Image Watermarking with Grouped Convolution and Residual Network

In this paper, we propose a grayscale image water-marking method using grouped convolution, residual network, and Rectified Adam (RAdam). Instead of simple convolution operator, we employ grouped convolution which is a technique of deviding the input into layers, convolving each layer, and then merging them back together. Residual layers plus skipped connections that serve as bypasses for the network is a network structure with a series of convolutional convolutional layers. RAdam is an optimizer that has the advantages of both momentum SGD and Adam, and is expected to stabilize learning. Incorporating these methods into ReDMark, an existing watermarking framework, can improve resistance to a variety of common noise attacks while preserving watermarked image quality.