Semi-Supervised Landmark-Guided Restoration of Atmospheric Turbulent Images

Amit Kumar

Abstract:

Abstract—Image degradation due to atmospheric turbulence
(AT) which is common while capturing images at long ranges
adversely affects the performance of tasks such as face alignment
and face recognition. To the best of our knowledge, there
does not exist any dataset consisting of turbulence-degraded

face images along with their annotated landmarks and ground-
truth clean images, making supervised training challenging. In

this paper, we present a semi-supervised method for jointly
extracting facial landmarks and restoring the degraded images
by exploiting the semantic information from the landmarks. The
proposed approach learns to generate AT images by combining
the content from a clean image and turbulence information
from AT images in an unpaired manner. Next, we use heatmaps
from the landmark localization network as a prior to the image

restoration module. Subsequently, we impose heatmap consis-
tency loss and heatmap confidence loss to regularize the restored

images. Extensive experiments demonstrate the effectiveness of
the proposed network, which achieves an NME of 2.797 on the
task of landmark localization for strong turbulent images and
yields improved restoration results compared to state-of-the-art
methods.