An Adaptive Thresholding Algorithm-Based Optical Character Recognition System for Information Extraction in Complex Images

Abstract

Extracting texts from images with complex backgrounds is a major challenge today. Many existing Optical Character Recognition (OCR) systems could not handle this problem. As reported in the literature, some existing methods that can handle the problem still encounter major difficulties with extracting texts from images with sharp varying contours, touching word and skewed words from scanned documents and images with such complex backgrounds. There is, therefore, a need for new methods that could easily and efficiently extract texts from these images with complex backgrounds, which is the primary reason for this work. This study collected image data and investigated the processes involved in image processing and the techniques applied for data segmentation. It employed an adaptive thresholding algorithm to the selected images to properly segment text characters from the image's complex background. It then used Tesseract, a machine learning product, to extract the text from the image file. The images used were coloured images sourced from the internet with different formats like jpg, png, webp and different resolutions. A custom adaptive algorithm was applied to the images to unify their complex backgrounds. This algorithm leveraged on the Gaussian thresholding algorithm. The algorithm differs from the conventional Gaussian algorithm as it dynamically generated the blocksize to apply threshing to the image. This ensured that, unlike conventional image segmentation, images were processed area-wise (in pixels) as specified by the algorithm at each instance. The system was implemented using Python 3.6 programming language. Experimentation involved fifty different images with complex backgrounds. The results showed that the system was able to extract English character-based texts from images with complex backgrounds with 69.7% word-level accuracy and 81.9% character-level accuracy. The proposed method in this study proved to be more efficient as it outperformed the existing methods in terms of the character level percentage accuracy.

Optical Character Recognition Engines Performance Comparison in Information Extraction

“…Haraj and Raissouni produced an average of 95.77% charcater accuracy using tesseract and opencv library over 4 sample images in 2015 [17]. Those research [14,15,16,17] only used relatively small samples (less than 50 documents), while our study used more documents (8,562 documents in 6 Categories and two document structures). Previous research [14,15,16,17], which also employed the Tesseract library, only used string matching to measure the OCR.…”

“…Similar research employed the Tesseract library [16] with only 11 images as input yielding 69.7% precision. On the other hand, our study produced 83.07% precision with 8,562 documents as the same library input.…”

“…Kumar and friends produced 97% accuracy for small scanned bill documents and 83% accuracy for small scanned bill documents using Tesseract OCR on 25 scanned bills in 2020 [15]. Akinbade and friends produced 81.9% character accuracy and 69.7% word accuracy on 11 sample images in 2020 [16]. Haraj and Raissouni produced an average of 95.77% charcater accuracy using tesseract and opencv library over 4 sample images in 2015 [17].…”

“…Those research [14,15,16,17] only used relatively small samples (less than 50 documents), while our study used more documents (8,562 documents in 6 Categories and two document structures). Previous research [14,15,16,17], which also employed the Tesseract library, only used string matching to measure the OCR. On the other hand, our study used four measurements, i.e., conversion time, NER time, string match accuracy as precision, and the number of entities acquired as recall.…”

“…An offline desktop-based application called Foxit, an online-based application called PDF2GO, and an open-source OCR library called Tesseract were used to convert all documents. Patel et al in 2012 [14] produce 70% accuracy, Kumar et al in 2020 [15] produce 97% accuracy and, Akinbade [16] produce 81.9% accuracy using the Tesseract library on scanned documents.…”