Multi-scale Fusion for Visible Watermark Removal

Abstract

Considering the diversified features of visible watermarks and the need for a lot of manual operations to remove visible watermarks, we propose watermark removal algorithm based on multi-scale fusion and deep learning, which is suitable for automatic detection and removal of various watermarks. We present a two-step framework: watermark detection and watermark removal. Firstly, we use target detection algorithm to extract the watermark. Then, we adopted the image-to-image translation method based on multi-scale fusion to remove the watermark, and the self-guided loss is added into the loss function. Finally, we conducts comprehensive evaluation on different datasets. Experimental results show that this framework is feasible in practical application, and the proposed network structure is superior to the existing methods in the accuracy of watermark removal.