Integrasi Tesseract OCR dengan Deep Learning U-Net untuk Ekstraksi Nutrisi dari Label Minuman Berpemanis Buatan

Rakhmanda, Venia Anisya (2026) Integrasi Tesseract OCR dengan Deep Learning U-Net untuk Ekstraksi Nutrisi dari Label Minuman Berpemanis Buatan. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6026241008-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (31MB) | Request a copy

Abstract

Peningkatan konsumsi minuman berpemanis buatan di Indonesia mendorong kebutuhan akan sistem otomatis yang mampu mengekstraksi informasi nilai gizi secara cepat dan akurat. Namun, label kemasan umumnya memiliki karakter berukuran kecil, tata letak yang beragam, serta latar belakang yang kompleks sehingga menurunkan akurasi Optical Character Recognition (OCR) konvensional. Penelitian ini mengembangkan sistem ekstraksi informasi nilai gizi dengan mengintegrasikan U-Net dan Tesseract OCR. U-Net digunakan sebagai pengganti proses adaptive thresholding pada Tesseract untuk melakukan segmentasi area teks sehingga hanya informasi yang relevan diproses oleh OCR. Dataset penelitian terdiri atas 1.500 citra label minuman berpemanis buatan yang merepresentasikan 725 produk dan dianotasi secara manual sebagai ground truth. Tahapan penelitian meliputi preprocessing, pelatihan model U-Net, segmentasi area teks, inversi citra, ekstraksi teks menggunakan Tesseract OCR v3 dan v4, serta ekstraksi hasil OCR menjadi 12 field informasi nilai gizi menggunakan regular expression. Hasil evaluasi menunjukkan bahwa model U-Net memperoleh Dice sebesar 0,9472 pada data pelatihan dan 0,9185 pada data validasi, dengan IoU sebesar 0,9009 dan 0,8565, serta Pixel Accuracy sebesar 99,81% dan 99,71%. Integrasi U-Net terbukti meningkatkan performa Tesseract OCR, terutama pada citra dengan kualitas segmentasi yang baik, sehingga mampu menghasilkan ekstraksi informasi nilai gizi yang lebih akurat dibandingkan Tesseract tanpa integrasi U-Net. Penelitian ini menunjukkan bahwa segmentasi berbasis deep learning merupakan pendekatan yang efektif untuk meningkatkan kualitas ekstraksi informasi gizi pada label kemasan dengan teks kecil dan latar belakang yang kompleks.
======================================================================================================================================
The increasing consumption of artificially sweetened beverages in Indonesia has created a growing demand for automated systems capable of extracting nutritional information quickly and accurately. However, product labels often contain small-sized text, diverse layouts, and complex backgrounds, which reduce the accuracy of conventional Optical Character Recognition (OCR) methods. This study develops a nutritional information extraction system by integrating U-Net with Tesseract OCR. U-Net replaces Tesseract's adaptive thresholding process by performing text area segmentation, allowing only relevant text regions to be processed by the OCR engine. The dataset consists of 1,500 images of artificially sweetened beverage labels representing 725 products, with manually annotated ground truth. The proposed methodology includes image preprocessing, U-Net model training, text area segmentation, image inversion, text extraction using Tesseract OCR versions 3 and 4, and the extraction of 12 nutritional information fields using regular expressions. Experimental results demonstrate that the U-Net model achieved Dice scores of 0.9472 and 0.9185 on the training and validation datasets, respectively, with Intersection over Union (IoU) values of 0.9009 and 0.8565, and Pixel Accuracy of 99.81% and 99.71%. The integration of U-Net significantly improved the performance of Tesseract OCR, particularly for images with high-quality text segmentation, resulting in more accurate nutritional information extraction compared with the original Tesseract OCR without U-Net integration. These findings indicate that deep learning-based text segmentation is an effective approach for improving nutritional information extraction from packaging labels containing small text and complex backgrounds.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	OCR, U-Net, Tesseract, Segmentasi Teks, Ekstraksi Nutrisi, Minuman Berpemanis Buatan, Text Segmentation, Nutritional Information Extraction, Artificially Sweetened Beverages
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 59101-(S2) Master Thesis
Depositing User:	Venia Anisya Rakhmanda
Date Deposited:	28 Jul 2026 01:22
Last Modified:	28 Jul 2026 01:22
URI:	http://repository.its.ac.id/id/eprint/138224

Actions (login required)

View Item