Robust audio watermark decoding by nonlinear classification

This paper introduces an audio watermark (WM) decoding scheme that performs a Support Vector Machine (SVM) based supervised learning followed by a blind decoding. The decoding process is modelled as a two-class classification procedure. Initially, wavelet decomposition is performed on the training audio signals, and the decomposed audio frames watermarked with +1 and -1 constitute the training sets for Class 1 and Class 2, respectively. The developed system enables to extract embedded WM data at lower than -40dB Watermark-to-Signal-Ratio (WSR) levels with more than 95% accuracy and it is robust to degradations including audio compression (MP3, AAC), and additive noise. It is shown that the proposed audio WM decoder eliminates the drawbacks of correlation-based methods.