SaveDoc: A Robust Technique for Degradation Classification on Ancient Document Images

Saddami, Khairun; Ramadhana, Rizka; Fathurrahman, Fathurrahman; Munadi, Khairul; Arnia, Fitri

doi:10.22060/eej.2026.24823.5773

SaveDoc: A Robust Technique for Degradation Classification on Ancient Document Images

Document Type : Research Article

Authors

Khairun Saddami ¹

Rizka Ramadhana ¹

Fathurrahman Fathurrahman ²^{, 1}

Khairul Munadi ¹

Fitri Arnia ¹

¹ Department of Electrical and Computer Engineering, Faculty of Engineering, Universitas Syiah Kuala, Banda Aceh, Indonesia

² Eletrical Engineering Department, College of Engineering and Physics, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia

10.22060/eej.2026.24823.5773

Abstract

The analysis of degraded ancient documents remains challenging because multiple degradation patterns may appear within a single document image and reduce document quality. In this paper, we propose a combination of convolutional neural networks and extreme learning machine for patch-level degradation classification in ancient document images. The dataset was constructed from 100 source document images and cropped into 700 labeled patches. Four pre-trained architectures, namely VGG19, ResNet101, MobileNetV2, and EfficientNetB0, were used as feature extractors. The number of hidden nodes in the ELM classifier was tuned, and two activation functions, namely radial basis function and sigmoid, were evaluated. The experimental results show that EfficientNetB0 features combined with the extreme learning machine classifier achieved the best performance, with a testing accuracy of 98.21% ± 2.08%. This combination also produced faster inference time compared with the other evaluated models. These results indicate that the EfficientNetB0-extreme learning machine combination is effective for patch-level degradation classification in ancient document images.

Keywords

Degraded Ancient Documents

Degradation Classification

Extreme Learning Machine

Pre-trained CNN features

EfficientNet

Subjects