Kernel Estimation Using Total Variation Guided GAN for Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Super-resolution (SR) kernel estimation is crucial for mitigating artifacts caused by degradation kernels in SR algorithms.
• Existing methods like KernelGAN and E-KernelGAN have limitations in estimating sizable and anisotropic kernels due to insufficient consideration of image structural information.

Purpose of the Study:

• To propose a novel kernel estimation algorithm, Total Variation Guided KernelGAN (TVG-KernelGAN), that effectively utilizes image structural information.
• To enhance the accuracy and stability of SR kernel estimation, particularly for challenging kernels.

Main Methods:

• Developed TVG-KernelGAN, a generative adversarial network (GAN)-based approach incorporating Total Variation (TV) regularization.
• The method guides the network to focus on the structural details of input images for improved kernel estimation.

Main Results:

• TVG-KernelGAN demonstrates accurate and stable estimation of SR kernels, outperforming previous methods, especially for sizable and anisotropic kernels.
• Qualitative and quantitative evaluations confirm the superiority of the proposed algorithm.
• Non-blind SR methods utilizing TVG-KernelGAN show significant performance improvements.

Conclusions:

• TVG-KernelGAN offers a robust solution for SR kernel estimation, addressing limitations of prior approaches.
• The proposed method enhances the performance of SR algorithms by providing more accurate kernel estimations.