DVSNet: Deep variance-stabilized network robust to spatially variant characteristics in imaging

What is it about?

In real imaging systems, unwanted noises or artifacts can be caused by an image reconstruction error during photon-to-digital conversion. The generated noises typically tend to have spatially variant characteristics in an acquired image due to their signal dependencies. In this letter, we analytically introduce the noise characteristics via a preliminary study and propose an image restoration scheme based on deep variance- stabilized network. Specifically, to improve the robustness of restoration performance to the noise properties, variance-stabilizing transformation and binning priors are properly combined with a deep neural network as a layer structure particularly designed for denoising.

Featured Image

Photo by Giammarco Boscaro on Unsplash

Photo by Giammarco Boscaro on Unsplash

Why is it important?

Spatially variant characteristics in imaging are one of the troublesome issues in the deep learning-based image restoration approach. In this letter, we propose a new network model by combining the strengths of variance stabilization and deep neural network. To increase the robustness to the spatially variant degradation, we newly design the layer structure composed of 7 cascade operations and learn the network parameters in the VST domain instead of the typical pixel domain. It is clearly validated in experimental results that the proposed network model achieves superior performance to the existing alternatives in terms of both objective and subjective qualities.

Perspectives