Visible Watermark Removal with Deep Learning Technology

Abstract

Watermarking is a technique used to assert ownership over an image, and can be categorized into visible and invisible forms based on the detectability of the watermark. Visible watermarking is more user-friendly and intuitive than invisible methods since it allows individuals to identify image ownership with their own eyes rather than relying on machine-based watermark decoders. To enhance the visual quality of watermarked images and ensure the original images can be fully recovered after visible watermark authentication, a visible watermark removal approach using deep learning-based inpainting is proposed in this paper. Experimental results demonstrate that the watermarked images carrying the visible watermark and auxiliary information achieve peak signal-to-noise ratios (PSNRs) ranging from 41.89 dB to 43.17 dB and structural similarity indices (SSIMs) up to 0.97 to 0.98. Furthermore, our hybrid recovery operations enable the complete restoration of the original images, making them easily readable.