An image watermarking scheme: combining the transform domain and deep learning modality with an improved scrambling process

This paper proposes a novel image watermarking scheme with the following two processes. In the embedding process, the cover image is subjected to NSCT to obtain the properties of multi-scale and directional characteristics and fed to DWT which decomposes the contourlet image into various sub-bands. The wavelet decomposition image is subjected to a region selection step, where the regions are selected based on the DL strategy for selecting the region from the contourlet of the cover image where the features, including LOOP feature, mean feature, Canny features, and ResNet features are extracted and subjected to the DNN classifier. Conversely, the watermark image performs NSCT and DWT as the cover image and is fed into further steps to eliminate the correlation of pixels. For that, it proposes the FTTS approach to scramble the contourlet of watermark images, in which tent map and dyadic transformation are employed to eliminate the correlation among pixels. Then the resultants of region selection and FTTS are embedded to obtain the watermarked image. On the other hand, the extraction process takes place, which extracts the embedded watermarked image and applies inverse scrambling and inverse an NSCT approach to recover the watermark image against any malicious attack.