Sistem Pencarian Dokumen Bahasa Indonesia Berbasis Konten Visual dan OCR Menggunakan Pendekatan Deep Learning

Abstract

The development of digital technology has led to the accumulation of a large number of documents, especially in the form of images or scans, which makes it difficult to search using traditional keyword-based methods. This research aims to develop a visual content-based automatic document search system for Indonesian language documents using Optical Character Recognition (OCR) technology and Deep Learning approach. The methodology used consists of three main components. Firstly, the Tesseract OCR module is used to extract text from document images, with preprocessing to improve accuracy. Second, the Word2Vec algorithm with Skip-gram architecture (vector size 300, window 10, min_count 2) is used to convert text into low-dimensional semantic vectors. Third, the Transformer paraphrase-multilingual-MiniLM-L12-v2 model is used as an initial filtering mechanism with a threshold of 0.2 to determine document relevance. The system was built using Django framework as the backend and PostgreSQL as the metadata management, with the data divided 80% for training and 20% for testing. Evaluation was performed on 50 searches with varying query lengths (16-472 tokens) against 1,000 PDF documents. The results showed excellent performance with Precision@10 of 1.0, average Recall of 0.911692, F1-score of 0.952634, and MAP@10, MRR@10, and NDCG@10 values each reaching 1.0. The search time ranged from 1.95 to 12.86 seconds, with an average of 4.43 seconds. This time variation is more influenced by the complexity of the OCR and semantic matching process, rather than the query length. The system proved effective and consistent in finding and ranking relevant documents.