ConvNet-HIDE: Deep-Learning-Based Dual Watermarking for Health-Care Images
Preetam Amrit, Naman Baranwal, Kedar Nath Singh, Amit Kumar Singh
Abstract
Recently, significant attention has been focused on the copyright protection of medical images, causing watermarking to become a prevalent research topic. In comparison to embedding a single watermark directly into cover images, inserting multiple marks through a deep neural network enables image authentication and verification of ownership, which is more suitable for health-care applications. Most of the existing methods have proposed watermarking through the insertion of a single mark in a deep learning environment, which does not consider an excellent relationship between robustness, imperceptibility, and capacity and, therefore, limits watermarking performance. In this article, we propose a new watermarking method based on a convolutional neural network (ConvNet) called ConvNet-HIDE as well as principal component analysis (PCA). First, the customized UNet3+ model is applied to segment the region of interest and background information of the image. Then, the ConvNet model is used with Mobi-ConvNet, a downsampler, to imperceptibly hide a mark in both segmented parts of the image. Finally, the ConvNet model, which is followed by PCA, is used to recover both marks robustly. Extensive experimental results demonstrate the effectiveness of our proposed ConvNet-based method compared to state-of-the-art schemes. The method is able to achieve a peak signal-to-noise ratio and structural similarity index of 39.59 dB and 0.9812, respectively, with capacity of 2 bit per pixel. Further, the average execution time for our method is found to be 0.22 s, demonstrating its efficiency compared to other similar techniques that require more time.