The Study of Tesseract OCR Performance for Image Processing in Financial Transaction Verification

Abstract

This research aims to enhance Tesseract OCR efficiency for financial transaction verification by examining five image preprocessor codenames: Image Preprocessor Alpha (IPPA), Image Preprocessor 2 or IPP2, IPP6, IPP7, and IPP12. (Disintegration, n.d.) The study employed a two-phase methodology: accuracy testing on 116 bank transfer receipts and performance evaluation using datasets of 250 to 1,500 transactions. The findings revealed IPP12 (incorporating Grayscale, Resize, Contrast, and Sauvola Threshold) as the optimal preprocessing approach. The Sauvola Threshold method effectively consolidated preprocessing techniques into a single, accurate method. Accuracy metrics using Confusion Matrix demonstrated impressive results: Accuracy (81.03%), Precision (82.14%), Recall (90.79%), and F1-Score (86.26%), indicating balanced performance in reducing false positives and maintaining comprehensive receipt detection. Performance testing using GO programming language benchmarking tools showed the system achieved an average processing time of 5.313 seconds, utilizing 0.277 GB memory and 5.47M allocated space. The research highlights the significance of strategic image processing technique selection in improving OCR performance for financial document verification.