Detection and Removal of Visible Watermarks Combined with Perceptual Attention

Abstract

In view of the diversity of visible watermarks and the high cost of human intervention in the process of removing visible watermarks, a method of removing visible watermarks including detection and removal was proposed. Watermarking detection based on YOLO V3 model adopts model fusion method and improves the loss function to improve the generalization of watermarking detection. Based on Generative Adversarial Nets (GAN), the structure similarity is introduced into the loss function by combining perceptual attention mechanism. The network generates images prior to repair images by learning regions other than the watermark. Experimental results show that the proposed method can improve the objective evaluation indexes such as peak signal to noise ratio (PSNR) and structural similarity.