Towards a Robust Feature-based Watermark

A Fourier domain watermark localized near image feature points is presented in an attempt to make an invisible watermark that is robust to common image processing attacks. Finding robust feature points allows for localized watermarking to be detected after attacks. The feature points are found using the Harris corner detector. Points that respond greatest to the Harris corner detector will often survive attacks to the image. Applying image normalization to non-overlapping disks around the feature points allows for a stable reference point to be found in the image assuming little local variation. A binary sequence is then embedded into the Fourier domain of the localized regions and the modified regions are replaced into the original image. Some information from the original disks and the secret key are used to correct the attacked image and detect the watermark. Experimental results show the scheme is robust to rotation, JPEG compression, scaling, median filtering, sharpening, Gaussian filtering, cropping and combination attacks.