Work Ex

Undergraduate Reasearch Assistant, Tufts University

I'm currently a part of the Tufts Neutrino Group, working under Dr. Taritree Wongjirad on the MicroBoone experiment.
I'm designing a continuous conditional Generative Adversial Network (cGAN) that can be used to simulate neutrino events.
I'm also a junior investigator at IAIFI.

I implemented Single Image Super Resolution Convolution Neural Networks to downsample Sea Surface Temperature (SST) fields of Bay of Bengal from a spatial resolution of 15 km to 5km and 5km to 1km.
I trained the Very Deep Super Resolution Convolution Neural Network (VDSR).
The network achieved a Peak Signal Noise Ratio (PSNR) Gain of 12 between input and expected resolution.
The Root Mean Square Error (RMSE) was in the order of 0.0001 degrees Celsius between the predicted and expected resolution.

I built a pipeline capable of detecting six classes of wrinkles on human faces irrespective of their age, gender or face tone.
Since wrinkles are faint, I experimented with image processing techniques like Binarization, Gaussian Filter and Hessian Line Tracking in order to highlight them.
Certain wrinkles (like those around the eyes) are more common than others. I selected the RetinaNet architecture to tackle the imbalanced dataset.
The pipeline achieved a Mean Average Precision (maP) of 0.5. It was deployed on the CureSkin application, with over a million downloads.

I worked on a service for JioLiv.
I devised an intelligent algorithm based on structural similarity index to filter out non-redundant movie frames.
I implemented a Tensorflow-based text detection and recognition pipeline to read words off the movie frames.
Subsequently, I formulated a logic using spaCy and TextRank to extract contextually significant phrases (single-worded or multi-worded) from the generated text.

I built a pipeline to detect price tags from images of supermarket shelves and interpret prices from them.
To detect minuscule price tags, the Single Shot Object Detector based on Feature Pyramid Network was implemented.
A YOLO-based Multi-Digit Classifier was trained to read digits off the price tags.
The entire pipeline achieved a Mean Average Precision (mAP) of 0.6, and is being deployed by Xplorazzi Technologies. [View Details]