Impact of Image Preprocessing on OCR Performance for English Printed Receipts: A Comparative Study of Tesseract and Easy-OCR

Abstract

Optical Character Recognition (OCR) plays an important role in extracting machine-readable information from scanned documents and receipts; however, recognition performance can be affected by image quality and document characteristics. Although image preprocessing is commonly used to improve OCR input quality, its effectiveness may vary across OCR systems. This study experimentally investigates the effect of image preprocessing on the performance of two open-source OCR engines, Tesseract and EasyOCR, for English printed receipt images. A controlled comparative experiment was conducted using 200 images from the Scanned Receipts OCR and Information Extraction (SROIE) dataset. Each OCR engine was evaluated under four image conditions: the original images as a baseline, Gaussian Blur, Contrast Limited Adaptive Histogram Equalization (CLAHE), and Fast Non-Local Means (Fast NL-Means) denoising. Recognition performance was assessed using Character Error Rate (CER) and Word Error Rate (WER), while OCR processing time was used to evaluate computational efficiency. The results demonstrate that preprocessing does not provide a uniform improvement across OCR systems. CLAHE produced the best recognition performance for EasyOCR, reducing CER from 27.10% to 26.23% and WER from 61.17% to 58.38%. For Tesseract, Gaussian Blur produced the lowest WER, reducing it from 49.46% to 48.54%, whereas the original images yielded the lowest CER of 26.19%. Fast NL-Means did not provide an overall recognition benefit for either system. Tesseract also exhibited substantially lower OCR execution time than EasyOCR under the tested conditions. These findings indicate that image preprocessing should not be regarded as a universally beneficial OCR step; instead, its selection should be considered in relation to the OCR engine, document type, and image characteristics.