Sistem Pengenalan Citra Dokumen Teks Terdistorsi menjadi Teks Menggunakan Metode Deep Learning

Abstract

A common issue in document image processing is the inability of OCR systems to accurately read text from blurred images. This study aims to develop a deep learning-based OCR pipeline capable of recognizing text in blurred document images. The process begins with image enhancement using the DnCNN model for deblurring, followed by character segmentation and classification of A–Z characters using a CNN trained on the EMNIST Letters dataset. The recognized characters are then reconstructed into complete text. Experiments were conducted on 300 blurred images with varying levels of blur (low, medium, and high). Evaluation using PSNR and SSIM metrics showed improvements in image quality, with an average PSNR of 29,56 dB and SSIM of 0.89. Furthermore, the character classification accuracy reached 95.64%. Compared to the baseline (direct Tesseract OCR without deblurring), the proposed system showed a significant improvement in text readability. These results demonstrate the effectiveness of CNN-based approaches in enhancing OCR performance on blurred document images.