3-D OCT data denoising with nonseparable oversampled lapped transform

This paper proposes 3-D OCT data denoising with nonseparable oversampled lapped transform (NSOLT), and examines the effectiveness through experiments. NSOLT is a lattice-based redundant transform which simultaneously satisfies the symmetric, real-valued and compact-support property. It is possible to apply a dictionary learning technique to the design by preparing examples. NSOLT is capable of having rational redundancy by controlling the number of channels and decimation ratio. In this study, a denoising technique is proposed by combining learned NSOLT dictionary and iterative hard thresholding (IHT), and the performance of the proposed method is evaluated for 3-D OCT data. It is verified through robust median estimator of noise variance and structural similarity index measure (SSIM) that the proposed technique yields effective denoising performance with moderate redundancy.