A Deep Neural Network Approach for Classification of Watermarked and Non-watermarked Images

Digital watermarking is the process of embedding an unique mark into digital data to prevent counterfeit. With the exponential increase in the data, the process of segregating a watermarked and non-watermarked images is very time consuming. It is necessary to automate the process of differentiating a watermarked and a non-watermarked images as well as identifying whether the given image is watermarked or not for identifying the authenticity. In this paper, we propose to use Deep Autoencoders, a form of deep neural networks for classification and identification of watermarked and non-watermarked images. The experiments are carried out using NWND dataset originally with 444 images. These images are watermarked using image, shape and text watermarking techniques to make the entire dataset to 1776 images. The experiment results show that, deep neural networks performed better that traditional feed forward neural networks. The classification accuracies with Original - IW for DAEN and ANN are 77.9% and 25.9 % respectively. Whereas for Original - SW and Original - TW, it is 82.1% and 32.7%, 64.2% and 20.06% respectively. The DAEN was able to identify 86 images correctly out of 100 images supplied which is 86% of accuracy with an average training rmse of 0.06423 and testing rmse of 0.0784.