Projects

Look Who's Talking : Reconstructing faces from voices


Dr Najim Dehak, ECE, Johns Hopkins UniversityAug 19 - Dec 19

Used Generative Adversarial Network to synthesize face images from voice samples using the EmoVoxCeleb dataset. Added an emotion recognition module which improved the reconstructed faces images' PSNR by 1.29 dB and SSIM by 0.051.

[Presentation] | Code coming soon...

Multi-Modal MRI Brain Segmentation & Survival Rate Prediction


Prof Jerry Prince, ECE, Johns Hopkins UniversityFeb' 20 - May' 20

Used Separable 3D U-Net with histogram matching for the brain tumour segmentation task to achieve a dice score of 0.67. Extracted hand crafted intensity, shape, and texture features with Tree Ensemble regressor for the survival task.

[Presentation] | [Report] | Code coming soon...

Lesion Classification in endoscopy videos using 3D features


Dr. Vishal Patel, ECE, Johns Hopkins UniversityFeb' 20 - May' 20

Employed Multi-View CNN (MV-CNN) and 3D-ResNet to perform 3D shape based polyp classification in colonoscopy videos. Added channel based attention for feature fusion in MV-CNN to improve accuracy by 3.13%.

[Presentation] | [Report] | Code coming soon...

Using Wavelets in Deep Learning Methods for Medical Image Segmentation


Prof. Trac Tran, ECE, Johns Hopkins UniversityAug' 19 - Dec' 19

Investigated approaches to integrate wavelets into U-Net architecture. Wavelets used as a pre-processing step improved the accuracy of U-Net by 5% with a reduction in training time for US fetal brain anatomy segmentation.

[Report] | [Code]

Analyzing Sparse Stochastic Gradient Descent for Optimization Tasks


Prof Trac Tran, ECE, Johns Hopkins UniversityFeb' 20 - May' 20

Analyzed performance of sparsified SGD for optimizing deep networks for classification, regression and reconstruction tasks. Showed a 10-40% reduction in the number of floating point operations (FLOPs) while maintaining accuracy comparable with vanilla SGD.

[Presentation] | [Report] | [Code]

Segmentation algorithm and performance evaluation for 2D Choroidal Optical coherence tomography (OCT) images

Internship Project, Guide : Prof. Sumohana S. Channappayya, IIT HyderabadMay 15 - July 15

  • Automated detection of Choroid-Sclera Boundary in 2D-OCT Choroidal images : Developed image segmentation algorithm based on gradient methods like edge detection and morphological operations.

  • Objective Performance evaluation of blood vessel detection algorithms for 2D choroidal OCT images : Four similarity measures based on Dice Coefficient, Hamming distance, mismatch area, van Dongen distance and pairs of pixels classification were proposed for comparing various vessel detection algorithm’s results.

[Presentation]

Classification of polyp in endoscopic images using deep learning

Course project Guide : Prof. M. K. Bhuyan, IIT GuwahatiJan’ 17 - May’ 17

Classified polyps into benign and malignant classes based on pit pattern classification using AlexNet inspired Convolutional Neural Networks(CNNs).

[Report]

Eye gaze controlled mouse pointer

Course projectGuide : Prof. Prithwijit Guha, IIT GuwahatiJul’ 15 - Nov’ 15

Developed a robust eye gaze controlled mouse pointer system using a webcam. The algorithm for gaze tracking used Viola-Jones algorithm for face detection, K-means clustering for thresholding and gradient descent for finding the pupil location. Similarity of triangles was used for mapping pupil movement to cursor movement.