Semi-Instance Normalization Network for Turbulence Degraded Image Restoration

In computer vision tasks, a variety of normalization methods are widely used. Compared with other normalization methods, Instance Normalization (IN) performs better in turbulence degraded image restoration. However, the simple application of IN to a degraded image restoration network can be suboptimal. In this paper, we present a novel block named Semi Instance Normalization Block (SIN Block), which can improve the performance of the image restoration network. SIN Block incorporates original features in the normalization layer, which can preserve contextual information. Furthermore, we designed a semi-instance normalization Network (SINet) consisting of a series of the SIN Block for restoring turbulence degraded images. Extensive experiment on simulation dataset demonstrates that SINet can effectively restore details of the turbulence degraded image and sharpen its edges.