A Deep Learning Approach to Document Recovery: High Performance with DenoiseU-Net

Authors : Salih Can Turan, Zeki Çıplak, Ali Sarıkaş, Kazım Yıldız

Pages : 1323-1335

Doi:10.35414/akufemubid.1628066

View : 163 | Download : 270

Publication Date : 2025-12-03

Article Type : Research Paper

Abstract :Image denoising, a crucial task in image processing, has consistently faced challenges despite ongoing research efforts. In this research, the dataset was created by extracting 20,000 images from 60 public sources, some of scanned or digitized documents. Each image was checked to see if it contained at least one plain text, image, table or mathematical expression. Common noises such as random black and white pixels, Gaussian blur, gray areas, speckle noise, random directional lines, Poisson noise and salt and pepper noise were applied to the images. To create the test set, the seven types of noise were individually added to 500 images excluded from the dataset to create a balanced test set of 3500 images. The dataset consists of 23,000 images with a training and test set ratio of 5:1. In particular, our DenoiseU-Net model targets the recovery of noisy scanned documents and achieves capability on various content types such as tables, images, mathematical equations, and text. Experimental results show that the average precision, recall and F1-score of DenoiseU-Net on the test set are 99.36%, 99.59% and 99.48%, respectively. In addition to these evaluation results, the average SSIM and PSNR values, which are commonly used parameters to indicate the quality of the images, were obtained as 0.9657 and 40.28 dB, respectively. From overall results, proposed DenoiseU-Net method shows the superior performance.
Keywords : Belge kurtarma, taranmış belge, belge görüntüsü gürültü giderme, gürültülü görüntüler, derin öğrenme, yapısal benzerlik indeksi

ORIGINAL ARTICLE URL