Research

1. Multimodal Blood Flow Velocity wave cycle classification with explanation of decisions

This project takes a multimodal deep learning approach to classify ocular Blood Flow Velocity (BFV) wave cycles (corresponding to each cardiac cycle) into "Acceptable" (i.e., suitable for further analysis in diagnostics) or "Noisy" (i.e., to be discarded) by using a combination of corresponding 2D retinal Laser speckle contrast image (LSCI) sequences and 1D blood flow velocity wave cycles as inputs. The prediction output of the multimodal deep learning model can be explained (to ensure its trustworthiness) with respect to each input modality, using Gradient-weighted Class Activation Mapping (Grad-CAM), which creates a heatmap to show which parts of each input are most important for the deep learning model's decision.

Keywords: Multimodal classification, Grad-CAM, Explainability in AI (XAI), TensorFlow, Keras, Deep Learning
GitHub Repo: Link

2. Multiple ellipse detection from noisy points using the RANSAC algorithm

This project develops an implementation of multiple ellipse detection from noisy points using the RANSAC (RANdom SAmple Consensus) algorithm.

Keywords: Multiple Ellipse Detection, RANSAC, Noisy points
GitHub Repo: Link

3. Defect Detection and Localization in Industrial Products with Class Activation Mapping

In this project the goal is to detect the presence of any metallic defects and locate them in images of industrial products using Class Activation Mapping (CAM). CAM highlights the discriminative regions in an image that a Convolutional Neural Network (CNN) uses to predict a specific class. By visualizing the CAM for a particular object class, one can localize where that object is located in the image.

Keywords: Class Activation Mapping, Defect Detection, Localization, Computer Vision, Deep Learning, Keras
GitHub Repo: Link

4. Real-time hand gesture recognition with Deep Learning

This project achieves the aim of recognizing the hand gestures of a person in real-time from the video provided by a webcam and displays the gesture label on the input camera frame in real-time, along with the probability of the detected label being correct. A modified LeNet-5 architecture is used to classify the 9 hand gestures.

Keywords: Gesture Recognition, OpenCV, LeNet, Computer Vision, Deep Learning, Keras
GitHub Repo: Link
Paper: Link

5. Real-time Facial Emotion Recognition using Deep Learning

This project uses a modified VGGnet architecture of deep learning and the fer2013 dataset. Given the images from a real-time webcam video input, the goal is to categorize the emotion expressed on each face in each frame into six distinct classes: angry, afraid, happy, sad, surprised, neutral.

Keywords: Emotion Recognition, Deep Learning, VGG Net, FER2013, Computer Vision, Keras
GitHub Repo: Link
Paper: Link

6. Drowsiness Detection System Using Deep Learning

Drivers drowsiness is the major problem that causes road accidents. The drivers' drowsiness state is estimated using the facial regions corresponding to the entire face. The algorithms employed for face detection are i) Viola Jones ii) DLib iii) Yolo V3. For the Classification and drowsiness detection, a CNN (Convolutional Neural Network) architecture is employed.

Keywords: Convolutional Neural Network (CNN), Drowsiness detection, Face Detection, Viola Jones algorithm, dlib, YoloV3, LeNet, DarkNet-53, Deep Learning, multi-modal analysis
Paper: Link

7. Eye Tumour Detection Using Deep Learning

In this project a deep learning algorithm with modified LeNet architecture has been developed to detect eye tumors. Due to limited availability of eye tumor images, data augmentation techniques were used to increase the number of images for training. For segmenting the circular boundary of the iris, Hough transform was used.

Keywords: Eye Tumor, Cancer, Deep Learning, CNN, LeNet, Hough Transform
Paper: Link

8. Brain Tumour Detection Using Deep Learning

In this study, MRI scans of the brain are used to analyse the health condition of the patients. Objective of this research is to identify the abnormal image and segment the tumour region of the brain.

Keywords: Medical Image Processing, Brain tumour, MRI, Artificial neural network, CNN, Keras
GitHub Repo: Link
Paper: Link