Researcher Resume

About

Hello, I am a Fellow Researcher in CSE at IIT Bombay. I am Pursuing MS by Research and My Previous Qualification is B.Tech. Getting fired up over learning things and problem-solving are some of my main aspects of life. I want to master every skill because nothing is impossible. Knowing and applying skills on various projects has helped me to achieve some of the best results.

Research

Here are some of my Research areas

U-NET AND YOLO: AIML MODELS FOR LANE AND OBJECT DETECTION IN REAL-TIME CV
The research focuses on a self-driving vehicle system utilizing computer vision, employing CNNs like U-Net for segmentation and YOLO for object detection. Evaluation in real driving conditions demonstrates U-Net's proficiency in lane detection (RMSE score) and YOLO's speed and accuracy in object detection. The study aims to advance autonomous driving technology for secure and efficient transport networks. Newer Method of approach using binary images for prediction at higher speed.

Methodology and Phases

Still from working Model with both Ground Truth and Predicted Mask
Paper Link
BLIND CAR: REINFORCEMENT LEARNING BASED SELF-DRIVING CAR CV, RL & Evolutionary Algo
This type of research tends to search for something which in the traditional sense. Usually in a car, we as drivers get inputs from sight and sound. A self-driving car focuses on using only sight. But what if we don't rely on sight or sound? My car uses the distance from other objects with the help of fixed angles radar which gather inputs from a certain distance. Since there is no previous data available, we have to work on Reinforcement learning modules like DQN and other policies. But this work well on same kind of environment but had a hard time while driving on a B-pline Hyper generation of tracks, So we went on with Neuro-evolutionary Networks for Self-conditioning itself with each generation having a fixed time but having the tracks generated randomly to ensure maximum fitness of the evolution.

Still from Generation of 3_90 angle Configuration
SKIN CANCER DETECTION AND GENERATION OF DATA USING GANS TARGETING MULTICLASS ACCURACY CV, GANS & Transformers
Skin Cancer is very dangerous if not identified early on, but the problem arises when we have little to no information about this and we tend to ignore it till we find out at the later stages. We tried to create a Deep learning-based module that tries to identify whether it's related to cancer and classify a lesion to the types of cancer it's related to. Based on various modules to find the best method, we decided to go with ViT-based architecture which has given us more accurate results. While working with GANs to point towards N-shot learning or meta-learning as the skin cancer data is limited to certain extents, so to increase the data, I have tried various GANs like DCGAN, STARGAN, ACGAN, WGAN, etc., and its 3 combinations like STARGAN without image-to-image translation, DC+ESRGAN and ACWGAN which were having its generative patterns.