Research and projects
Research and projects
Embedded Files
Weighted LASSO for Sparse Recovery With Statistical Prior Support Information
Weighted LASSO for Sparse Recovery With Statistical Prior Support Information
This is a course project done along with Keshava Kotagiri and Rajkumar for course fulfillment of 'Advanced Digital Image Processing' in Spring 2025 at IIT Bombay. We study how statistical Prior Support Information (PSI) of sparse signal improves the number of measurements for sparse recovery. We have studied, experimented and presented some of the results of this paper.
Results
Results
For our experiments it is inferred that information on prior support serves as a better initial starting point than the LASSO algorithm.
Bayesian Defogging
Bayesian Defogging
This is a course project done along with Anuj Agrawal for course fulfillment of 'Medical Image Computing' in Spring 2025 at IIT Bombay. We have studied, experimented and presented some of the results of this paper.
In adverse weather conditions like fog and haze, the images captured suffer from non-trivial degradation. We aim to remove such degradation and restore the true scene appearance i.e. how the scene would have looked on a clear day. This process is called defogging.
We estimate the scene albedo and depth from a single foggy image by alternating minimization strategy.
Results
We compute the scene albedo and depth for several images.
A foggy image
Vision Language Model Guided Knowledge Distillation for Time Series Anomaly Detection
Vision Language Model Guided Knowledge Distillation for Time Series Anomaly Detection
This is a course project done along with Zahir Khan for course fulfillment of 'Deep Learning - Theory and Practice in Autumn 2025 at IIT Bombay.
In this work, we have demonstrated the visual reasoning power of VLMs to detect anomalies in multivariate time series data. The fine-tuned teacher VLM model transfers multi-modal reasoning to a compact LSTM based student model. We did the experiments on the MIT-BIH ECG dataset. It is found that the models are able to detect the anomalies in unseen data with substantial accuracy (75-90%). We have trained the models for detecting both sparse anomalies and dense anomalies. The materials for this project including the report, code walk-through, code demo, poster, etc. can be accessed from this link.
Predictions of the VLM based teacher model
Predictions of the student LSTM model
Dense anomaly predictions(shown in green band)
Page updated
Google Sites
Report abuse