I am a Computational Imaging Scientist at Zeiss, Geramany (Feb 2024- till now) working on AI solutions for zeiss applications. Prior to joining Zeiss I worked as a Applied Scientist at Onfido (May 2022- Oct 2023) an identity verification company. During my time at Onfido we developed a document detection algorith for edge devices. I also did a PostDoc at TU Berlin in RSiM group under Prof. Begum Demir. Prior to joining postdoc I was an Algorithm Development Engineer with Ebeam division KLA+, Chennai, India (Jan 2019-Jan 2021). I worked on SEM image restoration (denoising, SR), classification and detection frameworks at KLA+. I did my PhD from Image Processing and Computer Vision lab under the supervision of Prof. A N Rajagopalan, IIT Madras. My thesis addressed the problems of hand-held imaging (mainly resolution and motion blur) and proposed several multi and single image approaches to overcome it. I did my Masters from NIT Calicut in Signal Processing (2011-13), where I worked with Dr. G Abhilash in Compressed sensing based signal recovery techniques. I am a passionate researcher and have worked on multiple image restoration, detection and classification tasks covering traditional image processing methods to to the latest deep learning techniques. Check out my CV here and links to some of the academics pages

During my Ph.D., I worked in area of low-level vision and image restoration. My research interest includes compressed sensing, Depth from defocus, Dictionary based reconstruction and Camera mapping. I am also interested in the of-late deep learning frameworks for vision.

We synthesize pan photos from motion blurred videos. Pan photography is an artistic photography intended to capture motion in images. It improves the aesthetic feel of an image. But capturing such images require great amount of skill and effort. We ease this by synthesising the same from a captured video.

Sparse representations has found great application in image processing community. The central idea here is that any natural signal can be represented sparsely in an overcomplete dictionary. We use this idea to estimate the latent image and depth map from a space variantly blurred image.

The photometric properties of a scene changes with varying camera and illumination. Finding a representation inavariant to these changes is of great importance in finding changes between scenes taken under different tiimes of day with different cameras.

Haze and color loss are the major problems in underwater imaging. When light propagates inside the water medium due to scattering particles, different wavelengths gets attenuated differently with depth. This leads to color loss and hazy affect in the captured underwater scenes. We propose here a method to color correct these images and produce its equivalent as seen from above water surface.

Estimating depth map and a high resolution image from a bunch of low resolution motion blurred images is dealt here. Given the LR frames we estimate the HR camera motion. This is used to iteratively solve for the depth map and HR clean frame using a cost function relating the two.

Single image blind-deblurring is a highly ill-posed problem. We propose a deep network which can learn blur-invariant features. Our network consists of two stages. Stage I is an encoder-decoder that learns clean data representation and Stage II consists of a Generative Adversarial Network (GAN) that learns to map a blurred frame to the clean representation.

Deep networks usually work with paired input-output samples. But in most of the cases capturing such data pairs is difficult or impossible. Hence, a network that can be trained with no data pairing is what we look forward for. We propose a GAN based network for class-specific deblurring. To constrain the solution space, we also add a reblurring and gradient modules to the network.

SEM images are innately noisy and getting a ground truth clean data to train a supervised deep learning denoiser is not possible. Noise-to-Noise denoisers exist that works irrespective of no clean GT data. We use the same model in SEM denoising with additional cost functions tailored to SEM data.

Collecting high resolution SEM images is time taking and might even lea to burning of the images site. To avoid this, we propose a SEM super-resolver that improves the throughput of the system without impacting the detection and classification accuracies.

Early detection and classification of defects in chips helps in improving the yield of semiconductor industries. A detector trained on a certain field of view (FOV) struggles to give results when the FOV during inference stage. A study on how to combat this is done. Similarly, in defect classification its important to quantify the data requirement, cross-class correlations etc to understand the performance of deep learning classifier. Research on these directions were carried out.