Abstract: This paper presented a novel adaptive video watermarking algorithm based on spatial transformer networks. An adaptive video watermarking was embedded into a block detected by the attention model from spatial transformer networks to resist noise and compress attacks, as well as the geometric attack. According to the attention mechanism derived from the masking behavior of the human perceptual model, a visual saliency block was extracted for further watermarking embedding. Experimental results show that the watermarked video is highly imperceptible and resistant to many degradations and distortions of video, and reaches over 256 bits in watermark capacity.

Key words: copyright protection, space transformation network, digital watermarking, discrete cosine transform, video blind watermarking