DWT-SVD Based Watermarking on CUDA

This paper presents a parallel DWT-SVD based watermarking algorithm using CUDA. We implement the Daubechies' wavelets by using a 2D version of the analysis and synthesis filter banks, which can be decomposed into a 1D filter bank to the columns of the image and the one to the rows, meanwhile use CUDA to reprogram the process. The GPU-accelerated GESVD function is adopted to compute the singular value decomposition (SVD) of a general rectangular matrix. With a 1Kx1K image and a programmable NVIDIA GTX 280 GPU, more than 20 times of speedup can be achieved for both one scale and three scale Daubechies' wavelet transforms, and the GPU-accelerated SVD method outperforms the OpenCV implementation with near 4 times of speedup. By using the general-purpose GPU computing technology, the proposed DWT-SVD based watermarking algorithm achieves 2.71 times of speedup, which demonstrates the significance of parallelism.