Bi-directional extreme learning machine for semi-blind watermarking of compressed images

Abstract

Bi-directional Extreme Learning Machine (B-ELM) is a newly developed single layer feed-forward network capable of fast training with few hidden neurons. It is also reported to show better generalization capability as compared to its old counterpart ELM. In the past, it has never been applied to image processing data-sets and particularly to any of its applications. In this work, B-ELM is successfully used to carry out watermarking of JPEG compressed images by inserting a binary watermark into it. Two invertible activation functions – Sine and Sigmoid are tested in this work. The RMSE is plotted as a function of number of hidden neurons. As observed in case of other applications, this plot indicates that Sigmoid is better placed in comparison to Sine function. The robustness of embedding scheme is examined by applying seven different attacks over signed images. These results prove that the proposed scheme is robust enough against the selected attacks. The computed processing time for embedding and extraction in milliseconds indicates that this scheme is suitable for developing real time watermarking applications for videos.