A blind digital audio watermarking scheme based on EMD and UISA techniques

Abstract

In this paper, a new perceptual spread spectrum audio watermarking scheme is discussed. The watermark embedding process is performed in the Empirical Mode Decomposition (EMD) domain, and the hybrid watermark extraction process is based on the combination of EMD and ISA (Independent Subspace Analysis) techniques, followed by the generic detection system, i.e. inverse perceptual filter, predictor filter and correlation based detector. Since the EMD decomposes the audio signal into several oscillating components–the intrinsic mode functions (IMF)–the watermark information can be inserted in more than one IMF, using spread spectrum modulation, allowing hence the increase of the insertion capacity. The imperceptibility of the inserted data is ensured by the use of a psychoacoustical model. The blind extraction of the watermark signal, from the received watermarked audio, consists in the separation of the watermark from the IMFs of the received audio signal. The separation is achieved by a new proposed under-determined ISA method, here referred to as UISA. The proposed hybrid watermarking system was applied to the SQAM (Sound Quality Assessment Material) audio database (Available at http://sound.media.mit.edu/mpeg4/audio/sqam/) and proved to have efficient detection performances in terms of Bit Error Rate (BER) compared to a generic perceptual spread spectrum watermarking system. The perceptual quality of the watermarked audio was objectively assessed using the PEMO-Q (Tool for objective perceptual assessment of audio quality) algorithm. Also, using our technique, we can extract the different watermarks without using any information of original signal or the inserted watermark. Experimental results exhibit that the transparency and high robustness of the watermarked audio can be achieved simultaneously with a substantial increase of the amount of information transmitted. A reliability of 1.8 10^− 4 (against 1.5 10^− 2 for the generic system), for a bit rate of 400 bits/s, can be achieved when the channel is not disturbed.

Appendix: Empirical mode decomposition

The EMD technique was firstly developed by Huang et al. [13] to decompose any nonstationary and nonlinear signal into oscillating components obeying some basic properties. The key benefit of using the EMD technique is that an automatic decomposition and fully data adaptive. This method makes it possible to decompose the audio signal x(n), to a set of $\textsf{IMF}$ components, as follows:

$$ x(n)=\sum\limits_{\iota=1}^{N}\,\textsf{IMF}_{\iota}(n)+r_{N}(n), $$

(12)

where N represents the number of the $\textsf{IMF}_i(n)$, and r _N (n) is the final residue. The EMD technique can decompose the host signal from the highest frequency components to the residue r _N (n), which represents the lowest frequency component. An audio signal example (Harpsichord sound) and some selected $\textsf{IMF}$s components are shown in Fig. 12.

Fig. 12

EMD of an audio signal (Harpsichord sound) showing selected $\textsf{IMF}$ components

The $\textsf{IMF}$s are the bases for representing the time series data. Being data adaptive, the basis usually offers a physically meaningful representation of the underlying processes. There is no need of considering the signal as a stack of harmonics and, therefore, EMD is ideal for analyzing nonstationary and nonlinear data [13]. The EMD process can also be considered as dyadic filter-bank especially for white noise and the $\textsf{IMF}_i(n)$ components are all normally distributed [13].

The EMD technique has also many interesting features [13]: it is characterized by its completeness, that is x(n) can be exactly reconstructed by linear superposition of all the components. It generates roughly orthogonal components, because the $\textsf{IMF}$s have different local frequencies at the same time. Each $\textsf{IMF}_i(n)$ has a well defined Hilbert transform from which the instantaneous frequencies can be calculated.