Search

When imaging a target in space, atmospheric turbulence might change the direction of the path of light. This might be caused by the random motion of the turbulent medium leading to drastic distortion of the image. In order to correct the temporally varying blur, reduction spatially, and geometric distortion, a blind deblurring atmospheric turbulence based on the convolutional network is proposed. This includes image reconstruction subnetwork, asymmetric U-net, and feature extraction noise suppression block (FENSB). In the U-net, instead of the traditional convolution layer that is normally used, a deblurring noise suppression block (DNSB) is used. The basic idea behind this methodology is to reduce the noise before deblurring occurs. Hence the aspect of noise suppression before deblurring is carried out by the DNSB block along with FENSB. At the time of convolutional encoding, DNSB and FENSB are used to identify and extract the rich features of the maps and further suppress noise. In order to fuse high-level and low-level features, the subnetwork is skip connected with FENSB thereby improving the integrity of the system during the process of reconstruction. This mechanism of increasing the training difficulty of the network in a step-by-step manner is used to slowly converge into a complex model facilitating better quality of images despite the turbulence. The results of simulation data, as well as that of real data, indicate that it is possible to suppress the noise and restore the image with high efficiency. [ABSTRACT FROM AUTHOR]