Large-Scale Visible Watermark Detection and Removal with Deep Convolutional Networks

Abstract

Visible watermark is extensively used for copyright protection with the wide spread of online image. To verify its effectiveness, there are many researches attempt to detect and remove visible watermark thus it increasingly becomes a hot research topic. Most of the existing methods require to obtain the prior knowledge from watermark, which is not applicable for images with unknown and diverse watermark patterns. Therefore, developing a data-driven algorithm that suits for various watermarks is more significant in realistic application. To address the challenging visible watermark task, we propose the first general deep learning based framework, which can precisely detect and remove a variety of watermark with convolutional networks. Specifically, general object detection methods are adopted for watermark detection and watermark removal is implemented by using image-to-image translation model. Comprehensive empirical evaluation are conducted on a new large-scale dataset, which consists of 60000 watermarked images with 80 watermark classes, the experimental results demonstrate the feasible of our introduced framework in practical. This research aims to increase copyright awareness for the spread of online images. A reminder of this paper is that visible watermark should be designed to not only be striking enough for ownership declaration, but to be more resistant for removal attacking.