BSS-Based Extraction For Additive Video Watermarking

We propose a method for extracting watermarks additively embedded in video frames by using blind source separation (BSS). A video consists of a sequence of multiple still images, and similarity between successive frames is high. By using the similarity between frames and the BSS technique, we show that the proposed method can perform blind decoding. After separating the original signal and watermark from the stego-frames by BSS, it is necessary to identify the frame containing the watermark from the separated signals. We propose a method to identify watermarked frames on the basis of correlation. We embedded a binary image, or logo, into frames as a watermark. Since the embedding strength β affects the quality of the video and bit error rate (BER) of the estimated image, the peak signal-to-noise ratio (PSNR) for stego-frames and BER for the image were evaluated. In the case where the strength was low, the PSNR was over 47 dB, but the BER was 0.3 at maximum. For a high strength, the PSNR was 30 to 42 dB and the BER was less than 0.1. As a result, an embedding strength of
$\beta=0.02$
was recommended, which achieved
$\text{PSNR}> 42\text{dB}$
and
$\text{BER} < 0.1$
.