Image Gallery of Results from my Research

Here are some of the results of my research, to give you a feel of what I was working with!

Recent Research Contributions:

In Visual Tracking, we have proposed a Correlation based Fusion tracker that has the complementary advantage over most of the existing recent best trackers. This work is published in IEEE Tran. on Image Processing (pdf), and was shortlisted in Visual object challenge (VOT) 2016. A video demonstrating its comparative performance with few of the best trackers is here.

Another area where I have been working is on Rotational invariant character / object classification. Recently two of the M.techs who worked on this problem have developed a set of algorithms (such as RICNN and RIMCNN) which can classify characters, texture and objects without reference to their orientation without data augmentation. Further, when combined with per-trained models like ImageNet they can be used for classifying 1000s of classes irrespective of their rotations. The challenge is to nicely integrate them with region proposal networks such as RCNN to come up with a quite general rotational invariant scene annotation / scene captioning. Below is a snapshot of a result on basic integration proposed.

During Post-doctoral research, we proposed Fluid flow estimation with a weighted version of Ensemble Transform Kalman filter (WETKF). The WETKF is formulated in a fully non-linear setting employing Navier-Stokes dynamics and image reconstruction error based observation model. This approach is demonstrated to extract motion fields and vorticities from different kinds of fluid flows including Meteorological, experimental fluids and Oceanic image sequences.

Here are few typical comparisons with state-of-art fluid flow analysis techniques: The state of art approaches considered for comparison are second order div-curl regularization approach of Yuan (07), Heas's (09) Bayesian selection with self-similar prior and a Variational assimilation approach of Papadakis (07).

On Particle image Velocimetry (PIV) images:

Scalar image sequences:

Experimental 2D Turbulence images:

Schlieren Photographic soap film images of turbulent water:

Simulated Ocean images with holes:

Real Ocean images with holes and coast:

During the PhD, we developed novel approaches for efficient denoising of different types of degradations.

This work is based on Kalman filter framework, Markov random field (MRF) modeling and Importance sampling approaches.

Incorporation of the MRF modeling of the images, and analyzing the MRF probabilistic models with Montecarlo approaches within the Kalman framework has given rise to a number of simple and efficient denoising algorithms.

Starting with the Additive white Gaussian noise (AWGN) case, the research considered the denoising of the logarithmic film-grain noise and multiplicative Speckle noise. The novel filter developed for the denoising of AWGN is referred to as the Importance Sampling Kalman Filter (ISKF), since it extracts the first two statistics of the prior discontinuity-adptive MRF model using Importance sampling principle, at each Kalman step for estimating each pixel in a recursive manner .

Here are the AWGN denoising results from our ISKF algorithm and a comparison with the state-of-the-art methods:

Denoising of logarithmic film-grain noise and multiplicative speckle were accomplished by similar modification to a non-linear counterpart of Kalman filter, the unscented Kalman filter, and the developed novel denoising algorithm in non-linear case is referred to as Importance Sampling Unscented Kalman filter (ISUKF).

In film-grain noisy case, the comparisons were made with Modified Wiener filter (MWF) developed by Tekelp et al. and Particle filter (PF) proposed by Ibrahim et al.

Here are the film-grain noise removal results from our ISUKF algorithm and a comparison with the state-of-the-art methods:

In case of Speckle noise the strandrad speckle filters such as Frost, Lee and recent Wavelet speckle filters are used for comparison.

Here are the Speckle suppression results from our ISUKF algorithm and a comparison with the state-of-the-art methods:

My PhD research later aimed at developing a simple and efficient approaches for simultaneous denoising and digital inpainting (filling in the missing regions) of the images. This is accomplished by developing a simple edge-linking-based inpainting approach and incorporating it into the above developed non-linear denoising approach (ISUKF).

Here are the results from our denoising and inpainting methods, and a comparison

with the state-of-the-art methods:

Here are the results from our Adaptive Particle Filtering (APF) technique,

and comparison with previous methods.