A Multilevel Shrinkage Approach for Curvelet Denoising

This paper suggests an image restoration technique when the image is corrupted by additive white Gaussian noise. Based on the fact that the discrete curvelet transform is redundant, it proposes a scale adaptive threshold design for curvelet denoising where at each scale of curvelet transform a different threshold is applied to the transform coefficients to restore a noise free image. The strategy is to generate a set of thresholds corresponding to the various subbands of the transform whereas the traditional soft/hard thresholding applies the same threshold to each scale of transform coefficients. It is demonstrated numerically that this scheme obtains comparable performance to the state-of-the-art denoising approaches for a wide range of noise levels. Due to the adaptive support, the edges are clean and the restored images are visually pleasant.