Digitization of Old Indonesian Documents using YOLO and OCR
Description
NOTE: The Dataset For TrOCR Is Available In HuggingFace NusaAksara https://huggingface.co/datasets/NusaAksara/NusaAksara/viewer/Image%20Transcription%20(OCR)/train?f%5Bscript%5D%5Bvalue%5D=%27jawa%27&views%5B%5D=image_transcription_ocr This data is a combination of data for TrOCR Text Transcription of Jawa Aksara from NusaAksara, and A collection data for YOLOv8 manually created and annotated from Robflow consisting of document images from NusaAksara Jawa Aksara Image Segmentation Dataset, and Single page images of historical documents from The British Public Library: Endangered Archives Program. Link To NusaAksara (OCR Transcription): https://huggingface.co/datasets/NusaAksara/NusaAksara/viewer/Image%20Transcription%20(OCR)/train?f%5Bscript%5D%5Bvalue%5D=%27jawa%27&row=4707&views%5B%5D=image_transcription_ocr Link To NusaAksara (OCR SEGMENTATION): https://huggingface.co/datasets/NusaAksara/NusaAksara/viewer/Image%20Segmentation/train?f%5Blanguage%5D%5Bvalue%5D=%27Jawa%27&views%5B%5D=image_segmentation Link EAP: https://eap.bl.uk/search
Files
Steps to reproduce
YOLO: 1. For the YOLO dataset, it is a YOLOv8 Dataset from Roboflow, all you need is to upload it to google drive, and copy the directory you saved the Folder in google drive for training/manipulation in google colab. TrOCR: 2. It is a dataset available in HuggingFace, you can load it using load_dataset from datasets script through python, and filter the dataset to OCR Transcription (Jawa Script) by filtering the train catalogue of data as seen below: dataset = load_dataset("NusaAksara/NusaAksara", "Image Transcription (OCR)") print(dataset) TARGET_SCRIPTS = {"jawa"} train_ds = dataset["train"].filter(lambda x: x["script"] in TARGET_SCRIPTS)