb is busy developing new DNN models for document analysis (among other things). (10/15/2016) A lot of excellent packages have become available for deep learning, vision, and GPU computing over the last few years. At the same time, it has become feasible now to address problems like layout analysis and text line following through attentional and reinforcement learning mechanisms. I (@tmb) am planning on developing new software using these new tools and techniques for the traditional document analysis tasks. These will become available as separate projects. Note that for text line recognition and language modeling, you can also use the CLSTM command line tools. Except for taking different command line options, they are otherwise drop-in replacements for the Python-based text line recognizer. OCRopy and CLSTM are both command line driven programs. The best way to contribute is to create new command line programs using the same (simple) persistent representations as the rest of OCRopus. Python and C++ models can not be interchanged, both because the save file formats are different and because the text line normalization is slightly different. Error rates are about the same. In addition, the C++ command line tool ( clstmctc ) has different command line options and currently requires loading training data into HDF5 files, instead of being trained off a list of image files directly (image file-based training will be added to clstmctc soon). The CLSTM project also provides LSTM-based language modeling that works very well with post-processing and correcting OCR output, as well as solving a number of other OCR-related tasks, such as dehyphenation or changes in orthography (see our publications). You can train language models using clstmtext . Generally, your best bet for CLSTM and OCRopy is to rely only on the command line tools; that makes it easy to replace different components. In addition, you should keep your OCR training data in .png/.gt.txt files so that you can easily retrain models as better recognizers become available. After making CLSTM a full replacement for ocropus-rtrain / ocropus-rpred , the next step will be to replace the binarization, text/image segmentation, and layout analysis in OCR