All-round improvement in DCT-based blind image watermarking with visual enhancement via denoising autoencoder

Abstract

Discrete cosine transform (DCT)-based watermarking has received great success in the copyright protection of digital images. This study exploited the orthonormal expansion properties of DCT to achieve superior performance in computational efficiency and watermarking efficacy. In particular, we reformulated the relative modulation to allow the modification of paired DCT coefficients as if operable in the spatial domain. To account for human visual characteristics, we also adjusted the embedding strength adaptively according to image entropy. The use of a range regulation mechanism enabled perfect watermark retrieval. The pursuit of an ideal balance between robustness and imperceptibility could be benefited from the grey wolf optimizer. Using the above-discussed techniques, the proposed watermarking scheme demonstrates superior robustness in terms of bit error rate and normalized correlation coefficient with the peak signal-to-noise ratio tuned at around 38 dB. Compared to the regular DCT approach, the computational acceleration is on the scale of N2/(log2N)2, where N signifies the side length of the image block under processing. Furthermore, for the type of character-style watermark, we explored the feasibility of using a denoising autoencoder (DAE) to enhance the retrieved watermark comprehensibility. The proposed DAE was found to be capable of contributing a 60% reduction of the bit error rate, thus making the recovered watermark visually more recognizable.