Jarin Ritu - Research Projects

Improved knowledge distillation via strcutural and statistical texture information

Sponsor: Massachusetts Institute of Technology Lincoln Laboratory

Paper: Structural and Statistical Audio Texture Knowledge Distillation (SSATKD) for Environmental Sound Classification

This project investigates the limitations of conventional knowledge distillation approaches for environmental sound classification, particularly their tendency to overlook low-level audio texture information that captures fine-grained local patterns in complex acoustic environments. To address this challenge, we proposed the Structural and Statistical Audio Texture Knowledge Distillation (SSATKD) framework. SSATKD enhances traditional teacher–student learning by transferring not only high-level contextual representations but also low-level structural and statistical audio texture features extracted from intermediate network layers. By integrating both global semantics and local texture information, the framework enables more robust and discriminative feature learning.

To evaluate generalizability, SSATKD was tested across four diverse datasets: two passive sonar datasets (DeepShip and Vessel Type Underwater Acoustic Data (VTUAD)) and two general environmental sound datasets (ESC-50 and TUT Acoustic Scenes). We further explored two teacher adaptation strategies, classifier-head-only adaptation and full fine-tuning, and evaluated the framework using both convolutional and transformer-based teacher architectures. Experimental results demonstrated consistent accuracy improvements across all datasets, model architectures, and training settings, confirming the robustness and effectiveness of SSATKD for real-world environmental and underwater acoustic sound classification tasks.

Histogram Layer for Improved Target Classification

Sponsor: Massachusetts Institute of Technology Lincoln Laboratory

Paper: Histogram Layer Time Delay Neural Networks for Passive Sonar Classification

Code: Github

Passive Sonar

Media links: MIT Lincoln Lab Project

Underwater acoustic target detection in remote marine sensing operations is challenging due to complex sound wave propagation. Despite the availability of reliable sonar systems, target recognition remains a difficult problem. Various methods address improved target recognition. However, most struggle to disentangle the high-dimensional, non-linear patterns in the observed target recordings. In this work, a novel method combines a time delay neural network and histogram layer to incorporate statistical contexts for improved feature learning and underwater acoustic target classification. The proposed method outperforms the baseline model, demonstrating the utility in incorporating statistical contexts for passive sonar target recognition. The code for this work is publicly available.

In addition, I contributed to active sonar data analysis by developing segmentation methods to improve target localization and interpretation.

Active Sonar

Ritu, Jarin, et al. "Histogram layer time delay neural networks for passive sonar classification." 2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2023.

Sethuraman, Advaith V., et al. "Machine learning for shipwreck segmentation from side scan sonar imagery: Dataset and benchmark." The International Journal of Robotics Research 44.3 (2025): 341-354.

Past Projects

Neonatal facial features for detecting acute pain using Machine Learning approach

Project : Neonatal facial features for detecting acute pain using Machine Learning approach

Paper: Facial Detection for Neonatal Infant Pain using Facial Geometry Features and LBP

Neonatal pain assessment is essential for infants concerning their health issues. There have been several studies to assess the pain of infants using image processing in the field of computer vision. In this paper, we propose a different approach to detect pain in infants that outperforms previous research in this field. We merged a face area-based feature collection method with a local binary pattern (LBP). Moreover, three different machine learning algorithms have been executed to find the best parameter to get a decent accuracy on the iCOPE dataset. The proposed method uses the SVM classifier to achieve 86% of testing accuracy compared to other methods.

This project focuses on automated neonatal pain assessment using computer vision techniques on the Infant COPE (iCOPE) dataset. Two face area–based feature extraction approaches were explored using Local Binary Patterns (LBP).

In the first approach, LBP features were extracted and evaluated using multiple classifiers, including Random Forest, Linear Regression, and SVM, with SVM showing strong performance compared to prior methods.

In the second approach, facial regions were detected using the MediaPipe framework combined with LBP features. Among four evaluated machine learning models, the Random Forest classifier achieved the best performance, reaching 95% testing accuracy.

Face Recognition-based Attendance Management System

Fig : Class Diagram

Objectives

>Making a platform to ease the attendance management.

>To overcome the problem of manual attendance and provide reliable record maintenance.

Lab Result Management System ( Using Java)

The project's major goal is to provide the exam results in an understandable manner. This initiative helps educators and institutions provide outcomes in an easy-to-understand manner. The Teacher is the target user of the system. And by entering a user name and password for a secure login, teachers are given the ability to read, write, update, delete, and execute results. The administrator will have complete control over the result analyzer and will have the ability to read, write, and execute the results.

Technical Used:

Programming platform : Netbean.IDE

Front-end & Back-end: Java

Database : MySQL.