vioft2nntf2t|tblJournal|Abstract_paper|0xf4ff02d82b000000a121060001000b00
Copyright protection of digital images is an important commercial requirement to individual artists and large organisations alike. Wavelet-based image watermarking methods have been in practice due to their robustness against standard geometrical and image processing attacks. Convolutional Neural Networks (CNNs)-based watermarking methods are becoming popular as they provide a new dimension to the generation of a watermarked image, which is perceptually close to the original image when trained over a large class of images, thereby eliminating the need to train on each image that is to be watermarked. However, the watermark extraction performance of CNNs when used in standalone mode reduces in the presence of adversarial examples. In this study, we combine the robustness of a multi-level Discrete Wavelet Transform (DWT) and the power of CNNs and propose a robust blind grayscale image watermarking method. In the proposed method watermark is of the same size as the original image thereby demonstrating the robustness under increased payload as well. The quality of the extracted watermark is measured using Structural Similarity Index Measure (SSIM), Peak-Signal-to-Noise ratio (PSNR) and Normalized Cross Correlation (NCC). Our proposed method provides high quality watermark extraction under geometrical, image processing and adversarial attacks including second watermarking by an attacker.