Undergraduate Projects -

The relevant projects from my undergraduate days.

Laboratory/Design Projects:

Multimedia Research Laboratory @ BITS Pilani

Mentors : Prof. Kamlesh Tiwari, Prof. Surekha Bhanot

Topic : Image to text translation with deep neural networks for the speech impaired

Duration: August - December 2017

  • The linear layers and the classifier of an AlexNet model, trained on the ImageNet dataset, were finetuned on the ASL Fingerspelling dataset A; data augmentation techniques were used on the ASL dataset to prevent overfitting by Alexnet
  • Suitable training hyperparameters were experimentally found through grid search
  • I learnt about language modeling with deep neural networks and experimented with common bag of words (CBOW) model for the Tiny Shakespeare dataset
  • A character RNN was trained on the Tiny Shakespeare dataset to complete sentences given a primer text of characters
  • Finally, the outputs/character labels of the fine-tuned CNN are used to provide the primer text for the RNN to generate the rest of the text
  • This work can have practical application in the form of voice generation system for the speech impaired
  • Most of the code was written in Pytorch and Tensorflow

Project Report

Multimedia Research Laboratory @ BITS Pilani

Mentor : Prof. Surekha Bhanot

Topic : Face recognition using convolutional neural networks

Duration: January - May 2017

  • I studied the working of CNNs and their use for different computer vision tasks
  • (VGG-D , VGG-E) were trained end to end on Color Feret and Youtube Faces datasets
  • The linear layers and the classifier of VGG-E models trained on Youtube Faces, were finetuned on the LFW dataset; the OpenFace toolbox was used for scaling and aligning the images from LFW dataset because the images from LFW are far more diverse in style due to variety in poses, lighting etc.
  • I experimentally found suitable training hyperparameters through grid search
  • Caffe and python were used for coding most part of the project
Project ReportCodebase

Collaboration Projects:

Nano-Bio Sensors Group @ CSIR-CEERI Pilani

Mentor : Mr. Soumendu Sinha

Topic : Temperature compensation with machine learning models in ISFET based pH-meter

Duration: January - May 2017

  • An ISFET based pH-meter was designed at the fabrication unit of the institute
  • pH data at different temperatures (pH sensitivity with temperature) was collected
  • Features from the collected data were centered around the feature means and scaled down by the inverse of the feature standard deviations
  • Several regression models (3-layered feedforward fully-connected neural network, Random Forrests, Support Vector Regression, Polynomial Regression) were trained on the collected data for temperature compensation
  • Majority of the code was written using scikit learn and numpy in python.
  • Two papers from this work have been published at IEEE RSM 2017 and IJCTA 2019
Publication 1Publciation 2

Minor Projects:

Multimedia Research Laboratory @ BITS Pilani

Mentor : Prof. Kamlesh Tiwari

Topic : Scale specific CNNs for scene recognition

Duration: March - April 2017

  • An ensemble of scale-specific CNNs to recognize scenes of variable scales was designed
  • I reproduced baseline results from the publication on SUN397 dataset

Project Report

Multimedia Research Laboratory @ BITS Pilani

Topic : Online traffic sign recognition system with Canny edge detection and CNNs

Duration: March - April 2017

  • Traffic signs from traffic scene images in GTSRB dataset were detected using Canny edge detection
  • A VGG-19 ImageNet model was fine-tuned on the data for traffic sign classification
  • I built a GUI using PyQT4 for online inference of traffic scene images using the model
  • The project was presented at APOGEE 2017, BITS Pilani annual technical festival

Wireless Sensor Networks Systems Laboratory @ CSIR-CEERI Pilani

Mentor : Dr. Kota Solomon Raju

Topic : Signal denoising of household power lines using wavelet transform in python

Duration: August - December 2016

  • An in-depth literature survey of signal denoising in house-hold power lines was done
  • Finally, wavelet transform was chosen as the method to denoise in combined time-frequency domain
  • Code was written in python for performing denoising on real-time simulations of household power signals