Performance Analysis of Tesseract and EasyOCR for Bangla Optical Character Recognition on the Novel Bangla CrossHair Dataset

This paper presents a comparative study of key metrics for OCR engines in Bangla language processing. PyTesseract (a Python wrapper for Tesseract OCR) and EasyOCR were benchmarked on a novel dataset, "Bangla-CrossHair," created for testing OCR engines. This dataset combines samples from "Bangla Text Detection and Recognition," "Bangla Handwritten Characters," and "Bangla Handwritten Words" datasets and includes diverse image types, such as blurred, clear, torn, and tilted. Results show EasyOCR outperforms PyTesseract in several scenarios, while PyTesseract consistently demonstrated faster processing. Since many OCR engines provide pre-trained capabilities, traditional metrics like training and validation accuracy are challenging to measure for some models, including PyTesseract. Instead, Character Level Accuracy, Word Level Accuracy, Levenshtein Distance, Character Error Rate, Word Error Rate, Precision, Recall, F1-Score, and Inference time were measured for an unbiased evaluation.