Visible Watermark Removal Based on Dual-input Network
Tong Li, Bingwen Feng, Guofeng Li, Xinzhen Li, Mingjin He, Peiya Li · 2021
A visible watermark removal network that combines partial and standard convolution is proposed in this paper. Existing deep learning-based watermark removal methods use a standard convolutional network over the watermarked image, using convolutional filter responses conditioned on both valid pixels as well as the substitute values in the watermark pattern covered area (typically the mean value). When the watermark is almost opaque, this often leads to artifacts such as color discolorepancy and blurriness. Based on this, we propose a watermark removal network that uses partial convolution, where the convolution is masked and re-normalized to only be conditioned on valid pixels, and then use a dual-input branch network based on ordinary convolution for optimization, one of the input branches is connected to the partial convolution network, and the other accept the watermarked image directly. With this mechanism, different convolutions can be combined to remove watermarks with different transparency. The scheme is evaluated through a large watermark image dataset. Experimental results show that this method can effectively remove watermarks of different colors and transparency.