Kartika Candra Kirana, Ira Kumalasari, Gulpi Qorik Oktagalu, Afwatul Maqbullah, Agung Faradiz Shobari, Bagus Hidayat
This study compares the performance of Tesseract, Easy-OCR, and Transformer OCR in recognizing crossed-out text in the Indonesian and English languages. The focus on crossed-out text aims to assess the ability of OCR methods to face the challenges of recognizing text with non-standard formats. This preliminary study provides an important finding for optimizing the OCR technology in crossed-out multilingual handwriting. Testing was conducted on the modified IAM Handwriting dataset and the UM-PTI-Handwriting dataset using the processing time, Character Error Rate (CER), and Word Error Rate (WER) metrics to compare Transformer (Tr-OCR, Donut), Tesseract, and Easy OCR. According to the WER and CER metrics, Transformerbased OCR (Tr-OCR) achieves the lowest error rate, achieving 0,63 WER and 0.63 CER on the modified IAM Handwriting dataset, while 1,92 WER and 3,05 CER on the UM-PTI-Handwriting dataset. In terms of the processing time, Tesseract OCR is the fastest, while Transformer-based OCR (Donut) is the slowest. It concluded that the Transformer OCR (Tr-OCR) excels in recognizing English handwritten. However, all compared OCRs have lower accuracy in Indonesian handwritten text compared to English handwritten text. © 2025 IEEE.
Universitas Negeri Malang, Department of Electrical Engineering, Malang, Indonesia; University of Tsukuba, Department of Computer Science, Tsukuba, Japan
