MantaRayWmark: An image adaptive multiple embedding strength optimization based watermarking using Manta Ray Foraging and bi-directional ELM

Abstract

Digital watermarking schemes based on a single value of embedding strength do not take into account the local characteristics of the host signal for watermark embedding. Consequently, striking a balance between imperceptibility and the robustness of the watermarking scheme becomes a challenging task. The use of multiple values of the embedding strength provides a solution to this problem. Multiple embedding strength-based watermarking schemes often deploy meta-heuristic techniques to look for locally optimal values of the embedding strength. However, meta-heuristic techniques being inherently slow, such watermarking schemes cannot be applied in real-time applications. This paper proposes MantaRayWmark—a novel image adaptive watermarking scheme that uses Manta Ray Foraging Optimization (MRFO) to optimize the locally relevant multiple embedding strengths (MES) for balancing the imperceptibility and robustness. However, the time-intensive MRFO-based optimization is carried out only once to train a bi-directional ELM (B-ELM) for a set of images. Each training image is subjected to 4-level DTCWT to produce approximation coefficients. For B-ELM training, the approximation coefficients act as the input features and the corresponding MES values (obtained by way of MRFO using imperceptibility and robustness as the optimization criteria) act as the target outputs. Subsequently, given a new image, the trained B-ELM is used to predict the optimized multiple embedding strengths. For watermark embedding, an image is transformed using 4-level DTCWT and the resulting approximation sub-band matrix is decomposed using SVD to get the principal component (PC) matrix. To avoid the false-positive problem, a binary watermark is inserted into the PC matrix using the MES. To ensure the robustness of the proposed scheme against geometric attacks, we have deployed a novel convolution neural network (ConvNet) based geometric correction procedure. Further, image authentication is achieved by applying SVD-based hashing on the watermarked image. To ensure the security of watermark, the watermark is scrambled using transposition cipher. Experimental results reveal that MantaRayWmark achieves a high degree of watermark recoverability from the signed image in the presence of a number of common geometric and image processing attacks, thus proving its robustness. The results also demonstrate the superiority of the proposed MantaRayWmark when compared to the existing state-of-the-art schemes in terms of robustness and imperceptibility. The proposed scheme being time-efficient may find use its in various multimedia applications.