ARF-GAN: An Arabic-aware Region-focused Adversarial Framework for Visible Arabic Watermark Detection and Removal

Abstract

This paper proposes ARF-GAN, an Arabic-Aware Region-Focused Adversarial Framework for visible Arabic watermark detection and removal.The proposed framework consists of a two-stage pipeline comprising a deep learning-based detection module and a region-focused adversarial restoration module.The main contributions include a novel region-focused adversarial formulation that directs learning toward watermark-affected regions and an Arabic-Aware Structural Preservation Loss that combines weighted reconstruction, structural similarity, and texture consistency constraints to preserve image fidelity during restoration.Experiments were conducted on Arabicwatermarked versions of BOSSBase, MS COCO, and ImageNet datasets, and all results were reported as mean ± standard deviation over 30 independent runs.For watermark detection, MobileNetV2 achieved accuracies of 96.57± 0.61% on BOSSBase and 96.48 ± 0.66% on MS COCO, while ResNet50V2 obtained the highest accuracy on ImageNet with 93.54 ± 0.82%.For watermark removal, the proposed ARF-GAN achieved 30.41 ± 0.64 dB PSNR and 0.941 ± 0.010 SSIM on BOSSBase, 20.37 ± 0.81 dB PSNR and 0.612 ± 0.032 SSIM on MS COCO, and 28.36 ± 0.71 dB PSNR and 0.914 ± 0.014 SSIM on ImageNet.Comparative experiments showed that ARF-GAN achieved competitive restoration performance, with an average of 26.38 ± 0.59 dB PSNR, 0.822 ± 0.012 SSIM, and 0.160 ± 0.015 LPIPS, while using a compact 8.9M-parameter inference model specifically designed for visible Arabic watermark removal.These findings demonstrate that the proposed Arabic-aware optimization strategy provides an effective and computationally efficient solution for visible Arabic watermark detection and removal while preserving structural and perceptual image quality.