Learning to Restore a Single Face Image Degraded by Atmospheric Turbulence using CNNs
This work addresses image restoration for face recognition systems affected by atmospheric turbulence, representing an incremental advance in domain-specific computer vision.
The paper tackled the problem of restoring face images degraded by atmospheric turbulence, which causes geometric deformation and blur, by proposing a deep learning framework that estimates prior information and uses a novel loss function to correct distortions and improve visual quality, achieving significant improvements on both synthetic and real images.
Atmospheric turbulence significantly affects imaging systems which use light that has propagated through long atmospheric paths. Images captured under such condition suffer from a combination of geometric deformation and space varying blur. We present a deep learning-based solution to the problem of restoring a turbulence-degraded face image where prior information regarding the amount of geometric distortion and blur at each location of the face image is first estimated using two separate networks. The estimated prior information is then used by a network called, Turbulence Distortion Removal Network (TDRN), to correct geometric distortion and reduce blur in the face image. Furthermore, a novel loss is proposed to train TDRN where first and second order image gradients are computed along with their confidence maps to mitigate the effect of turbulence degradation. Comprehensive experiments on synthetic and real face images show that this framework is capable of alleviating blur and geometric distortion caused by atmospheric turbulence, and significantly improves the visual quality. In addition, an ablation study is performed to demonstrate the improvements obtained by different modules in the proposed method.