Unknown-box Approximation to Improve Optical Character Recognition Performance

This code repository contains the scripts needs to recreate the results mentioned in the manuscript "Unknown-box Approximation to Improve Optical Character Recognition Performance" (Springer publication of ICDAR 2021 | arxiv)

Contents of the repo

Scripts

• train_crnn.py - Script to train the CRNN model to avoid cold start problem
• train_nn_area.py - Script to train the NN-based preprocessor with VGG dataset
• train_nn_patch.py - Script to train the NN-based preprocessor with POS patch dataset
• train_sfe_area.py - Script to train the SFE-based preprocessor with VGG dataset
• train_sfe_patch.py - Script to train the SFE-based preprocessor with POS patch dataset
• eval_prep.py - Evaluate the preprocessor with two datasets and the two OCR engines
• properties.py - Contains global properties used by the scripts

Directories

• trained_models - Pretrained preprocessor models and CRNN models
• datasets - Contains data loader scripts
• ocr_helper - Contains codes to connect with OCR engines
• models - Contains the two models

Steps to run

1. We have used Anaconda package manager in the Linux environment (Ubuntu 18.04.5 LTS) and recommends to use the same. Use the following command to create a conda environment named ocr-test with all the dependencies.

conda env create -f environment.yml

2. Run init_workspace.py to create necessary directories. Download the three dataset zip files from "http://bit.ly/approx-ocr-grad" and put them in a directory named data and unzip them.