Watermarking is a method in computer security by which identifiers of sources or copyright owners of digital signals are embedded into the respective signals themselves in order to keep track of where a signal comes from or who the copyright owners are. In general, a watermarking system must have two characteristics: perceptual transparency and robustness. This thesis proposes a method for transparently and robustly embedding a watermark into the curvelet transform of grayscale images. The image is partitioned into small blocks; Fast Discrete Curvelet Transform (FDCT) via Unequally-Spaced Fast Fourier Transforms (USFFT) is employed to decompose each block into curvelet domain. We embed the watermark into the selected blocks, scale and curvelet coefficients based on the edge map of the cover image. The embedding strength is adjusted by a Just Noticeable Distortion (JND) model computed for each curvelet coefficient. Robustness is tested against a variety of types of image attacks. Since the curvelet transform enables most of the energy of the object to be localized in just a few coefficients, the optimally sparse representations of image edges allows for the embedded watermarks to recover from severe image degradation. However, the block-based watermarking algorithm in curvelet domain provides low robustness against geometrical distortion because geometrical distortion (e.g. rotation) desynchronizes the embedding location in the cover work. A scheme relying on the radon transforms and edge detection is developed to synchronize embedding location before the watermark detector is applied. The proposed scheme estimates the geometrical distortion the cover image was subjected to and restores the distorted image to its original state. Thus, the improved watermarking system provides high tolerance to geometric attacks as well as normal image processing. We also propose a technique for selecting the threshold for watermarking detection based on statistical analysis over host signals and embedding schemes. Experiments show our scheme is capable of keeping the probability of false positive and false negative both low and is generally robust against a wide range of image attacks. Finally, we present a new quality measure, M-SVD, which expresses the perceived distortion of image/video. We show our measure to be strongly correlated with evaluations by the Human Visual System. The quality of the watermarked images marked by our proposed curvelet based algorithm is evaluated with this approach. The evaluation demonstrates the transparency of our watermarking system, the performance of JND modeling is also confirmed with M-SVD.