An adaptive framework to image watermarking based on the twin support vector regression and genetic algorithm in lifting wavelet transform domain

摘要

A novel adaptive gray scale image watermarking approach based on the combination of machine learning (ML) algorithms in wavelet domain is presented. Based upon fuzzy entropy information, non-overlapping and significant regions are selected. Lifting wavelet transform (LWT) is performed on selected significant regions in order to obtain low frequency sub band and underwent through the QR factorization. Prominent low frequency features of each region are supplied as input features for the training purpose of Lagrangian twin support vector regression (LTSVR) model. Then the optimal value of watermark scaling factor (strength) obtained using genetic algorithm (GA) is used to embed the watermark in the test data of output wavelet coefficient obtained by trained LTSVR. Arnold transformation is performed for the security of watermark along with the imperceptibility and robustness. The experimental results as well as the comparison between traditional methods and the proposed one showed a significant improvement in robustness in terms of image processing attacks which makes it suitable for implementing copyright protection applications.