Noise Adaptive Channel Smoothing of Low-Dose Images

Many nano-scale sensing techniques and image processing applications are characterized by noisy, or corrupted, image data. Unlike typical camera-based computer vision imagery where noise can be modeled quite well as additive, zero-mean white or Gaussian noise, nano-scale images suffer from low intensities and thus mainly from Poisson-like noise. In addition, noise distributions can not be considered symmetric due to the limited gray value range of sensors and resulting truncation of over- and underflows. In this paper we adapt B-spline channel smoothing to meet the requirements imposed by this noise characteristics. Like PDE-based diffusion schemes it has a close connection to robust statistics but, unlike diffusion schemes, it can handle non-zero-mean noises. In order to account for the multiplicative nature of Poisson noise the variance of the smoothing kernels applied to each channel is properly adapted. We demonstrate the properties of this technique on noisy nano-scale images of silicon structures and compare to anisotropic diffusion schemes that were specially adapted to this data.