dikubab/Amharic_OCR

Amharic OCR based on MMOCR ## Detection and Recognition of Amharic Scene Text using the toolbox This toolbox is an open-source toolbox based and for details of installation and related information see (github.com. Generally, Geʽez/Abugida/Ethiopic script has up to 519 characters. For Amharic, we use 289-319 characters depending on whether we use Ethiopic numerals and punctuation. ## Amharic Text Detection dataset preprocessing We have two datasets for the detection task. HUST-ART is the real word dataset, and HUST-AST is the synthetic dataset. HUST-ART consists of 1500 training images and 700 test images. HUST-AST comprises 75,904 training images. To convert the dataset labels to MMOCR format, use tools/data/textdet/icdar_converter.py as follows python tools/data/textdet/icdar_converter.py det_datasets/HUST-ART -o det_datasets/HUST-ART -d icdar2015 --split-list training test ## Amharic Text Recognition We have two training sets and two test sets datasets. Tana (TN) and Waliya (WL) training sets consist of 2.85 and 6M cropped words, respectively. HUST-ART and ABE test sets consist of 4039 and 5218 text images. We also have a validation dataset composed of 14835 text images, which is the training part of HUST-ART and ABE. All five datasets are in LMDB format. The toolbox usage 1. In the directory configs/_base_/recog_pipelines/, you have different pipelines you must change dict(type='LoadImageFromFile') to dict(type='LoadImageFromLMDB'), 2. In the directory configs/_base_/recog_datasets/, you need to modify the path of test and train datasets. 3. In the directory mmocr/models/textrecog/convertors/ base.py define the dictionary using the 314 Amharic characters. No need to worry we have modified it. Based on your character set, modify dict_type in all other related files. We have modified the configs/textrecog/satrn/satrn_small.py settings. You can use it as an example. The datasets for both detection and recognition can be downloaded from the website g …