Research on watermarking removal algorithm based on deep learning

Noise2noise is a classical algorithm for image denoising in neural network deep learning. In this paper, some improvements are made to Noise2Noise (N2N for short), such as changing the underlying network of the original network to empty convolution, using SRResNet in SRGAN to improve the image quality, and selecting an adaptive parameterized Relu activation function. At the same time, the improved N2N algorithm is trained and tested by using self-built watermark data set. Experimental results show that the improved N2N algorithm can achieve good watermark removal effect, and can not only remove single visible watermark, but also remove large-scale watermark images in batches, which has strong practical value for image restoration and other fields.