Projects

• Unpaired Image Denoising

  August 2019 - May 2020 M. Tech Thesis
  Priyatham Kattakinda, A N Rajagopalan
  IIT Madras, Chennai, India
  Paper Code Presentation Project Page
  Deep Learning algorithms for image denoising are predominantly either fully supervised or fully unsupervised. Using flow-based generative models, we developed a semi-supervised method for image denoising that is applicable in scenarios where clean images are available, except that they are not ground truths for any noisy image (note that it's not possible to use supervised learning here). This work was published in IEEE International Conference on Image Processing (ICIP), 2020.

• SEM Image Super-resolution

  May 2019 - July 2019 Internship
  Priyatham Kattakinda
  KLA Corporation, Chennai, India
  The quality of a scanning electron microscope (SEM) image is largely determined by the scan rate of the electon beam. Low scan rates produce high resolution images but take a long time. On the other hand, high scan rates are fast, but the image quality is poor. Any downstream tasks (such as defect detection) are also affected by this tradeoff. To alleviate this tradeoff, I synthesized a GAN-based super-resolving CNN that reproduces any defects in the input faithfully. Interestingly, the outputs from this network perform better than even the ground truths, presumably, due to the smoothing properties of the CNN. Jointly mentored by Prof. A N Rajagopalan, Nimisha T M.